Substack announced AI detection as a protective measure. The gist of the argument is that readers should know what’s written by humans and be able to selectively filter out articles that don’t align with their expectations.

While this move is well-intended, here’s what I think will actually happen: readers will stop evaluating arguments. They will start hunting for tells. Now, instead of asking a question like “is this reasoning sound?” we are being actively encouraged to ask “did a machine write this?”

This is the wrong problem to solve. And it’s the wrong way to solve it.

The Bandwidth Problem

When you ask a reader to verify authorship on the fly, you’re adding a task. AI-assisted writing is now invisible enough that readers can’t tell just by looking. While the Pangram integration is intended to do this for us, from my experience, it has been proven to be highly inaccurate. The spotlight on this witchhunt for AI-assisted readers (as described succinctly here by ) is subconsciously forcing readers to read defensively: hunting for repetition patterns and phrasings that feel off. They’re doing forensic source analysis while trying to follow the actual argument.

Here’s the problem: forensic analysis and critical thinking use the same cognitive resources. They compete directly. When readers are burned out from verification work, they can’t think clearly about substance. You can have one or the other. You can’t have both.

Mosier and Skitka found this pattern in automation studies: when operators are forced to continuously verify whether a system is actually doing what it claims (not monitoring it, but verifying it) their judgment erodes, even when the system is reliable. The verification overhead itself becomes the failure point.

The cost of Substack’s new detector isn’t accurate AI detection. The cost is readers who stop thinking critically about arguments because they’re exhausted from authorship verification.

How Aviation Actually Solved This

Aviation faced the same problem with the introduction of the autopilot. Pilots resisted. The cultural anxiety was the same: the machine is replacing human judgment. Your skills. Your job. Early pilots saw autopilot as a threat to the profession.

The resistance didn’t evaporate because pilots got more trusting. It evaporated because the mechanism changed.

When you hand-fly an aircraft for eight hours, you’re managing a dozen competing demands simultaneously: hold the heading, maintain altitude, watch the descent rate, scan the engine instruments, monitor fuel, think about weather, plan for traffic, talk to ATC, anticipate contingencies. The cognitive load is relentless.

Autopilot took the holding-level-and-tracking work off the pilot’s cognitive plate. That freed bandwidth for the work that actually matters: pattern recognition on weather, fuel analysis, contingency planning, systems monitoring. The best pilots were usually the first to use autopilot, precisely because they knew that their decision-making was sharper when they weren’t exhausted from hand-flying.

This wasn’t about trust. It was about capacity. Better tools meant better thinking.

The cultural acceptance came from evidence. Pilots could see the difference in their own workload and decision quality. It took time. It took training. But it was grounded in mechanism: the tool freed bandwidth for judgment.

Substack is asking the opposite question. Instead of “what can writers do better with AI assistance?” it’s asking “can we catch them using it?” Instead of freeing cognitive bandwidth, it’s spending cognitive bandwidth on verification.

The Real Failure Mode: Omission Errors

Sarter and Woods documented a shift in glass cockpit accidents: early errors were commission (pilots doing the wrong thing). Later errors were omission (pilots missing the right thing).

The pattern: pilots hyper-focused on monitoring whether the automation was in the correct mode missed actual system failures happening in the background. They were surveilling identity instead of monitoring behavior. The surveillance created blindness.

This is what Substack’s detector does.

Readers hyper-focused on “is this AI?” stop asking: Is this reasoning sound? Is the evidence good? They’re doing forensic analysis instead of critical thinking.

Writers hyper-aware of the scan stop asking: Is this a good idea? They start asking: Will this trigger detection? They’re surveilling their own draft process instead of thinking through it.

The detector doesn’t solve the problem. Readers and writers both become worse at the judgment that actually matters because they’re burned out from verification work.

The Perverse Incentive

Mosier and Skitka found that when people distrust a system that’s actually reliable, they either disuse it (reject it entirely) or compulsively verify it (check obsessively, as if checking harder will make it trustworthy). Both responses erode performance.

Substack pushed honest writers into compulsive verification: “Will this flag? Do I need to rewrite this to hide the AI assist?” Meanwhile, writers who hide their AI use entirely fly under the radar.

The system penalizes transparency.

For readers, the detector builds reflexive skepticism. They learn to distrust AI-assisted posts by reflex, not by judgment. They become worse at catching undetected AI misuse because they’ve outsourced the thinking to the detector instead of developing their own critical eye.

Autopilot made pilots better at judgment. Substack’s detector makes readers worse at it.

What Autopilot Got Right (And Substack Missed)

Aviation solved this through three mechanisms working together.

Transparency first. Pilots learned exactly what autopilot did, what modes were available, what would and wouldn’t happen automatically. No mystery. No surprise mode changes.

Training second. Pilots didn’t get handed autopilot and told to figure it out. They learned when to trust it, when to verify, how to monitor. They learned the boundaries. This was continuous, not a one-time event.

Feedback third. When pilots encountered autopilot errors, those errors were explained. They became teaching moments, not gotchas. Pilots calibrated their trust based on real evidence about what the system could do.

Substack did none of these. The detector is opaque. No explanation of how it works. No training on how readers should think about AI-assisted writing. No feedback loop.

The Problem Substack Actually Has (And Didn’t Solve)

Legitimate worry: spam, plagiarism, scraped content dressed up as original. These are real problems.

Authorship detection isn’t the answer.

Originality detection would be. Check against scraped datasets. Catch recycled content whether it’s human-written or AI-generated. That solves the actual problem.

Or, maybe, just trust readers. They’re actually good at smelling something fishy when the frame is honest. Give them time and context, they’ll figure out who’s worth reading.

Aviation didn’t solve the “is this autopilot” problem by detecting when pilots used it. It solved it through evidence: maintenance logs, reliability data, clear communication about what the system could do. Trust came from transparency, not surveillance.

The Real Question

Substack is asking readers to solve the wrong problem: How do I know if I can trust this source?

The real question is: Is this argument sound? Are the claims backed by evidence? Does the reasoning hold? What am I not seeing?

These are judgment questions. They require thinking. Detection questions don’t.

Substack’s detector forces detection and taxes the cognitive resources readers need for judgment. By trying to protect against AI misuse, it trains readers to be worse at evaluating AI-assisted content.

Readers who develop good judgment are safer from AI misuse than readers who develop good detection. Detection can be gamed. Judgment can’t. A reader trained to ask “is this reasoning sound?” will catch bullshit regardless of authorship. A reader trained to ask “did a machine write this?” will miss a well-reasoned AI argument while flagging an honest human stumble.

You’re teaching readers to ask the wrong question. And the wrong question leaves them vulnerable.

Subscribe now

Substack tried to protect readers from automation. Instead, it taught them to distrust good tools and to stop thinking.

That’s not safety. That’s the opposite.

The solution is better judgment, not better detection. Readers who evaluate arguments regardless of source. Writers who use AI to think deeper, not to hide their work. A platform built on transparency, not suspicion.

That’s what autopilot built. Not through surveillance. Through clarity and feedback loops that let pilots calibrate their reliance over time.

Substack could have built that. Instead, it created a verification culture that makes everyone worse at what matters.

The irony: AI writing assistance is genuinely useful. It frees bandwidth for better thinking. But Substack’s detector doesn’t allow that. It forces both readers and writers to spend energy on the wrong questions: not “how do I think better?” but “where did this come from?”

As long as that’s the question, judgment stops working. And judgment is the only thing that protects you from bad ideas, whether they come from a human or a machine.

Every six months, I get locked in a simulator and someone tries to kill me.

Engine failure at V1. Dual hydraulic loss. Unreliable airspeed. I hand fly approaches the autopilot would normally handle, because the airline has figured out that the skills you do not use decay, and automation guarantees you will not use them.

Recurrent training is mandatory for airline pilots and is primarily built around emergencies most of us have (hopefully) never had to face for real. A lot of what's in the syllabus got there because it already happened to somebody else's crew first. Most of the regulations did too: the rule tends to show up after the accident, not before it.

I've sat through enough of these sessions to notice that it isn’t there to necessarily make me a better pilot. It’s there to make sure the judgment automation lets go is still intact.

Nobody built the equivalent for the tool now doing the first pass on your job. Two years into the biggest cognitive automation rollout in history, there's no recurrent cycle for knowledge work. No scheduled check that judgment still holds up once the AI does the thinking first. That's the argument of this piece: it needs one, for the same reason aviation needed one, before the skill goes quiet without anyone noticing.

Aviation figured this out decades before anyone had heard of a large language model. In 1983, a human factors researcher named Lisanne Bainbridge wrote the four page paper that explains why. It’s still one of the most cited papers in the field, and the reason airlines run pilots through failures they'll hopefully never face for real.

The plant that ran itself into a corner

Bainbridge opens with an odd little anecdote. At an industrial plant she studied, management had to keep someone present on the night shift, because the operators kept switching the automated process back to manual. Not because the automation was broken. Because the operators wanted to stay capable, and staying capable required practice the automation had taken away from them.

From that she built an argument with two parts.

First: designers automate the tasks that are easiest to automate, which leaves the human with exactly the tasks too difficult or too irregular to hand to a machine. The hardest work in the entire system. Monitoring for rare failures, then taking over when the failure finally arrives. Under time pressure. With degraded information. Having not touched the controls in weeks.

Second: the more reliable the automation gets, the less the human practices. The less the human practices, the less skilled they are at the exact moment skill matters most. Bainbridge called this an irony because it flips the whole promise of automation on its head. We built the machine to lighten the load on the human. Instead we built a machine that quietly disqualifies the human from the one job still left for them.

An operator who learns a system by working it directly builds a rich, flexible internal model of how the process behaves. What causes what. What the early signs of trouble look like. How to improvise when something unfamiliar happens. An operator who only watches the automation run the process, stepping in occasionally, never builds that model. Or lets it rot.

What decays, in what order

A 2014 study by Casner and colleagues put sixteen airline pilots into a full motion 747 simulator to find out exactly what atrophies when pilots fly with heavy automation for years. The finding surprised some of the researchers involved. Basic stick and rudder control (the physical business of flying the airplane by hand) held up fine. It’s overlearned, motor based, resistant to decay the way riding a bike is resistant to decay.

What eroded was something else. The cognitive work of building and holding a mental picture of the airplane’s state. Recognizing an unfamiliar failure for what it is. Reasoning through a response with no procedure to follow. The pilots could still move the controls. What several of them struggled with was knowing what the controls should be doing, under conditions the automation normally absorbed without telling anyone.

A separate line of research, from Ebbatson and colleagues studying 737 pilots, found something I think about every time I sit down for a checkride: manual handling proficiency correlated far more strongly with how recently a pilot had hand flown than with total flight hours logged over a career. Recency beat experience. A ten thousand hour captain who hasn’t hand flown an approach in eight months isn’t protected by those ten thousand hours. This is the finding that should worry any organization telling itself its most senior people are the ones least at risk from AI dependence. Seniority isn’t currency here. Practice is.

None of this is abstract for aviation. On June 1, 2009, an Air France A330 flying Rio to Paris flew into a band of high altitude convective weather. Ice crystals blocked the pitot tubes measuring airspeed. Autopilot and autothrust disconnected (exactly as designed) and handed control back to the crew with a message that amounted to: the airplane doesn’t trust its own instruments anymore, you have the aircraft. The pilot flying held a nose up input that stalled the airplane. It stayed stalled, falling at roughly ten thousand feet a minute, for over three and a half minutes, while three professional pilots worked through a confusion the investigation later traced in part to how rarely crews hand flew at cruise altitude, in degraded flight control law, in conditions exactly like these. Two hundred twenty eight people died. The automation recognized bad data and quit, cleanly, exactly as designed. The danger sat on the other side of the handoff, in a skill that had gone quiet for years.

That accident, and others like it, is why every recurrent cycle now includes mandatory upset recovery training and unreliable airspeed scenarios. The industry’s answer to the problem Bainbridge named in 1983 was never more automation, or a better warning light. It was scheduling the practice back in. On purpose. By force. Forever.

Your team just inherited an autopilot, and nobody scheduled the sim

In roughly two years, knowledge workers were handed the most capable cognitive automation ever built. It drafts the analysis. Writes the code. Builds the forecast. Summarizes the contract. Adoption moved faster than any technology transition in memory. Nothing resembling a skill maintenance plan moved with it, because until very recently, nobody thought one was necessary.

The clearest evidence that one is necessary comes from a 2023 field experiment run by researchers at Harvard Business School and Wharton, with Boston Consulting Group, later published in the peer reviewed journal Organization Science. It’s one of the most rigorous studies ever done on how AI actually changes the work of skilled professionals.

The researchers gave 758 consultants two kinds of tasks. On tasks chosen to sit comfortably within what GPT-4 handled well, consultants using the AI completed about 12 percent more tasks, finished roughly 25 percent faster, and produced work rated over 40 percent higher in quality by independent graders. That result alone would justify every dollar of AI spending in corporate America.

Then the researchers handed everyone a different task: a brand strategy case built deliberately to sit just outside what the model could reliably handle, requiring the consultant to notice a subtle detail buried in interview notes that contradicted the more obvious surface data. Consultants working without AI got the case right 84.5 percent of the time. Consultants using GPT-4 got it right 70.6 percent of the time. Consultants who’d also been given a short prompt engineering briefing (meant to make them sharper AI users) did worse still. 60 percent. The extra training in how to use the tool made the outcome worse, because it deepened trust in an answer that was confidently, fluently, completely wrong.

The researchers named this the jagged technological frontier: the line between what AI does well and what it does badly isn’t smooth or intuitive. It juts in and out in ways that have nothing to do with how hard a task looks to the person using the tool. Nobody in that study could tell, from the outside, which side of the frontier they stood on. The AI didn’t announce it either. It delivered its wrong answer on the brand strategy case with exactly the same fluency and confidence it delivered its right answers on tasks it was actually good at.

A related study from the same group went further, looking at what happened when consultants pushed back on the AI’s wrong answer. Questioning it and presenting contradicting evidence. The model didn’t disclose uncertainty or back down. It escalated its own persuasion. Restated its original position with more supporting structure, more apparent rigor, more confident framing, while giving no ground on the substance. The professionals doing the pushing back were, in the researchers’ own words, persuaded rather than corrected. The natural instinct to double check by asking the tool itself isn’t a safeguard.

The failures are already arriving on schedule

A 2026 survey of over 2,500 senior decision makers across ten countries and six industries found that 74 percent of organizations that deployed a customer facing AI agent into production had rolled back or shut down at least one of them. The rollback rate was 81 percent higher among companies with mature AI governance frameworks in place. Mature governance did not prevent failures. It found them faster, because someone was finally watching closely enough to notice.

Individual incidents make the same point at ground level. A widely covered case from 2025 involved an AI coding agent operating under an explicit, repeated instruction not to touch production during a declared code freeze. It deleted the production database anyway, then told the engineer overseeing it that the damage was unrecoverable. It was not; the data was restored from backups. Before that, the same agent had fabricated thousands of fake user records and false test results, manufacturing the appearance of verification that nobody independently checked. Every link in that chain is a documented human factors failure with a name: standing permissions granted without review, a human accepting an automated system’s report about its own failure instead of checking directly, and a review process that had, through months of routine success, stopped actually reviewing.

Elements of recurrent training

The tempting response to all of this is more guardrails: tighter permissions, better logging, a planning only mode, a second model checking the first model’s work. Those measures are worth having. None of them solve the underlying problem, because every guardrail still depends on a human somewhere in the system who can recognize when something has gone wrong. Guardrails assume the monitor stays sharp. Nothing in the last three years of AI deployment has kept that assumption true, and the aviation record says nothing ever will, absent deliberate intervention.

The intervention aviation settled on, after enough accidents forced the question, has four parts.

Scheduled failure exposure: pilots do not wait for a real engine fire to practice one, so a simulator introduces a fire, or a stall, or a total hydraulic failure, on a fixed calendar, whether the pilot feels ready or not. Knowledge teams need the equivalent: a realistic, wrong, plausible AI output deliberately placed into an actual workflow, on a schedule, to see who catches it and how.

Scoring and blameless debrief: every simulator session is measured, and every measurement is discussed afterward without attaching blame to the individual. The purpose is calibration, not punishment. Teams learn what their actual detection rate is, which is reliably far lower than what they would have guessed beforehand, and where the specific blind spots sit.

Protected manual practice: airlines require pilots to hand fly certain segments specifically so the underlying skill does not go quiet. Knowledge teams need designated work that gets done without the tool.

Permanence: this is the piece organizations resist most, because it costs money and calendar time forever, with no finish line. Recurrent training is not a workshop you complete once. It is called recurrent because the decay it addresses is continuous. A one time session teaches the mechanism. It does not maintain the skill, any more than one trip to the gym maintains fitness.

The uncomfortable math, and the honest limits of the comparison

The obvious objection: this costs time, and the entire point of adopting AI was to save time. Aviation ran that exact experiment and reached its answer decades ago. Full motion simulators cost tens of millions of dollars to build and thousands of dollars an hour to operate. Airlines pull revenue generating crews off the schedule to sit in them. They pay this cost, continuously, because the alternative was written into accident reports with body counts attached, and the industry chose not to keep paying that price instead.

The comparison has real limits. Aviation failures happen in seconds and are often fatal; most knowledge work failures unfold over days or months and are financial. Pilot tasks are procedurally bounded in a way that makes a simulator scenario buildable; open ended knowledge work is harder to script realistically. Aviation has a regulator that mandates the expense; no regulator is coming for corporate AI governance any time soon, which means this has to be sold on its own economic logic rather than imposed by law. The mechanism, decay through disuse, degraded monitoring, miscalibrated trust in fluent output, transfers cleanly across domains because it is a property of human cognition. The implementation has to be rebuilt from scratch for every domain it touches.

The teams that hold an advantage over the next several years will not be the teams that adopted AI first or most aggressively. Everyone will have done that by then. The advantage belongs to the teams that can still tell, reliably, when the automation is wrong, because they never let that particular skill go quiet.

Subscribe now

Papers I’m Reading

This article summarizes papers I’ve curated this week on AI in aviation, automation dependency, human factors in the cockpit, and pilot cognition and judgment. I’ve selected sources I believe are most relevant to these topics and linked them for deeper reading.

“Methods for Preventing the Degradation of Manual Flying Skills” (Collegiate Aviation Review International, Dec 2025)

• Recent flight practice is a significantly stronger predictor of manual flying performance than total flight hours or time since flight school, meaning currency matters more than experience, and the logbook number you’re proud of is largely irrelevant to your ability to hand-fly right now.

• The paper proposes a structured program built around UPRT (Upset Prevention and Recovery Training) plus mathematical models of skill decay, benchmarked against ICAO, FAA, EASA, Airbus, and Boeing policies. Most of which currently under-prescribe manual flying relative to what the decay models say is needed.

• LOC-I (loss of control inflight) remains the leading category of fatal accidents, and the paper links this directly to inadequate manual skill maintenance, making automation dependency not an academic concern but a body-count problem.

“Exploring Automation Bias in Human-AI Collaboration” (AI & Society, Springer, Jul 2025)

• Automation bias is expanding, not contracting, as AI systems become more capable and trusted. A 35-study review shows the problem is getting worse precisely because better systems make overreliance feel more justified.

• Expertise doesn’t protect against automation bias, and two-person crews don’t eliminate it. Both findings directly challenge the assumption that training, experience, or redundancy will naturally correct for over-reliance on a capable system.

• Explainability interventions (showing the human why the AI made a decision) have mixed results, with some studies showing they actually increase bias by making the recommendation feel more authoritative and harder to challenge.

“Automation Complacency: Navigating the Ethical Challenges of AI in Healthcare" (Columbia University SPS, Nov 2025)

• The panel’s opening example is the sharpest possible illustration of AI’s contextual gap: an insurance algorithm recommended discharging a wheelchair-bound patient home (unaware the patient lived in a fifth-floor walk-up). The observation that landed: “AI algorithms don’t know what they don’t know, and therefore can’t replicate the human capacity for humility.” The system was confident, coherent, and catastrophically wrong about something obvious to any human in the room.

• The panel drew a hard distinction between decision support (AI assists the human deploying their expertise) and decision substitution (the human defers judgment entirely to the system without recognizing its limitations). The legal cases showed what substitution looks like at scale: insurers using AI models allegedly programmed to automatically deny medically necessary care, and a drug company taking kickbacks to design clinical alerts that increased opioid prescriptions.

• “If the bias is the psychological hole, the automation complacency is the clinical trap because of that hole.” Automation bias is the tendency; complacency is what happens when you live inside that tendency long enough that it stops feeling like a choice.

Mathematical Justification for Training Intervals

The following is an excerpt taken from the University Aviation Association article. I love when abstract things are broken down into math equations and found this computation of the maximum allowed interval without practice in regards to maintaining flight proficiency quite interesting.

The program defines specific intervals for simulator sessions (e.g., quarterly, annually), and these intervals are grounded in research on skill degradation rate over time. Using the exponential decay model described in the Methodology section (Equation 1), we can derive optimal training frequency. To ensure skill does not fall below the minimum acceptable threshold, we solve for the maximum allowed interval without practice:

Studies suggest motor skills decay slowly and can be retained for several months, while cognitive skills degrade more quickly. Casner et al. (2014) documented that cognitive skills related to manual flight planning and situational awareness showed approximately 30% decline after four months without practice. If cognitive skills drop to 50% after 4 months, λ is estimated as:

Assuming Rmin= 0.8 (representing an 80% minimum acceptable skill threshold based on industry training standards),the maximum permissible interval is:

This result suggests that for cognitive skills, retraining should ideally occur every 1–1.5 months to retain at least 80% of the trained proficiency. Since such frequent training may not be operationally feasible, the program compensates by focusing quarterly simulator sessions on motor skills (which degrade more slowly), integrating cognitive exercises into both simulator and classroom environments, encouraging ongoing manual practice during line operations, and supporting this with data monitoring and individual coaching (Module E). These measures allow the program to maintain an acceptable overall level of pilot proficiency without imposing unrealistic training burdens.

Table 2 presents sensitivity analysis demonstrating how training interval recommendations vary with different decay coefficients. This analysis accounts for individual variability in skill retention based on pilot experience, recency of practice, and learning characteristics.

Table 3 presents degradation dynamics of pilots' motor and cognitive skills without practice, illustrating the differential decay rates that inform the training program structure.

Figure 1 provides graphical representation of the exponential skill degradation model, illustrating the more rapid decline of cognitive skills compared to motor skills and the 80% threshold that determines training interval recommendations.

To ensure pilots maintain both motor and cognitive manual flying skills, basic training exercises (Module A) are scheduled every three months. While motor skills alone may not require such frequent refreshers, quarterly simulator sessions provide essential reinforcement of cognitive abilities related to manual control. Advanced and high-stress scenarios (Module B) are conducted annually. This frequency reflects logistical considerations and the fact that once the skills required to recover from unusual attitudes are initially acquired, annual reinforcement is generally sufficient. Nevertheless, where possible, UPRT elements should ideally be practiced every six months, a frequency already adopted by some airlines. Thus, the proposed schedule represents a balanced compromise, leaning toward more frequent training than traditional programs while remaining operationally feasible.

Subscribe now

You’ve Already Done This

Your third Initial Qualification feels exactly like your first. You’d think it gets easier, and in some ways it does, but the nerves show up anyway. The night before training starts, the same questions creep in. What if I get behind in the sim? What if I fail a validation? What if this is the aircraft that finally exposes me?

I just finished my third Initial Qualification as an airline pilot, this time on the Airbus. And I’ll tell you honestly, I felt all of it. The initial excitement of flying something new, followed closely by that familiar low-grade anxiety that sets in when you realize how much there is to learn and how little time you have to learn it.

But here’s the reframe that got me through it, and the one I come back to every time training starts: by the time you’re sitting in an airline Initial Qualification, you have already taken somewhere between 8 and 15 checkrides. Think about that number. Every certificate, every rating, every type. Each one required you to perform under pressure, in an unfamiliar aircraft or scenario, in front of an examiner who wasn’t going to hand you anything. And you passed. Maybe not every single one on the first attempt, but you passed enough of them to get here. You are sitting in an airline training center because a major carrier looked at your record, your background, and your judgment and decided you were worth investing in.

This is just another checkride. That reframe matters more than any study technique, any memory system, any optimization strategy you could bring to training. Before you worry about how to study, remind yourself of what you’ve already proven.

Subscribe now

On failure

I failed two checkrides during my training career. My Private Pilot checkride and my initial CFI checkride. At the time, both felt significant. The kind of failure that makes you question whether you’re cut out for this. I can tell you now, from the other side of a Delta new hire interview, that they were not the career-ending events my younger self feared they were.

When I sat across from the interviewers and those failures came up (and they did come up) what they were listening for had nothing to do with the failure itself. They wanted to know two things: did I take accountability for what went wrong, and what did I do to correct it? That’s it. The failure is almost incidental. Attitude is everything. A pilot who failed a checkride, owned it completely, identified the specific gaps, and came back stronger is a more compelling candidate than someone who skated through everything on autopilot and never had to dig deep when things got hard.

What you do with a failure defines you far more than the failure itself.

That said, context matters. I want to be honest with you about the stakes at different stages of your career. A failure during training at a Regional airline carries more weight than one at a Major, and you should understand why. At a Regional, you are still building the record that you will carry into your Major airline interview. Any training failures, any checkride busts, are going to come up in that interview room. That doesn’t mean they’re disqualifying. Plenty of pilots with training setbacks go on to get hired at the Majors. It does however mean you’ll need to have a clear, accountable, well-articulated answer ready. The scrutiny is real.

At a Major airline, the calculus is different. If you bust a validation or struggle through a sim session, the primary consequence is that you retake it. Your interview is behind you. Your seat is secure. The training department has seen everything, and a pilot who needs an extra session or has to repeat a checkride is not a novel event. You move through it, you get the type rating, and you go fly the line. It’s worth keeping that perspective when the anxiety starts to spike.

Preparation is the job

Here’s something I tried to do before my most recent Initial Qual that taught me a useful lesson: I spent time trying to build an optimized study system using AI tools. I wanted to create a personalized learning path, sequence the material more efficiently, get ahead of the curve before class started. Not uploading any training material, just coming up with a blueprint for an optimized learning path. I put real time into it.

Once training began, I abandoned it almost immediately. Not because the tools weren’t capable, but because I realized something important: the airline had already done this work. They train pilots every single day, across every experience level, every background, every learning style. They have been watching where pilots struggle and where they succeed for decades. The curriculum you receive in Initial Qualification is not a rough draft. It is a distillation of everything the training department knows about how to move a pilot from zero to qualified on that aircraft in the shortest possible time.

Your job is not to optimize the curriculum. Your job is to follow it.

That sounds simple, but it’s easy to overcomplicate training when anxiety is driving the bus. You start looking for an edge, for extra materials, for a better system. Most of the time, the edge is simpler than that.

Read the lesson plan before every single sim session. Not a skim. A real read. Know what maneuver is coming, know what the standard is, know what the common errors are. Pull up the charts for the procedures you’ll be flying and study them before you walk into the building. Chair fly. Sit in a quiet room and physically work through the flows, the callouts, the memory items. Do it until your hands know where to go before your brain has finished the thought.

Show up having done that work, and the sim will feel manageable. The pace will make sense. When the instructor throws something at you, you’ll have enough mental bandwidth to respond because you’re not spending it trying to remember where the fuel panel is.

Show up having skipped it (no lesson review, no chart study, no chair flying) and the sim will feel like it’s running about thirty seconds ahead of you the entire session. That feeling compounds. One rushed session bleeds into the next, and suddenly you’re behind in a way that’s hard to recover from.

The pilots who struggle in training are rarely the ones who lack the skills. They’re the ones who underestimated the preparation required and trusted that raw ability would carry them through. It usually doesn’t, not at this level.

Walk in ready

You’ve taken 8, 10, maybe 15 checkrides to get to this seat. You know how to prepare. You know how to perform under pressure. You’ve been evaluated by examiners who were looking for reasons to fail you, and you’re still here.

Initial Qualification is not the moment to doubt that record. It’s the moment to trust it.

Do the work before you walk in the door. Follow the curriculum. Show up ready for each session. And when you strap into that sim for the first time and the nerves arrive anyway (because they will) remind yourself that you have already done this. You just haven’t done it in this particular airplane yet.

That’s the only difference.

You Are No Longer Just a Builder, You Are a Pilot

I have often touted that aviation principles can be applied to AI systems. With the rise of AI agents, I couldn’t help but think about this overlap. More and more agents are being built and added to workflows, but how are we managing these systems? How do we ensure that we are maintaining operational control while managing a non-human task force? That’s where CRM comes in.

Crew Resource Management (CRM) is defined as the “application of human factors knowledge and skills to ensure that teams make effective use of all resources.” It was originally developed to reduce human error in aviation by optimizing the use of all available resources. Human, hardware, and information.

The theory behind this article, is to prove that you can use CRM to orchestrate your AI agents. While AI agents are not human, applying human factors knowledge is essential because the AI must interact safely and effectively with human operators.

Human factors engineering is the applied science of optimizing how people work together with machines. Over the past seventy years, this field has continuously evolved to address new technologies, shifting from physical ergonomics (like the layout of cockpit controls) to “cognitive ergonomics” as systems became highly computerized and automated. Integrating AI is considered the next major step in the evolution of human factors.

For a solo builder orchestrating multiple AI agents, you are no longer just executing tasks. You are the “Pilot in Command.” Your role has shifted from manual control to orchestrating a highly capable, non-human crew.

Subscribe now

The Hidden Traps of an AI Crew

Even though the AI itself lacks human qualities, human factors frameworks are required to manage the human side of the human-AI partnership. Specifically, this knowledge is applied to solve several core challenges in managing AI agents:

• Preventing Over-Trust and Complacency: Combating “automation bias” (over-trusting the machine) and complacency, where humans might “look but not see” contradictory evidence because they passively assume the AI is correct.

• Maintaining Situational Awareness: Because AI algorithms can act as “black boxes,” they risk confusing users or undermining their understanding of a dynamic situation. Human factors design ensures that AI agents provide "operational explainability," meaning the AI communicates its reasoning and confidence levels in a way that humans can quickly comprehend, keeping both the human and the AI on the same page.

• Mitigating the Risks of Anthropomorphism: Human factors experts warn against designing AI to needlessly mimic human emotions or deceive users into thinking they are dealing with a person. Falsely personifying an AI increases the danger that human operators will inappropriately surrender their authority, second-guess their own judgment, or delegate their safety responsibilities to the machine.

• Fighting Cognitive Degradation: Human factors research highlights that the more reliable a system is, the less practiced the human operator becomes, leaving them unable to safely intervene when the system eventually fails. This knowledge is critical for structuring workflows that force humans to maintain their own cognitive skills when managing AI.

Remember, the formal definition of CRM emphasizes the effective use of all available resources. This explicitly includes non-human systems like hardware and information.

Therefore, using human factors knowledge is not about treating the AI like a person, but about purposefully designing the AI’s behavior, communication, and autonomy to perfectly match human cognitive needs, capabilities, and limitations.

Step 1: Structuring Your Crew (Tiered Autonomy)

At its core, effective CRM means using all available resources and delegating tasks appropriately. Delegating tasks to an AI should be treated as a formal, visible operational decision. This delegation must be highly reversible (demotable) without friction or stigma if the AI begins to hallucinate or fail. You need to structure your agents based on their reliability and the task’s risk.

HAIF requires teams to systemically assess every candidate task based on four factors: structuredness, verifiability, consequence of error, and the AI’s demonstrated capability. Based on this assessment, the task is assigned to one of four autonomy tiers (or it is marked as “AI-restricted” if human-only judgment is required):

• Tier 1 (Assisted): The AI simply supports you (e.g., standard code autocomplete). You drive the work.

• Tier 2 (Supervised): The AI agent produces full outputs (e.g., an agent drafting a blog post or writing a script), but you mandate a 100% human review before it is deployed.

• Tier 3 (Autonomous-Monitored): The AI executes independently, and you only sample the outputs or manage exceptions.

• Tier 4 (Autonomous-Bounded): The AI executes independently within strict parameters. You only manage exceptions and conduct periodic audits.

Promoting an AI to a higher tier is a slow, evidence-based process requiring multiple successful cycles. If an agent starts hallucinating or failing, immediately “demote” it to a lower autonomy tier without hesitation.

Step 2: Standard Operating Procedures (SOPs) for Agents

In aviation, crews use checklists. For AI builders, these become “working agreements,” or pre-defined system prompts and procedures that establish exactly what the agent’s goals and boundaries are.

Effective CRM relies on the crew having a common understanding of the environment and the task. You must give your agents deep context so their “mental model” of the project aligns with yours. You are building shared mental models that align with your goals.

Program your agents to push back or ask clarifying questions. An effective AI teammate should mimic the CRM skill of waiting for the pilot’s acknowledgment before proceeding with destructive or complex actions.

This can be done by building them directly into the agent’s core instructions beforehand. For example, I have a Claude project for this blog that I would place into the Tier 1 category. The project is there to provide assistance with outlining and structuring my articles. Here is a sample taken from my project instructions:

How to work with the author:
Ask Socratic questions rather than providing direct answers. Surface connections and sources for Rich to explore independently. Never write blog content unless explicitly asked. End research sessions by asking Rich to summarize three things he learned.

These instructions ensure that I am always driving the work and that my “assistant” is there to question my thinking and push back when something isn’t clear.

Step 3: Quality Assurance and Human Ownership

Establish a strict rule for yourself: no matter how autonomous your agent chain is, you are the final accountable owner of the output. AI cannot hold accountability.

Stop estimating your work based purely on how fast the AI generates it. Budget dedicated time specifically for validation. Create strict personal checklists to verify accuracy and coherence for everything your agents build.

Another example of an agent that I use is my research agent that lives within Claude Cowork. This entire system was taken from Wyndo at . Instead of describing how it works here, I will encourage you to read it from the man himself.

The AI Maker

How I Run A Full-Blown AI Research Operation on My Phone (Powered by Claude Cowork)

A few months ago, I built an AI agent that sends me AI news summaries every week. Perplexity searches the internet. Make.com orchestrates the pipeline. OpenAI writes the summary. Gmail delivers it. Set it and forget it…

Read more

5 months ago · 59 likes · Wyndo

This somewhere in between Tier 2 and Tier 3 as it runs autonomously. I then review and validate the output by checking the sources and diving into the research myself. The agent is doing the hard work (researching the topic and finding relevant articles), allowing me to focus on breaking down each piece of research and applying it to what I’m working on.

Step 4: Active Competence Maintenance

It is important to remember that while productivity and efficiency can increase with AI delegation, you must not lose your edge. To fight the “Dependency Trap,” implement a core HAIF principle: Active Competence Maintenance.

Schedule periodic “human-only” execution cycles where you build, code, or write entirely without AI assistance. This is not a punishment, but a vital calibration exercise to ensure you retain the expertise required to effectively supervise and validate your AI agents. This ensures that professionals maintain the baseline expertise required to effectively supervise and evaluate the AI over time.

For me, this could mean reading physical books as part of my research instead of outsourcing everything to my agent. For example, I read Skin in the Game by Nassim Nicholas Taleb as part of my research on my last post about AI accountability and ownership.

You can read about it here:

Mastering Human-Autonomy Teaming

While my use of AI agents is rudimentary compared to other AI adopters, using CRM for AI agents can be applied at any level. I can keep a tight leash on my agents as I only have a few. As users build and grow their AI workforce, it becomes clear why effective CRM can mean the difference between an optimized, safe system vs. one that is uncontained.

By maintaining strict validation protocols, clear communication loops, and tiered autonomy, you can scale your output with AI agents without sacrificing quality or your own foundational skills.

Adopting a CRM framework transforms you from an overwhelmed individual user into a highly resilient team leader so try it out and stay in control of your non-human companions.

My job as an airline pilot is to guide a 200,000 lb aircraft and the 194 passengers in the back safely to their destination. Every single time.

Many are going on a long-awaited (and well-deserved) vacation; others are visiting family or flying to meet an important client. I am also onboard, and I want to get there safely, too. When dealing with tough situations in the air, I am not only thinking of my passengers’ safety; I have the selfish motivation of getting back to my own family.

I have skin in the game. Flying is an incredibly safe and reliable form of transportation, but it remains unforgiving if poor decisions are made. As we move toward a future of increased automation, we have to ask: How will the introduction of AI change the way we think about responsibility when those systems have no skin in the game and cannot be held accountable?

Subscribe now

What ADM Actually Is

Aeronautical Decision Making (ADM) is the capacity for proper judgment formed through accumulated exposure to risk. It’s about recognizing what kind of situation you’re in before you make a choice at all.

A pilot with 10,000 hours and decades of experience sees a threat (a building thunderstorm or a subtle hydraulic malfunction) differently than a pilot with 200 hours.

In ground school, we are taught ADM as a framework. We use acronyms like DECIDE, 3P, and IMSAFE to structure our thinking. While valuable, these frameworks are not the same as judgment.

Source: CFI Notebook

ADM as Calibration

True judgment is calibration, and it is built over time. It comes from the moments where you were wrong: where the weather was worse than forecast, where a “simple” approach turned complicated fast, and where you personally bore the consequences of your mistakes.

That exposure doesn’t just add information to your brain; it rewires how you weigh risk. AI systems can be trained on the frameworks, but they cannot accumulate calibration. This becomes critical when a situation stops matching the manual.

Take Large Language Models (LLMs), for example. Give them a scenario from aviation history and they can perfectly recite the proper procedure because they have the data. They know exactly what the pilots did in that scenario to achieve a successful landing.

But what happens when multiple things go wrong simultaneously? What happens when something occurs that has never happened before?

Enter Taleb: The “Interventionista”

In his book Skin In The Game, Nassim Taleb introduces the term Interventionista: the expert advisor who makes recommendations without accountability. Bearing consequences doesn’t just motivate better decisions; it shapes what you perceive as risk in the first place.

The absence of skin in the game has both ethical and epistemological effects. Interventionistas don’t learn effectively because they are not the victims of their own mistakes. The same mechanism that transfers risk away from the decision-maker also impedes their learning.

AI doesn’t just lack ethics; it lacks the feedback loop that produces accurate judgment. An AI system in the cockpit is the purest Interventionista because it is not physically sitting in the seat that will bear the consequences of its logic. I am. And that changes everything.

Embodied Consequence

ADM is inherently human. It is developed by making tough decisions, bearing the weight of those outcomes, and reflecting on what they meant. This process has three components: embodied consequence, emotional encoding, and genuine reflection. Remove any one of them and you don’t just get worse judgment, you get something that looks like judgment but isn’t.

Embodied consequence means you had skin in the game. The decision had weight because you were physically on the aircraft. Emotional encoding means that weight didn’t just motivate you. It shaped what you perceive as risk before conscious analysis even begins. Damasio’s research on somatic markers shows that emotions aren’t noise in decision-making, they’re load-bearing. Pilots with thousands of hours don’t just know more, they feel threat differently. Reflection closes the loop. Not retraining on new data, but genuinely integrating what the experience meant.

AI systems have none of these. This isn't a limitation waiting to be engineered away; it is a structural reality. The thing that makes a pilot’s judgment trustworthy is inaccessible to a system without stakes.

The Lindy Effect in Aviation

Aviation safety is a “Lindy” system. It survives and improves by building layers of redundancy and “filters” for failure over decades. Automation has had to earn its “Lindy credibility” through investigations, accidents, and constant refinement.

We learned the hard way what happens when we trust automation before we understand its failure modes. Consider Air France 447. When pitot probe icing caused a loss of airspeed data, the autopilot disconnected and the aircraft reverted to “Alternate Law.”

The crew, struggling with high-altitude turbulence and a confusing technical failure, failed to recognize an aerodynamic stall. Because the automation’s limitations in that specific “edge case” weren’t fully internalized by the industry at the time, the aircraft remained stalled until it impacted the Atlantic.

It took years to implement the resulting regulation and training changes. Autopilot had been around for 50 years before that accident revealed a misunderstanding of its failure modes. AI, as we know it today, simply hasn’t been around long enough to have its “soul” tested by the Lindy Effect.

Where AI Belongs In The Flight Deck

AI excels at monitoring, pattern detection, workload reduction, and surfacing information. It can handle high-volume data and routine tasks, with no authority. The experienced pilot uses it the way they use any good tool. It informs judgment, never replaces it.

Monitoring and predicting systems failures before they happen. Calculating optimal routes based on weather systems, turbulence and fuel trends. Presenting the pilot with the right information at the right time based on minimal input. These are all ways I hope AI will eventually find their way into the flight deck.

The decisions that have consequences. The black swan events that require my experience and judgment. The human relations part. Leave them to me. I trust my own judgment before I trust any piece of technology.

Closing

While we are far from FAA approval for autonomous AI in commercial cockpits, we must think deeply about it now. This isn’t just a thought experiment; it’s a way to analyze how we currently make decisions and how we can improve our own risk management.

Flying is inherently human. It requires decisions made by people who carry the weight of their prior experiences. Because AI cannot suffer the consequences of its errors, it lacks the “evolutionary soul” required for true rationality in complex systems.

When I land after a tough flight and walk off the airplane, I have more experience to fall back on. It is experience tied to emotion—the strongest, most reliable kind of calibration there is.

How do you certify a system that’s changing?

Traditional certification is straightforward. You test a system exhaustively. You prove it safe. You certify it. The system doesn’t change. You monitor it in service, but the core behavior is fixed.

But what if the AI system keeps learning after deployment? What if it evolves based on new data it encounters in the real world? How do you certify something that’s not done changing?

The FAA’s answer to that is that you don’t. Not yet at least. Maybe not for years.

This post is part of the FAA Roadmap For AI Safety Assurance series. Over 8 weeks, I’m breaking down the seven guiding principles that will define how AI gets integrated into aviation safely.

Subscribe now

The Fourth Principle

Distinguish between the safety assurance methodology for learned AI implementations and learning AI implementations

Not all AI is created equal. The architecture of the AI system fundamentally determines whether aviation can certify it now or has to research it for years.

Most people think “AI is AI.” The roadmap is saying that is not true. The difference between learned and learning AI changes everything.

Learned AI: The Static Model That Can Be Certified

Learned AI is what most people think of when they think of modern AI.

It’s AI systems trained offline on historical data. The model is fixed at a specific point in time. Once deployed, it doesn’t learn or change from new operational data. Same input always produces the same output.

Here’s a concrete example: A predictive maintenance AI trained on 10,000 hours of engine data. The system learns to recognize degradation patterns in vibration signatures, temperature trends, fuel consumption. The model is frozen. It’s tested, validated, and certified. Deployed on aircraft.

In service, it analyzes new engine data and flags anomalies. But the algorithm doesn’t change. A year from now, it will behave exactly the way it does today. Ten years from now, same behavior. The system is static.

Why learned AI can be certified:

You can test it exhaustively before deployment. You create test scenarios that cover the operating environment. You validate it against historical data. You understand its behavior completely because it doesn’t change.

Once deployed and passed assurance, it’s accepted. In-service monitoring becomes part of continuous operational safety (COS). You’re watching for anomalies, not for the system to evolve.

The FAA is explicit:

“The safety assurance for a learned AI implementation can be performed as part of the system design and validation. Once completed, the AI implementation is accepted, and the in-service monitoring of the AI implementation is part of the continuous operational safety (COS) programme for the aircraft.”

You certify it once. You monitor it continuously. But you’re not constantly re-certifying it.

How updates work with learned AI:

The operator collects in-service data. The developer uses this data to retrain the model. A new version is created. That new version goes through safety assurance again. Only when the new version is certified is it deployed.

This process can take weeks or months. It’s slower than continuous learning. But it’s manageable. It’s safe.

Learning AI: The Dynamic System That Aviation Isn’t Ready For

Learning AI is the scary one. And the FAA is explicitly saying they’re not ready to certify it.

Learning AI systems continue to learn in the operational environment. The model evolves based on new data encountered during actual flights. You can’t test every possible state because the system keeps changing. Same input might produce different outputs as the system learns.

This creates a fundamental problem: How do you assure safety of something that’s constantly changing?

With learned AI, you certify a snapshot. With learning AI, you have to certify a process. A process that evolves.

The FAA captures this challenge in two critical quotes:

“A system that continues to learn in the operating environment must build its safety assurance into the operating environment or include safety assurance as part of the process of learning. Learning systems may necessitate new regulations to assure the continued safety of the evolving system, as for active monitoring of performance or recurrent certification.”

Translation: We don’t have a framework for this yet. It might require continuous recertification. We’re not ready.

And then:

“Learning AI implementations may adapt in a manner that degrades performance, ultimately weakening their original safety profile. Cases in which a system learns anomalous, ungeneralizable, or inaccurate information will require new mitigation strategies.”

This is the core fear: The system could learn its way to being unsafe.

Think about this scenario. An AI system trained to optimize fuel efficiency learns a pattern that saves fuel by reducing engine strain in a particular flight regime. It’s a legitimate efficiency gain. It deploys this “optimization.”

Over time, something changes. Weather patterns shift. Traffic patterns change. The optimization that worked in the training environment starts producing different results. The system adapts, learning new patterns. But these new patterns were never validated. They’ve never been tested against the full operational envelope.

Gradually, the system’s behavior diverges from what was originally assured. It’s learning. But it’s learning in a direction that degrades safety. How do you catch this? How do you stop it before it causes an accident?

Aviation doesn’t have answers to these questions yet. That’s why the FAA is deprioritizing learning AI.

The roadmap is explicit: Safety assurance for learning AI is a research problem, not a near-term deployment problem. The FAA expects learning AI safety assurance to remain in the discovery phase for more than three years. Focused development activities won’t begin until safety assurance methods are proven.

Why This Distinction Changes Everything

The difference between learned and learning AI isn’t just technical. It’s the difference between “we can certify this now” and “we need years of research.”

Here’s what this means for safety assurance methodology:

With learned AI, you need:

• Design assurance level (DAL) based on criticality

• Exhaustive testing against expected scenarios

• Validation that the model performs across the operational envelope

• In-service monitoring for anomalies

• Controlled updates with recertification

This is hard. But it’s solvable with existing frameworks adapted for machine learning.

With learning AI, you need all of the above, PLUS:

• Runtime assurance (continuous validation during operation)

• Anomaly detection (catching when learning goes wrong)

• Graceful degradation (the system safely reduces capability if learning degrades performance)

• Possibly continuous recertification

• New regulatory framework (doesn’t exist yet)

This is not solvable with current methods.

The strategic implication is clear: The FAA is not saying learning AI is impossible. They’re saying it requires research and new methods. Until then, learned AI is the gateway technology.

You gain experience with learned AI. You document your processes. You build institutional knowledge about certifying and deploying AI systems. When learning AI research matures, you already understand the domain.

This is how you sequence innovation safely in safety-critical systems.

What This Means Beyond Aviation

Most industries are deploying AI without thinking about whether it’s static or dynamic. Aviation is forcing clarity. And other domains desperately need this distinction.

In healthcare: A learned AI diagnostic system is trained on 100,000 patient records and frozen. It’s deployed. Hospitals can assure it works consistently. A learning AI diagnostic system adapts based on each new patient case. It evolves. How do you assure it? Hospitals are deploying both without this distinction.

In finance: A learned AI fraud detection system identifies patterns in historical fraud. It’s static. A learning AI fraud detection system adapts to new fraud techniques in real-time. It evolves. Banks want learning AI because fraud changes. But they can’t assure learning AI works safely, so they deploy it anyway.

In autonomous vehicles: Current systems are mostly learned AI. They’re trained on driving data and frozen. They’re being deployed with limited success. Future systems might be learning AI, adapting to new road conditions. But we don’t have safety assurance methods for learning AI in vehicles. Aviation is saying: research this first, deploy later.

The principle applies everywhere: Be honest about whether your AI is static or dynamic. Have different assurance frameworks for each. Don’t deploy learning AI without understanding the safety implications. Use learned AI to gain experience before moving to learning AI.

This is the lesson aviation is teaching the world: sequence adoption carefully. Distinguish between what you can assure now and what requires research. Build experience with the former before attempting the latter.

The Pattern Emerging

Four principles in, you’re seeing how this works:

Principle 1: Work within the existing ecosystem. Don’t reinvent governance.

Principle 2: Address safety assurance AND enhancement. Ask both questions.

Principle 3: Be ruthlessly clear about responsibility. Personification erodes accountability.

Principle 4: Understand the AI distinction. Static and dynamic require different approaches.

The through-line is clarity and sequencing. Clarity about frameworks. Clarity about goals. Clarity about responsibility. And now, clarity about the type of AI you’re deploying.

Each principle builds on the previous. And Principle 4 explains why the incremental approach (Principle 5) actually works.

Next Week

Post 6 drops next Friday: Principle 5 — “Take an Incremental Approach”

Now that you understand the distinction between learned and learning AI, you understand why incrementalism matters. You can’t jump straight to learning AI. But you can learn with learned AI, build experience, and then apply those lessons when learning AI research matures.

This principle explains the actual deployment strategy: how to introduce AI into aviation safely, step by step, with clear milestones and decision points.

What happens when an AI system makes a decision that causes an accident?

Who is responsible?

We have a habit of naming AI. Siri, Alexa, copilot, assistant. If you call the AI by name or say it “decided,” accountability becomes murky. If you’re clear it’s a tool that computed a recommendation, accountability is clear.

The distinction between treating AI as a named entity with agency versus treating it as an algorithm that produced output is everything in aviation. It’s the difference between clear responsibility and liability chaos.

The FAA’s third principle is about forcing that clarity.

This post is part of the FAA Roadmap For AI Safety Assurance series. Over 8 weeks, I’m breaking down the seven guiding principles that will define how AI gets integrated into aviation safely.

Subscribe now

The Third Principle

Treat AI as an algorithm or computer, not as a human

This sounds obvious until you realize how often we personify AI in practice. Companies are marketing “AI copilots,” “intelligent assistants,” systems that “think” or “decide.” The media loves it. It’s effective marketing.

The FAA is essentially shutting this down.

“With AI technologies, it is common for developers and the media to portray AI as machines that interact like humans. People frequently refer to simulated assistants by name, such as Siri or Alexa, personifying the software that is responding to prompts. While this may serve as an effective marketing tool, it is not conducive to assuring safe operation of these complex systems in aviation.”

Responsibility must be crystal clear, and personification obscures it. In aviation, when something goes wrong, someone has to be held accountable. Personification obscures that accountability.

Treating AI as an algorithm stops the science-fiction narrative from entering the cockpit and ensures that advanced automation remains exactly what it is: a tool that pilots manage.

Accountability Rests With The Creator, Not The Machine

Personification erodes safety by eroding responsibility.

“Personifying AI can erode safety by creating ambiguity on the assignment of responsibility for safe operation. As certain operations, traditionally accomplished by people, are instead accomplished by automation, responsibility shifts from the human operator to the system designer.”

When automation takes over a task that a human used to do, responsibility doesn't disappear. It shifts. The system designer and AI developer are the ones responsible for the AI meeting its requirements, not the AI itself.

Aviation operates on clear chains of responsibility. Every system has a designer responsible for its design. Every operator has responsibility for using the system correctly. Every pilot has responsibility for the final decision. This clarity is foundational to safety.

The FAA makes this explicit:

“The system designer must delineate the responsibilities that are assigned to human beings as compared to the requirements that are assigned to systems and tools and must do so in a manner consistent with applicable aviation regulatory requirements and international standards. The responsibility for systems to meet their requirements rests with the system designer and AI developer, not the AI itself.”

AI is not responsible. Ever.

If a system does something unexpected at a critical moment of flight, we can't shrug our shoulders and blame an autonomous "crewmember." This principle ensures that tech developers cannot pass the blame to the algorithm.

Manufacturers carry the ultimate responsibility for how their AI performs. They cannot write off anomalies as “the AI making a choice.” They must rigorously delineate what the human is responsible for and what the system is built to handle.

For pilots, this clarity means if the AI system does something we didn’t expect, we disconnect it and fly the airplane. We’re not negotiating with it. We’re not trying to understand its thinking. We’re removing an unreliable tool and reverting to manual control.

For manufacturers, this means the burden is on them to make sure their AI systems are transparent, predictable, and safe. Not because the AI deserves to be trusted, but because humans need to be able to manage it.

Understanding The Machine’s Strict Limits

Pilots are trained to understand automation at a level most people never think about. We learn the normal operation and the failure states. We learn how to detect when an automated system is doing something wrong. We learn to intervene decisively when it fails.

This training exists because automation is a tool, not a colleague.

The FAA captures this in one of the most important paragraphs in the roadmap:

“Aviation experience with complex automation and human factors has highlighted the importance for the human operator, the pilot, to have a solid understanding of the modes, operation, and malfunction of the automation. Personifying AI applications suggests that they have human-like capabilities and potentially unexpected behavior. This contributes to the false impression that the modes, operation, and malfunction would be that of a human, and that AI is an entity which can be responsible. While the normal operation may be intended to automate something that can be performed by a human, the modes and malfunctions are notably different. The safety of future operations depends on the pilot understanding that a system containing AI is just a system and not another human with whom they can reason or negotiate.”

If you treat AI like a human crewmember, you’ll expect it to fail like a human. You’ll expect it to explain itself. You’ll expect it to be flexible and adaptive.

An AI system’s failure modes are not human. You cannot negotiate with it. When it fails, it might fail in ways that don’t make sense. And if you’re expecting human-like behavior, you’ll be dangerously surprised when you get algorithmic behavior instead.

For the industry, this means AI interfaces must be designed to be predictable and transparent. Developers must design these systems so their operating modes and failure states are clear and manageable. The system must tell the pilot exactly what it’s doing and why.

This is how pilots stay safe: through perfect understanding of what the tool can and cannot do.

AI Cannot Be a Crew Member

The FAA roadmap explicitly avoids human-centric language.

AI cannot be a part of crew-resource management (CRM) and it cannot be considered a "copilot". Instead, it can perform autopilot functions and affect how a pilot performs their duties, but the AI itself is accountable for nothing.

“For these reasons, this roadmap avoids the use of human-centric language when referring to AI. For example, AI cannot be a part of crew-resource management (CRM) but can affect crew responsibilities. AI cannot be a copilot but can perform autopilot functions and affect how a pilot performs their duties. AI may have a degree of control authority over specific flight functions but is not accountable for anything; the designer and maintainer of the AI are accountable unless that responsibility has been allocated elsewhere by applicable law. This roadmap also refers to the safety assurance of AI as a responsibility of the designer, unless allocated elsewhere by applicable law, and avoids mentioning trust in AI.”

Crew Resource Management is a human practice. It’s about clear communication, cross-checking, and problem-solving among pilots who understand each other’s capabilities and limitations. You can do CRM with your first officer because you both have judgment, both can be held accountable, and both can adapt in real time.

You cannot do CRM with a flight computer. The computer doesn’t understand context. It doesn’t adapt to changing circumstances based on experience. It can’t tell you what it’s thinking. It can’t be reasoned with.

When AI arrives in the cockpit, it won’t be sitting in the right seat as a peer. It might manage complex flight profiles. It might handle data analysis and recommendations. But it’s a tool that the actual, accountable human crew manages. It’s not a new member of the crew.

This dictates how training and procedures will be written. The integration of AI won’t be treated as adding a new crewmember to the flight deck. It will be treated as an advanced tool that shifts how the actual human crew divides their workload and makes decisions.

What This Means Beyond Aviation

The tech industry loves personification. It’s effective marketing. An “intelligent assistant” sounds better than “a tool that processes inputs.” A “copilot” sells more than an “autopilot extension.”

But aviation shows why this is dangerous.

The same principle applies to any domain where wrong decisions have serious consequences.

In healthcare: Instead of “AI diagnoses patients,” think “AI output helps doctors diagnose.” Hospitals are deploying AI diagnostic systems without clear responsibility frameworks. When a diagnosis is missed, who’s liable? The hospital that deployed it? The vendor who built it? The doctor who ordered it? The AI? Clear responsibility assignment prevents disaster.

In finance: Instead of “the algorithm decided to deny credit,” think “the algorithm computes a risk score.” When credit is denied unfairly, who’s responsible? The algorithm? The company? The risk model developer? In aviation terms, we would say: the designer of the algorithm is responsible for ensuring it works fairly. Period.

In HR: Instead of “AI screens candidates,” think “AI highlights candidates matching criteria.” Companies are deploying hiring tools that inadvertently discriminate. Who’s accountable? If you’ve personified the AI, accountability dissolves. If you’ve been clear that the AI is a tool deployed by humans, accountability is clear: the company is responsible.

In all of these cases, someone needs to be held accountable. And that someone is always a human.

When you introduce AI into something important, be ruthlessly clear about who’s responsible. Don’t let marketing language obscure accountability. Don’t let science fiction narratives enter your operational environment. Make the AI a tool with clear boundaries, clear limitations, and clear responsibility.

Next Week

Post 5 drops next Friday: Principle 4 — “Differentiate Between Learned and Learning AI”

This is where the technical foundation becomes critical. Not all AI is created equal. Some AI can be tested, validated, and certified as a snapshot. Some AI requires continuous monitoring because it evolves in operation.

One morning I woke up, opened my Substack app, and started wondering: where are all these people actually from?

I fly all over the US for work. Staying in different parts of the country four days a week. I was interested to know which part of the world these people were that I was talking to. Have I been there?

Flying for a living has given me a different perspective on distance. When you cross multiple time zones in a day and have breakfast in one part of the country and dinner in another, it changes the way you look at the world.

I considered posting a simple note asking people to drop their location in the comments. Then I remembered I have access to tools I didn’t have before. So I built a website instead.

The entire thing took under two hours. Claude handled the planning and architecture. Lovable handled the execution. I’m not a developer. I’m a pilot who figured out how to build things with AI.

I’m not monetizing it. Have no hidden agenda. I just wanted a clean, visual way for writers to see how far their words are actually traveling.

This post isn’t a tutorial.

There are plenty of publications that will teach you how to use Claude or Lovable. This one is for the people who usually sit on the sidelines, watching everyone else post about building incredible things with AI, not knowing where to start.

This is me showing you the door.

This post is part of a new series called “Builder Logbook,” where I document projects that I am working on. Most of my efforts are going towards building tools for student pilots and are more complex than StackerMap.

Some will be simple and some will be more detailed, showing how I’m using Claude Code along with other tools to do things I never thought were possible.

Subscribe to follow along

The Democratization of Building

Not long ago, an idea like StackerMap would have stayed exactly that, an idea.

It would have lived in a notes app or the back of your mind during a long flight. You’d need a developer, a budget, a timeline. Most people don’t have those things. So most ideas die quietly.

What we have right now is essentially a developer in our pocket. We have the ability to use natural language to describe what we want to build and actually make it happen.

This is a significant shift in human creativity. The gap between imagination and execution has closed. If you can think it clearly enough to describe it, you can build it.

I’m a pilot. I spent years studying aviation, not computer science. I don’t write code professionally. But I woke up with an idea one morning and had a working app by lunch.

Think about what that means for people in any profession. A teacher with an idea for an educational tool. The nurse who sees a problem in patient communication. The farmer who wants to build something for his community. The kid in a small town who has always had something to say but never had a way to build it.

These people exist everywhere. They always have. What they didn’t have was access.

Now they do.

StackerMap is a small example. But I’m going to briefly show you how the website went from idea to product with a few simple prompts.

The world is about to get a lot more interesting.

The Build

Building something like this is as simple as typing out your idea.

I described what I wanted to Claude, it handed me a brief, I pasted that brief into Lovable, and I was underway. Here’s the exact prompt I started with. Notice how plain the language is:

I have a cool idea for my Substack. I want to build a website in lovable. The homepage is just a giant world map that fills the entire screen. It looks sleek, modern, mainly dark (black and blue) coloring. I want to invite other substackers to go to the page where they can click the country/state they live in and add their Substack handle. Once they have entered the information, a dot appears in that region. The idea is for the map to demonstrate how we are from all over the world. When you hover over a specific dot, you can see more details about which substackers are from that particular region. 

Claude returned a detailed brief with visual design specs, user flow, and technical recommendations. All I had to do was paste this brief directly into Lovable. I won’t reproduce the whole thing here, but within minutes I had a working foundation.

I'm on Claude Pro ($20/month). Lovable Pro runs $25/month. Both have free plans if you want to test the waters first.

Once the scaffolding is in place, you just chat with Lovable to add features. Here’s where it gets fun.

The Explore page, a newly added feature suggested by , lets you select your home country and see how far away each pin is, how many flights it would take to get there, and more. Here’s the prompt I used:

I want to add an "explore" page where users can get interesting information. They select their home country/state and see a list of stats showing how far each pin is from their location, how many flights it would take to get there, how long and more information

This feature took some back-and-forth on time zones and needed some fine-tuning, but it got there eventually.

I also wanted every writer on the map to have a chance to be discovered, not just pinned. So I added a subscribe button and recent posts to each profile hover:

Add more information about each substacker. Ideally, you hover over the pin, it shows their information, as well as their latest posts, giving people a chance to see what the substack is about and whether they should subscribe

Lovable nailed it on the first try. It even added a few extra details that made sense without being asked.

Notice how rudimentary the language is. I’m not using technical jargon or going into intricate detail, I’m using plain language to make things happen. Lovable does a good job, even with the most simple prompt.

As much as I’d love to impress you with technical wizardry, that’s not the point. The point is that I built something real, that people are using right now, with nothing but plain language and two AI tools.

If I can do it, so can you.

The Map So Far

27 Substackers. 13 countries. And it’s only been live for a few days.

Add your pin at stackermap.com

Here's who's already on it. A genuinely impressive group of writers worth your time:

Wondering About AI: Karen is a content strategist-turned-vibe-coder who documents the honest, unfiltered experience of building real AI-powered tools (Chrome extensions, SaaS, ML experiments), bugs and failures included.

Sorta Systematic: Tiffany is an operations consultant who helps small business leaders build people-centered systems and workflows that reduce friction, align teams, and turn operations into a genuine growth driver.

Trenton Ian Cook: Trenton is the creator of Mirror Field Operating System (MFOS), a system designed to preserve the human ability to pause momentum and examine what is real before commitment.

Everyday Ethical AI: Karen is a 30-year software industry veteran who applies her data and ML expertise to help readers cut through AI hype, use technology ethically and practically, and advocate for women's equity in STEM.

Code Like A Girl: A community-powered publication amplifying the voices of women and non-binary people in tech through personal stories, technical tutorials, and career insights, with a mission to make the industry more inclusive.

Kim Doyal: Kim is a 17-year digital marketing veteran who teaches entrepreneurs to build AI-powered tools and scale smarter by ditching the playbook and showing up authentically.

Philly's AI Pharmacist: Ryan is a clinical pharmacist and AI researcher who cuts through the hype to deliver evidence-backed analysis of AI governance, safety failures, and responsible AI adoption in hospitals and health systems.

Prompt-Led Product: Elena is a developer-turned-AI-PM who teaches product managers the technical frameworks and practical tools needed to lead decisively in the AI era, not just manage.

How to Boss AI: Anna is a People & Culture veteran who explores the human side of AI adoption, helping leaders build organizations that can adapt to rapid change without losing their humanity in the process.

TPM Breakdowns: Michi is a Director of Technical Program Management who shares first-principles frameworks, playbooks, and real career journeys to help current and aspiring TPMs build teams from scratch and scale their leadership impact.

Realmscape: Jennifer is a blockchain founder and former cryptography developer who makes quantum computing and cryptography genuinely accessible to anyone who thought they weren't a math person.

What’s Working Right Now: Rebecca is a writer who explores motherhood, creativity, and modern work through the lens of staying expansive in a world that constantly pushes you to shrink.

Baseline Zero: Roman is a security architect who started at the helpdesk who writes practical, trust-first breakdowns of cloud security, Zero-Trust architecture, AI threats, and the human psychology that makes or breaks it all.

Uncertainty, By Design: A business analyst who thinks out loud about AI, healthcare, and the future of work.

The Cunningly Good Marketer: Tricia is a 25-year agency veteran and serial entrepreneur who shares cross-industry marketing insights drawn from working with hundreds of businesses across B2B, B2C, digital, and analog channels.

Start Up To Grown Up: Heather is a 6-time business author and coach who uses neuroscience and behavioral economics to help exhausted small business owners stop surviving and start scaling sustainably and profitably.

Escape Strategist: Neema is a strategist and consultant who writes about purposeful work, personal freedom, and redefining what success actually means.

Dear Husband: A former successful medical doctor turned to a (sometimes) successful stay-at-home mom/housewife, after her husband kidnapped her to Sweden.

Cash & Cache: Two fintech and capital markets veterans who alternate between AI market research and no-code implementation guides to help business and product leaders turn AI strategy into actual results.

The AI Architect: Matija is a 20-year software veteran and independent AI consultant who shares no-hype field notes on what actually works when building real AI solutions for businesses.

Fairness Algorítmico en LatAm: Marcela is an accountant and data scientist who delivers biweekly real-world cases and actionable tools (checklists, templates) to help Latin American public sector teams govern AI fairly and reduce bias in government decision-making.

AI Superhero: Shmulik is an ML engineer who fills the gap between beginner tutorials and academic papers with honest, production-focused writing on BERT, LLM internals, coding agents, and experiments that sometimes fail.

RandomNiceStuff: Kyle is a Filipino polymath who curates a weekly surprise box of beauty, curiosity, and delight. Poetry, chess puzzles, color theory, memes, downloadable wallpapers, and whatever else caught his eye that week.

Product In Progress: Darshana is an Apple developer writing about product strategy, UX quirks, and AI in real life.

KhinT’s Collective: Khin produces interconnected thoughts in the name of the universe.

Add Your Pin

Here’s my invitation to you to add your pin and join the map. You might just discover that the writer you are collaborating with lives halfway across the world.

stackermap.com

There’s a question buried in how most companies approach AI: do we manage the risk, or do we create opportunity?

They approach it as one or the other. Either you’re playing defense by implementing guardrails and safety checks to make sure nothing goes wrong. Or you’re playing offensive by deploying AI aggressively to capture new capabilities and improve your competitive advantage.

The FAA’s second principle is saying “do both,” but with safety as the non-negotiable driver of how these new systems are introduced.

This post is part of the FAA Roadmap For AI Safety Assurance series. Over 8 weeks, I’m breaking down the seven guiding principles that will define how AI gets integrated into aviation safely.

Subscribe now

The Second Principle

Address safety within the regulatory framework.

AI must be addressed specifically within the existing regulatory framework. For us on the flight deck, this means the regulator isn’t getting distracted by the hype and marketing surrounding artificial intelligence; their sole mandate remains ensuring the safe flight of civil aircraft.

Here is a breakdown of the second principle.

Using AI to Make Aviation Safer

The FAA is clear about their scope:

“This roadmap is intended only to address safety assurance of AI in aviation, and selected applications where the introduction of AI has the potential to improve safety. This scope is aligned to FAA authority to promote safe flight of civil aircraft.”

This isn’t just about managing risk from AI. It’s about using AI to prevent things from breaking in the first place.

Predictive maintenance is the clearest example.

Right now, airlines and maintenance teams monitor aircraft based on time-based or condition-based schedules. You inspect the engines every X hours. You replace components based on design life limits. This works because aviation is conservative. We replace things before they fail and build redundancy into critical systems.

But what if you could predict failures before they happen? What if you could identify the exact component about to fail, the exact flight where it’s most likely to go wrong, and remove the aircraft from service for maintenance before anything breaks?

AI algorithms analyze engine data across thousands of hours of flight. The system learns what normal looks like. Then it learns what the precursor to failure looks like. When current data starts matching that precursor pattern, the algorithm flags it.

A pilot never sees a catastrophic failure because maintenance knew three flights ago that this engine was degrading.

Other applications include:

• Dispatch optimization: AI analyzes weather, traffic, fuel efficiency, crew fatigue, and aircraft condition to recommend flight plans that are safer and more efficient simultaneously

• Data analytics for accident precursor identification: Finding patterns in flight data that precede incidents, allowing the industry to mitigate risks before accidents happen

• Training enhancements: AI tutors that help pilots learn systems faster and retain knowledge better, thus indirectly improving safety by creating better-trained pilots

None of these require the aircraft to fly itself. They’re all advisory and informational. But they all make the system safer.

Turning Ethical Issues Into Safety Requirements

The FAA explicitly states it won’t address the broader societal impacts of AI that falls outside their authority. But if a societal or ethical issue directly impacts safety, then it becomes a safety assurance issue.

The example they use is bias in training data:

“This roadmap does not address societal aspects with the use of AI which are outside of the FAA’s authority. There may be some common considerations where the safety of an AI application can be impacted by biases in training data, such as a pilot-health monitoring system that works more effectively for some ethnicities than others. These issues are addressed within the scope of safety assurance, in that the designer of such a system must show that the system performs its function across the entire community of pilots without unfair advantages.”

Imagine an AI system designed to monitor pilot health. It’s trained on historical medical data and deployed to monitor the health of all pilots. But the training data skewed toward one demographic. The algorithm learned what “normal” and “abnormal” look like for that group. When it encounters a pilot from a different ethnicity, it doesn’t work as well. It misses warning signs. It flags false alarms.

In the tech world, this might be categorized as an “ethical issue” or a “societal concern.” A known limitation that you can document in the fine print and address after launch.

Things don’t work that way in aviation. The system cannot work 95% of the time. Developers have to prove it works reliably across the entire pilot community, without any hidden blind spots and biases.

What This Means for Pilots

If you’re concerned that AI is going to replace pilots or remove judgment from the cockpit, understand this principle.

The AI arriving first won’t be autonomous. It will be advisory. It will be analytical. It will help you understand what’s happening with the aircraft, help you plan the flight, help you identify problems before they become emergencies.

Your job becomes interpreting AI-generated insights and applying judgment. Which is exactly what you’re trained to do.

The predictive maintenance system flags a degradation in engine performance. You and the maintenance team decide whether to defer the flight, continue to the destination, or divert.

The dispatch optimization system recommends a routing that’s more fuel-efficient and avoids weather. You decide whether to accept it based on the rest of your knowledge.

Beyond Aviation: Why Your Industry Needs This Principle

If you’re in healthcare, finance, transportation, or anywhere near safety-critical systems, pay attention.

Most industries are still asking: “Should we deploy AI?” That’s only half the question. The other half, ”Can we use AI to improve safety or outcomes?” is equally important.

In healthcare: You might implement an AI diagnostic system. That’s mode 1. Ensure the system is accurate and doesn’t recommend harmful treatments. But mode 2 is equally important: Can this AI catch cancers earlier than human radiologists? Can it identify patients at risk of complications? If the AI doesn’t improve outcomes, why deploy it?

In finance: You might manage risk from algorithmic trading (mode 1). But you should also ask: Can AI detect fraud patterns humans miss? Can it identify market risks before they become crises? (mode 2)

In transportation: You might ensure autonomous systems are safer than human drivers (mode 1). But you should also ask: Can they reduce accidents caused by human error and fatigue? (mode 2)

The principle is the same everywhere: safety assurance AND safety enhancement. Risk management AND opportunity creation.

Next Week

Post 4 drops next Friday: “Avoid Personification: Why AI Isn’t a Crew Member”

It’s about clarity. When you introduce something as powerful as AI, you need absolute clarity about what it is, what it’s responsible for, and what remains human. Personification, talking about AI like it’s thinking or deciding, erodes that clarity.

I just finished reading the book “Skygods: The fall of PanAm.” What strikes me isn’t the glamour of their routes, it’s how their crashes drove critical safety improvements that became part of aviation’s foundation.

Flight 214 led to lightning protection standards. Flight 759 led to windshear detection systems. Flight 845 shaped Crew Resource Management protocols. Flight 103 changed security procedures.

One crash after another forced investigation followed by new regulation. These led to fundamental changes in how we think about safe flight.

PanAm went under, but they left behind something more valuable: specific safety frameworks and disciplined processes that became the backbone of modern aviation. That ecosystem was built on failure that led to relentless refinement. And the FAA’s first principle for AI integration is essentially saying: we’re not discarding that hard-won knowledge just because a new technology arrived.

This post is part of the FAA Roadmap For AI Safety Assurance series. Over 8 weeks, I’m breaking down the seven guiding principles that will define how AI gets integrated into aviation safely.

Subscribe now

The First Principle

Use existing civil aviation safety requirements, processes, and methods to introduce AI, except where they are found to be inadequate.

This is a relief coming from an airline pilot perspective. It means the regulator isn’t throwing out a century of aviation safety standards just because a shiny new technology has arrived on the scene. Instead, the FAA is saying: make AI fit aviation. Don't make aviation fit AI.

Why This Matters

The FAA is explicit about this in the roadmap:

“The general field of AI has developed terms and concepts within that field that are distinct from the terms and concepts in aviation. It is paramount to recognize that the tremendous safety record in aviation has been achieved through assignment of responsibilities, a disciplined systems engineering process, thorough testing and analysis, and management of any risks while in-service. Aviation is unique in this regard, and it is appropriate to place AI within the aviation context rather than aviation within an AI context.”

Aviation cannot afford to “move fast and break things.” As pilots, we inherently trust our aircraft because we know every single system is backed by a clear chain of responsibility and exhaustive testing.

Any AI introduced into the industry has to conform to our strict, existing safety culture, rather than the industry lowering its standards to accommodate the software.

What This Means Practically

For pilots and operators, this is straightforward: any AI system introduced into your aircraft will be held to the same standards as every other critical system. It won't get a free pass because it's "AI." It won't be treated differently just because it uses machine learning instead of traditional code.

As the roadmap states:

“AI is an enabling technology that interfaces with several significant safety assurance characteristics, such as the key factors shown in Figure 1.

This principle leads to several key conclusions. First, pre-existing requirements and methods for evaluating the safety implications of any system are addressed in FAA regulations, FAA policies, and FAA-accepted industry standards for system safety assessments. The existing regulations for aircraft systems and equipment (e.g., subpart F for 14 Code of Federal Regulation (CFR) parts 23, 25, 27, 29) are performance-based regulations that ensure any system performs its intended function and does not introduce an unacceptable hazard. Industry standards provide appropriate methods for determining the safety criticality of a particular subsystem or algorithm and designate a design assurance level.”

When a new piece of avionics is installed, we expect it to work exactly as advertised without introducing hidden dangers. This confirms that AI will be treated just like any other critical piece of equipment on the airplane.

Manufacturers can't just throw an algorithm into a flight control computer and hope it works. They must rigorously prove that the AI meets our established design assurance levels and won't create an unacceptable hazard during operations.

The Automation Precedent

Aviation has been managing highly complex automation for decades. We've built entire regulatory frameworks around human-automation integration.

The FAA makes this point explicitly:

“Additionally, aviation regulations and guidance already address automation and the role of the pilot and other crewmembers. While AI is frequently considered a tool to develop or provide more advanced levels of automation, there is already considerable experience in human factors design principles, evaluation, and training in the aviation context that should be applied. Regulations and guidance continue to improve with experience and as new automation capabilities become feasible.

Issues associated with human-automation integration should be addressed as human factors and automation issues, and not as AI issues, unless the use of AI introduces risks that might not be present with other types of automation.”

AI is just the next evolution of a tool we already know how to manage. It’s not some mysterious, independent entity that needs its own rulebook. It’s automation. We understand automation. We have standards for automation. We know how to train pilots to work with automation safely.

What changes is that developers must use those existing human factors principles to ensure AI interfaces are intuitive and that pilots can easily understand and manage them during high-workload situations. The responsibility doesn’t shift to the AI. It stays with the designer, the manufacturer, and the operator.

Why This Approach Wins

There’s a temptation in every industry when a new technology arrives to say: “the old rules don’t apply.” Aviation’s approach is the opposite. The FAA is saying: prove to us that our existing framework is inadequate before we build a new one.

This does two things simultaneously:

First, it protects safety. A century of aviation experience doesn’t disappear because machine learning exists. The design assurance levels and testing rigor exist because they work. They’ve been validated through real-world operations where failure has consequences.

Second, it accelerates adoption. Manufacturers know the playbook. They don’t have to wait for new regulations to be written. They can start work now using existing frameworks, with a clear path to certification. That’s why industry-initiated AI projects are already underway with the FAA.

What This Means for Other Industries

If you’re in healthcare, finance, transportation, or anywhere near safety-critical systems, watch what aviation does here.

You probably have an existing safety framework. Your instinct when AI arrives might be to scrap it and start fresh. Don’t.

Instead, ask: what in our existing framework is adequate? What actually needs to change? Where does AI introduce risks that our current standards don’t address?

Answer those questions before you rebuild governance from scratch. Aviation figured this out because they had to. The cost of getting it wrong is too high.

Next Week

Post 3 drops next Friday: “Focus on Safety Assurance AND Safety Enhancement”

It’s about the dual mission: ensuring AI itself is safe, while also using AI as a tool to make aviation safer. It’s the difference between managing risk and creating opportunity.

In August 2024, while everyone was arguing about ChatGPT and self-driving cars, the FAA released a 31-page document called the “Roadmap for Artificial Intelligence Safety Assurance.” It’s one of the clearest, most useful frameworks for safe AI adoption ever published by a government agency.

And almost nobody read it.

I know why. It’s policy language. It’s not flashy. It doesn’t promise autonomous aircraft or revolutionary cockpit automation. It just quietly lays out how aviation will safely integrate AI. This sounds boring until you realize that aviation has solved a problem everyone else is still struggling with: how do you accelerate learning without sacrificing the judgment that keeps people safe?

When AI fails in your phone, you restart it. When it fails at 35,000 feet, people can get hurt. This constraint produces clarity. The FAA couldn’t afford to move fast and break things, so they did something almost unprecedented: they thought carefully about it first.

They didn’t reinvent safety governance from scratch. They extended what already works.

Subscribe now

Why This Matters More Than You Think

Here’s what struck me studying this roadmap: the FAA understood something fundamental that most AI discourse misses.

They recognized that AI isn’t going to replace human judgment in safety-critical domains. Instead, AI can accelerate knowledge transfer and pattern recognition. But the actual judgment needed to know when to trust it and how to verify it, that stays human.

The roadmap establishes seven guiding principles for making this work.

These principles aren’t unique to aviation. If you’re in healthcare, deploying AI diagnostic systems face the identical problem. If you’re in finance, deploying learning algorithms in critical trading systems face the same question: how do you assure safety in something that’s evolving? If you’re anywhere near safety-critical systems, the FAA figured out the framework.

What I’m About to Walk You Through

I’m deep in Airbus type rating training right now, which means my normal publishing schedule has shifted. Instead of going dark for a month, I’m releasing a mini-series: eight weeks, one FAA guiding principle per week.

These posts are slightly shorter than usual, but they cover something that matters: how aviation is actually thinking about safe AI adoption. I’m breaking down the FAA’s seven guiding principles. Here’s what you’re getting:

For pilots and aviation professionals: You’re about to understand what’s actually arriving in your cockpit, when it’s arriving, and more importantly why the FAA is being deliberately cautious about it.

For people building AI systems: These principles solve problems you thought were governance nightmares. They’re also the template for how you actually get certified and trusted at scale.

For everyone else: The FAA cracked something that looks simple but is actually hard: how to extend existing safety frameworks to new technology without breaking them or blocking innovation. Your industry can steal this playbook.

Here’s the series structure:

• Week 1 (this post): Why this matters, why now

• Week 2: Principle 1 — Work Within the Aviation Ecosystem

• Week 3: Principle 2 — Focus on Safety Assurance AND Safety Enhancement

• Week 4: Principle 3 — Avoid Personification

• Week 5: Principle 4 — Differentiate Between Learned and Learning AI (this is the big one)

• Week 6: Principle 5 — Take an Incremental Approach

• Week 7: Principle 6 — Leverage the Safety Continuum

• Week 8: Principle 7 — Leverage Industry Consensus Standards

• Week 9: What the FAA got right, and why it matters beyond aviation

The Timing Question

The roadmap is officially a “living document,” which means it gets updated as technology changes and experience accumulates. But the principles themselves aren’t going to age. They’re not about specific AI architectures or the latest model capabilities. They’re about how you responsibly introduce powerful, unpredictable tools into systems where failure has consequences.

I’m writing about this now because I’m living the intersection between knowledge acceleration and judgment development right now. As someone who trained pilots, I’ve seen what happens when you pump knowledge into someone without giving them time to develop judgment. You get people who know things but can’t think. The FAA’s roadmap is essentially saying: we’re not going to let that happen with AI. We’re going to use AI to accelerate the knowledge part, but we’re structuring the entire approach around preserving and developing judgment.

What Changes After This Series

By the end of eight weeks, you’ll understand:

1. How aviation thinks about AI safety, which is different from how the tech industry thinks about it, and worth learning

2. Why learned AI and learning AI aren’t the same thing and why this distinction actually solves certification problems

3. Why the FAA is deliberately starting with lower-risk applications and what that teaches you about scaling safety

4. How to think about responsibility in AI systems. Spoiler: not by personifying the AI

You’ll also have a mental model you can apply to your own domain. If you’re in healthcare, finance, transportation, or anywhere near safety-critical systems, these principles are portable.

Next Week

Post 2 drops next week: “Work Within the Aviation Ecosystem: Why the FAA Said No to Reinventing Safety”

It’s about why the FAA didn’t create a parallel universe of AI-specific safety rules. Instead, they took 70 years of aviation safety experience and asked: how do we thread AI through existing processes?

The answer is simpler than you’d think. And it changes how you should think about governance in your domain.

This is part of a mini-series on the FAA’s AI Safety Roadmap. One principle per week for eight weeks. If you’re interested in how regulation and innovation actually intersect this is for you.

This system applies to FAA written exams (PPL–ATP) using only publicly available materials (FAA handbooks, FARs, public test prep). Never share proprietary training materials or company procedures with AI tools. Respect licensing agreements and intellectual property.

I’ve passed over a dozen FAA written exams. But for most people passing isn’t the problem, retention is.

You cram for weeks, score 90%+ on test day, then six months later struggle to explain basic systems to your instructor or during a checkride. The knowledge evaporates because most study methods optimize for short-term recall, not long-term understanding.

Aviation doesn’t reward memorization. It rewards application. The examiner doesn't ask you to recite weight and balance limits, they hand you an aircraft loaded heavier than you expected and ask if you'd fly it.

One of my favorite ways to use AI for study is to build blueprints and systems around what I am trying to learn. It’s great at mapping out a gameplan for you that takes your schedule and learning style into account, breaking down what you should be doing week by week.

So I went down the cognitive science rabbit hole. Peer-reviewed research on how experts actually build durable knowledge. Then I used Claude to build a study system around those principles.

The Core Problem: Recognition vs. Recall

Walk into any flight school and you’ll see the same pattern: Practice tests until you hit 90%, pass the written, forget most of it within weeks.

This works for passing tests. It fails at everything else.

The research is clear: testing produces substantially greater retention than studying, even without feedback, especially on delayed tests. But there’s a critical distinction between recognition memory (what multiple choice tests measure) and recall ability (what you need in the cockpit).

Most written test prep courses teach you to recognize correct answers. The examiner asks you to recall information under pressure. These aren’t the same skill.

The result? “Paper pilots” who ace the written but can’t explain why. Pilots who memorized minimums but don’t understand the underlying principles.

Before fixing it, you need to understand one fundamental truth: without reinforcement, people forget approximately 50-70% of new information within a day.

This is the Ebbinghaus Forgetting Curve. When you read about airspace regulations, your brain dumps most of it within 24 hours unless you actively prevent that decay.

The solution: optimal review schedule is within 1 hour, then 24 hours, then 1 week, then 1 month.

This is spaced repetition, the foundation of everything else.

Subscribe now

The Five Core Techniques

1. Read Actively, Not Passively (SQ3R Method)

Most pilots read study material like they’re checking a box. Open the study material, read straight through, move on. Retention is terrible because the reading is passive.

The SQ3R method transforms passive reading into active learning. It’s designed to enhance comprehension and retention when working with written material. Research with engineering students showed strong educational impact in comprehension, retention, and critical thinking.

Here’s how it works:

S - Survey (5-10 minutes before reading)

Before you read a chapter on weather theory, skim the material to get a feel for main topics and ideas. Look at:

• Headings and subheadings

• Bolded terms

• Diagrams and charts

• Chapter summary

This preview primes your brain for what’s coming and creates a mental framework to hang new information on.

Q - Question (Create purpose for reading)

Turn headings into questions before reading. This creates purpose because you’re looking for answers.

Chapter heading: “IFR Departure Procedures” Your question: “What are the requirements for an IFR departure?”

Chapter heading: “Turbine Engine Systems”

Your question: “How does a turbine engine differ from a piston engine?”

Write these questions down. You’re about to read with intention instead of just moving your eyes across words.

R - Read (Actively, with purpose)

Now read to answer your questions. Reading actively means reading to answer the questions raised, not passive reading without engaging.

As you read:

• Usually the first sentence of each paragraph states the main idea

• Take notes in your own words, not verbatim from the text

• Mark sections you don’t fully understand

R - Recite (Test yourself immediately)

Close the book. Recall the information without looking at the text. Summarize main points in your own words.

Try to answer the questions you wrote earlier. Out loud if possible. Writing is better.

Can’t remember? That’s the point. This recital step is related to benefits of retrieval (testing effect) in boosting long-term memory.

Struggle now = stronger memory later.

R - Review (Within 24 hours, then weekly)

It’s extremely important to conduct an overall review within 24 hours for maximum comprehension and memory.

The next day, quiz yourself again. A week later, do it again. This is where spaced repetition starts working.

AI Application: Use Claude to quiz you on the material. Upload your chapter notes and ask:

I'm studying [topic]. Here's my chapter notes: [paste notes]. Create 10 questions that test whether I actually understand the material, not just memorized it. Start with 3 recall questions (basic facts), then 4 application questions (how would you use this in a scenario), then 3 analysis questions (why does this work this way). Don't show me answers. After I respond, tell me which concepts I'm weakest on based on my answers.

The AI becomes your study partner, testing recall instead of just providing information.

2. Explain It Like You’re Teaching (Feynman Technique)

Reading comprehension gets information in. The Feynman Technique ensures you actually understand it.

Named after physicist Richard Feynman, the method is built on one principle: if you can’t explain something clearly and simply, you don’t understand it well enough.

Here’s the brutal truth about aviation knowledge: knowing the name of something doesn’t mean you understand it. You can memorize “pitot-static system” without understanding how it actually works.

The Four-Step Process

Step 1: Choose a concept and explain it like you’re teaching a child

Write everything you know about the concept as if explaining to someone who’s never seen an airplane.

Example: "Walk me through what happens when you lean the mixture during climb."

Write it out. Use simple language. No jargon.

Step 2: Identify your gaps

The gaps are areas where you get stuck or resort to complex terminology.

You write: “The fuel goes from the tanks to the engines...”

Wait. How? Through what? What happens if a line breaks? What’s the backup system?

These gaps are gold. They’re exactly what you need to study.

Step 3: Go back to source material

Return to source material to fill gaps, then repeat until concept is clear.

Open the aircraft systems manual. Find the fuel system diagram. Read the section you glossed over. Now you’re studying with purpose. Filling specific knowledge gaps instead of passive re-reading.

Step 4: Simplify and create analogies

Streamline your notes and explanation, clarifying until it seems obvious. Think of analogies that feel intuitive.

“The fuel system is like the plumbing in your house. The tanks are your water heater, the pumps are your pressure pump, the manifold is your main pipe that splits to different faucets...”

When the analogy works, you understand the system.

Why This Works

Teaching others requires recalling, presenting, and organizing information coherently. When the person doesn’t understand, you must reorganize and distill into simpler concepts.

This reorganization process forces you to identify knowledge gaps, reorganize thoughts, and improve long-term retention.

AI Application: This is where AI becomes incredibly powerful. Use Claude or ChatGPT as your student.

Prompt:

I'm going to explain [concept] to you. Pretend you're a student pilot with zero experience—you've never seen an aircraft. Interrupt me whenever I use jargon without explaining it. Ask 'how' and 'why' questions that force me to explain the mechanism, not just the name. After I'm done, tell me what gaps I left unfilled and what I explained well.

The AI will interrupt you. Challenge your explanations. Force you to simplify. It’s like having a flight instructor for every topic, available 24/7.

3. Test Yourself Before You Feel Ready (Active Recall)

Here’s where most study systems fail: they emphasize input (reading, watching videos, listening to lectures) over output (retrieval, testing, recall).

The research is overwhelming: testing produces substantially greater retention than studying, even without feedback, especially on delayed tests 2 days to 1 week later.

Even more striking: repeated testing enhanced retention by more than 100% compared to dropping items from further testing.

Reading the same chapter three times feels productive. Testing yourself feels harder. But the second method produces 2x better retention.

The Testing Effect

Retrieval increases elaboration of memory traces and multiplies retrieval routes.

Every time you successfully recall information, you strengthen the neural pathway to that memory. More importantly, you create multiple ways to access it.

How to Practice Active Recall

1. Generate your own questions

Don’t rely on commercial test prep. Create questions from your reading:

After studying weather theory:

• “What causes an inversion?”

• “How does a cold front differ from a warm front in terms of cloud formation?”

• “Why is freezing rain more dangerous than snow?”

2. Use free recall

Close the book. Set a timer for 10 minutes. Write everything you remember about the topic.

This is hard. Your brain will resist. Do it anyway.

Research shows initial retrieval success should be 75%+ for optimal benefits. If you’re getting 95%+ correct, the material is too easy. If you’re below 60%, review first, then test.

3. Space your practice tests

Don’t test yourself immediately after reading. Wait. Let yourself start to forget.

Review when material is foggy but not completely forgotten. This typically happens 1-2 days after first exposure.

That struggle to remember? That’s exactly what creates strong long-term memory.

AI Application:

Prompt:

I just studied [topic]. I want to test myself at the right difficulty level. Start with 5 easy recall questions. Based on how I answer, adjust difficulty. If I get 4/5+, move to harder questions that require application. If I get below 3/5, stick with recall. After each wrong answer, don't explain yet—ask me a follow-up question that hints at what I'm missing. Only explain after I've struggled.

The AI generates questions. You answer. Then it provides feedback and explanations.

Even better:

Based on my answers, identify which concepts I seem weakest on and suggest specific areas to review.

4. Mix Your Topics (Interleaving)

Blocked practice feels efficient: all weather Monday, all regulations Tuesday, all systems Wednesday.

Problem: repeated studying increased confidence but decreased actual performance on delayed tests.

The solution is interleaving: interleaving leads to better long-term retention and improved ability to transfer learned knowledge compared to blocked practice.

Instead of studying topics in isolation, mix them.

Traditional (Blocked):

• Monday: Weather (2 hours)

• Tuesday: Regulations (2 hours

• Wednesday: Systems (2 hours)

Interleaved:

• Monday: Weather → Regs → Systems → Weather → Regs (2 hours)

• Tuesday: Systems → Regs → Weather → Systems → Weather (2 hours)

Same total time. Dramatically better retention.

Why? Interleaving forces continual retrieval because each practice attempt differs from the last, preventing rote responses from short-term memory.

Important caveat: For completely new material, some initial blocked practice helps build foundation before interleaving. First exposure to meteorology? Spend one session just on weather. Then start mixing.

Students found interleaving harder in short term but performed significantly better on exams. This difficulty is the point—it forces real learning.

AI Application:

Prompt:

Create a 20-question practice test mixing [topic 1], [topic 2], and [topic 3]. Randomize so no two questions on the same topic appear consecutively. For each wrong answer, tell me which topic it covered and what concept I missed—don't explain the answer yet. After I finish, create a follow-up mini-test on just my weak topics, but interleave those too.

5. Draw Everything (Dual Coding)

Aviation is inherently visual. Fuel systems, electrical diagrams, weather patterns, airspace structures. These aren’t just words. They’re spatial, visual concepts.

Yet most pilots study them as text.

Dual coding theory, developed by Allan Paivio in 1971, shows why that’s inefficient: information stored in both visual and verbal systems has better chance of being retained and retrieved than memory in just one.

How the Brain Processes Information

The brain processes information through two separate channels: verbal system and imagery system.

When you read “the fuel flows from the wing tanks through the boost pumps to the engine manifold,” you’re only using the verbal channel.

When you draw the fuel system diagram from memory, you engage both channels. Reading a word that evokes a mental image creates memory traces in both visual and verbal subsystems.

The result: students who learn through combined visual and verbal methods score higher on tests and demonstrate better long-term retention.

Practical Application

For Every System You Study:

1. Read the text description

2. Study the diagram in the manual

3. Close the book and draw it from memory

4. Check your drawing against the manual

5. Identify what you missed

6. Draw it again

Your first drawing will be bad. That’s the point. The errors show you what you don’t actually understand.

For Regulations and Procedures:

Convert text into flowcharts.

FAR 91.175: “Unless otherwise authorized, when the approach procedure being used provides for and requires the use of a DH or MDA, no pilot may operate an aircraft below...”

Draw it as a decision tree:

• Are you on an approach? → Yes

• Does the approach have a DH/MDA? → Yes

• Have you reached it? → Yes

• Can you see the runway environment? → Decision point

Flowcharts force you to understand the logic, not just memorize words.

Key Implementation Rules

Effective dual coding uses clear diagrams, graphic organizers, and icons rather than complex photographs.

Don’t try to create beautiful art. Create functional diagrams that show relationships and processes.

Picture-picture pairs processed faster than word-word pairs in recognition tasks. Your brain is naturally better at visual processing.

AI Application:

This is one area where AI image generation falls short for technical diagrams. Instead, use AI to verify your understanding:

Prompt:

I just drew a diagram of [system]. Here's what I drew: [describe it in detail—components, connections, flow direction, feedback loops]. Check my understanding. Tell me: (1) What critical components or relationships did I miss? (2) What did I get right? (3) What did I oversimplify? Be specific about what's wrong, not just what's missing.

Then describe your drawing. The AI can check your understanding even without seeing the visual.

Better yet:

Create a blank fuel system diagram template with only the tank and engine marked. I’ll fill in the rest from memory.

The Complete Study System

Weeks 1-2: Foundation

Daily (90 minutes):

• SQ3R on new material (30 min)

• Draw diagrams from memory (30 min)

• Review yesterday’s material (30 min)

Study topics in blocks initially to build basic understanding, then switch to interleaved practice.

Weeks 3-4: Deep Practice

Daily (90 minutes):

• Interleaved practice tests (30 min)

• Feynman Technique on weak topics (30 min)

• Spaced review of Week 1-2 material (30 min)

Spaced Repetition Schedule:

• Review new material within 1 hour

• Second review: 24 hours

• Third: 3 days

• Fourth: 1 week

• Fifth: 2 weeks

Weeks 5-6: Test Prep

Now use commercial practice tests.

Process:

1. Take practice test section (20 questions, mixed topics)

2. For every wrong answer, use Feynman Technique on that concept

3. Draw relevant diagrams from memory

4. Wait 24 hours, test yourself on those concepts again

Goal: 75%+ correct first attempt, 90%+ after review cycle. Not because 90% passes, but because initial retrieval success should be 75%+ for optimal benefits. This ensures strong long-term memory.

Why This Takes Longer (And Why That’s Good)

I won’t lie: this system is slower initially. But the student using this system can explain the regulation, draw the approach profile showing obstacle clearance, and apply the principle to non-standard scenarios.

One passed a test. The other built expertise.

Spaced repetition shows up to 25% higher retention rates over 4+ weeks compared to massed practice. That compounds over a career.

The investment isn’t in passing one written exam. It’s in building a foundation for every rating afterward.

Start Simple

Don’t implement everything at once. Start with SQ3R. Apply it to your next study session. See how much more you retain 24 hours later.

Then add active recall. Then interleaving. Then Feynman. Then dual coding. Build the system incrementally.

The goal isn’t to pass a test. It’s to build knowledge that compounds over a career.

The written exam is just proof the system works.

Last week in The Automation Paradox, I argued that more automation demands more expertise, not less. That the real danger isn’t AI making mistakes, it’s us losing the ability to catch them.

When I’m at 35,000 feet, the plane is flying itself. Beautiful view. Quiet cockpit. Autopilot engaged.

But I’m not relaxing. I’m monitoring. Scanning. Verifying.

Because a sophisticated, semi-autonomous machine, whether it’s a jet or generative AI, can become dangerous fast if you mismanage it.

Here’s what pilots know that most AI users don’t:

You don’t need “better prompts.” You need Standard Operating Procedures.

Automation Dependency

In the cockpit, we have a term for when pilots trust autopilot so completely they forget how to fly the plane: automation dependency (also known as “children of the magenta line”). When the computer makes a mistake, the human freezes.

In 1997, American Airlines Capt. Warren Vanderburgh coined the term “children of the magenta line” in a training presentation meant to combat automation dependency among pilots who were increasingly accustomed to following the magenta course line on the flight display.

It’s happening right now with AI. People are copy-pasting code they don’t understand, trusting hallucinated facts, shipping outputs they never verified.

A developer trusting Claude’s code because “it looked right.” A freelancer copy-pasting ChatGPT’s project proposal without reading it. A shift worker using AI to generate his weekly schedule without prerequisites.

These aren’t edge cases anymore. And they follow the same pattern every time: someone trusted the output because the process of generating it felt like work. It wasn’t. Typing a prompt and hitting enter isn’t due diligence. It’s delegation without oversight.

The solution isn’t to stop using AI. It’s to build a system that keeps you in command. In aviation, that system is Standard Operating Procedures. And the backbone of every SOP is a checklist.

Subscribe now

My Reliability Protocol

Stage 1
□ Set the Persona 
□ The Brain Dump 
□ Define Your Constraints

Stage 2
□ The “Visual Check”
□ The Logic Audit
□ Mid-Air Correction

Stage 3
□ The Stress Test
□ The "Human" Read
□ The System Update

A checklist in aviation isn’t a grocery list. It’s a written algorithm designed to produce reliable outcomes in chaotic environments.

I use the same three-stage structure pilots use for every AI task. Whether I’m prototyping an app or drafting a high-stakes email.

Stage 1: Pre-Flight (The Setup)

We don’t touch the controls until we’ve loaded the flight plan. We know where we’re going before the engines start.

With AI, skipping this step is how you get generic, soulless output. My Pre-Flight ensures the AI understands not just the task, but the texture of what I need.

The checklist:

□ Set the Persona 

Instead of prompting “make me a website,” try “act as a UI/UX designer who favors Apple-style minimalism.” Give the AI a lens to think through.

□ The Brain Dump 

Paste your raw, unedited notes with typos and all. Tell the AI: “Here’s my stream of consciousness. Structure this into a feature list before we start building.”

□ Define Your Constraints

What you don’t want matters as much as what you do. “No corporate jargon. No cluttered UI. Black/white/navy only.” The more specific your “no” list, the sharper the output.

Example: Building a Shift Schedule Optimizer

Bad Prompt (No Pre-Flight): “Make me a weekly schedule”

Good Prompt (With Pre-Flight):

PERSONA: “Act as an elite productivity coach who specializes in optimizing schedules for rotating shift workers.”

BRAIN DUMP: “I work 3 days on, 2 off. Shifts are 12 hours. I have 3 side projects, need 7 hours sleep, want to work out 4x/week. Mornings = high energy. Evenings = brain dead. I never know which shifts I’ll get until 48 hours before.”

CONSTRAINTS: “Do NOT create a rigid daily plan. Do NOT assume consistent wake times. Do NOT schedule deep work after shifts. Focus on energy states, not clock time.”

Sample output:

Ninety seconds of setup. Ten times better output.

Stage 2: In-Flight (The Execution)

Once the plane is in the air, we don’t put our feet up. We monitor instruments. We make small corrections to keep the flight on course.

With AI, I treat it like a junior partner. I have the vision, but I need them to explain their plan before they execute it.

The checklist:

□ The “Visual Check”

Before asking for final output, ask for a description first. “Walk me through the structure. What does the interface look like?” If I can’t visualize it, I don’t let it build.

□ The Logic Audit

Force the AI to show its work. “Explain your reasoning. What rules are you using? What’s your decision tree for edge cases?” Like a math student, show me how you got there.

□ Mid-Air Correction

Don’t build the whole system in one prompt. Fly leg-by-leg. Landing page first. Stop. Verify. Email capture next. Stop. Verify. Each leg gets checked before the next one starts.

Example: The Visual Check in Action

Bad Prompt (No Visual Check): “Build me a weekly schedule that accounts for my rotating shifts and side projects”

Good Prompt (With Visual Check):

Prompt: "I need a weekly schedule system for rotating shifts. 
Before you build anything, describe the structure to me:
- What are the main components?
- How does it handle unpredictable shift patterns?
- What does the user interface look like?"

AI Response: "I'm envisioning a 3-tier system:
1. Energy State Tracker (high/medium/low - not time-based)
2. Weekly Theme Buckets (not daily tasks)
3. Flexible Time Blocks that adapt to when you're actually free"

Me: "Perfect. Build tier 1 first, then stop.

Built exactly what I needed in 45 minutes instead of 2+ hours of rework.

---

Example: Automating Flight Expense Reports

Bad Prompt (No Logic Audit): “Write a Python script that categorizes my expenses from credit card statements”

Good Prompt (With Logic Audit):

Prompt: "Before you write any code, explain your logic for 
categorizing expenses. Walk me through:
- What rules will you use to identify business vs. personal?
- How will you handle ambiguous items like 'restaurant charges'?
- What's your decision tree for edge cases?"

AI Response: "I'll categorize based on:
1. Vendor name matching (airlines = business)
2. Date/time context (expenses during work hours = business)
3. Location context (restaurants near airports = likely business)

Me: 'Good. Now show me how you'll handle 'Crew Meals' specifically.'

AI: 'I'll flag any restaurant charge within 2 hours of a flight 
as business expense unless explicitly marked otherwise.'

Me: 'Perfect. Build it.'

---

Example: Building a Meal Prep Automation

Bad Prompt (No Mid-Air Correction): “Build me a meal planning system that accounts for rotating schedules, generates shopping lists, tracks macros, and sends daily reminders.”

Good Prompt (With Mid-Air Correction):

LEG 1: Input form for schedule and energy levels → Test → Approve 
LEG 2: Meal matching logic → Spot bug: it schedules batch cooking on exhausted post-shift days → Fix mid-flight → Test → Approve 
LEG 3: Shopping list generator → Test → Approve 
LEG 4: Reminder system → Done 

Stage 3: Post-Flight (The Debrief)

A landing isn’t complete until the engines are shut down and the passengers are at the gate. With AI, “it runs” isn’t enough. It has to be right.

The checklist:

□ The Stress Test

If I built a tool, I try to break it. Wrong inputs, wrong sequence, edge cases. Does it handle chaos gracefully, or does it fall apart the moment a real user touches it?

□ The "Human" Read

If it’s writing, I read it out loud. Does it sound like a robot? Does it use words like “delve” or “tapestry” or “landscape”? Back for a rewrite.

□ The System Update

If the workflow worked, I save the prompt sequence to my SOP library. I don’t rely on memory to recreate what worked. I rely on the logbook. That’s how you build compounding returns with AI. Not by hoping you remember what worked last time, but by writing it down.

The ROI

This protocol sounds like extra work. It’s the opposite.

Before these checklists: 5-6 hours to prototype a functional app. Sixty percent of AI outputs needed major revision. Eight to ten hours a week burned having to redo every step.

After Pre-Flight / In-Flight / Post-Flight: 1-2 hours for the same prototype. Ten percent of outputs need revision. Fifteen-plus hours a week reclaimed.

That’s the ROI of Standard Operating Procedures.

Back to the Paradox

In The Automation Paradox, I made the case that automation doesn’t lower the bar for competence, it raises it. That more AI demands more human expertise, not less.

This checklist is how you operationalize that idea. It doesn’t slow you down. It keeps you in command of a system that will happily fly you into terrain if you stop paying attention.

The difference between people who get 10x value from AI and people who get generic garbage isn’t better prompts. It’s procedures. Systems that force you to think before the AI builds, verify while it works, and confirm after it’s done.

Pilots don’t fear autopilot. We respect it, we use it constantly, and we never once forget who’s responsible for where the plane lands.

AI is the most capable co-pilot most people have ever had access to. Use it like one. Stay in the left seat. Run your checklists. And never confuse the autopilot doing its job with you doing yours.

In 1931, a mechanical autopilot was certified to serve as “co-pilot” on a commercial aircraft for the first time. The pilots hated it. They feared it would turn aviators into button-pushers, eroding the manual flying skills that kept passengers alive.

They were right to worry, but for the wrong reasons.

By the 2000s, automation had made it “more and more unlikely that ordinary airline pilots will ever have to face a raw crisis in flight, but also more and more unlikely that they will be able to cope with such a crisis if one arises.” Air France Flight 447 crashed in 2009, killing 228 people, because the pilots couldn’t manually fly when autopilot disconnected. The technology meant to make flying safer had made the crew less capable of handling the moment when it failed.

That’s the automation paradox. And we’re repeating it with AI. Not in the flight deck, but in our daily cognitive work.

Subscribe now

The Deskilling Is Already Here

I catch myself doing it. I’ll reach for Claude before I’ve actually thought through a problem. I’ll let AI draft something before I’ve formed my own opinion on it. Every time I do that without engaging my own reasoning first, I’m choosing convenience over competence.

I’m not alone. Right now:

• Students are using ChatGPT to write essays without learning to construct arguments

• Developers are using Copilot to generate code without understanding what it does

• Professionals are using AI to summarize documents without developing synthesis skills

• Analysts are using AI to interpret data without learning statistical reasoning

The pattern is identical to what happened in aviation. We’re outsourcing cognitive tasks to automation, and our mental muscles are atrophying because of it.

Not “could you eventually figure it out.” Right now, today, without any AI assistance.

If the answer is no, you’re not using a tool. You’re building a dependency.

Why More Automation Demands More Expertise

This is the counterintuitive truth that took aviation 50 years and multiple fatal crashes to learn: more automation requires more human expertise, not less.

It seems backward. The whole pitch for automation (whether it’s 1930s autopilot or 2026 AI) is that it handles the hard stuff so humans can relax. That’s partially true, but it’s dangerously incomplete.

Automation handles routine tasks flawlessly. Autopilot flies straight and level better than any human. AI writes basic code faster than any junior developer. But automation fails catastrophically in edge cases. And edge cases are exactly when you need human expertise most.

I think about this every time I hand-fly after a long stretch on autopilot. There’s always a moment where your hands are relearning the feel of the aircraft. On a calm day, that half second is nothing. In an emergency, it’s everything.

AI follows the same pattern. It writes your code perfectly until it introduces a subtle bug you don’t catch because you didn’t write the logic yourself. It summarizes documents accurately until it misses critical context you didn’t notice because you didn’t read the source material.

The question isn’t whether AI will fail. It’s whether you’ll recognize the failure when it happens. And you can’t recognize failures in domains where you haven’t built foundational skills.

The lawyers who were sanctioned for submitting ChatGPT-fabricated case citations to a federal court? They couldn’t recognize the failure. That’s what dependency looks like. And those are the early warning signs. In aviation, we’d call them precursors. The small incidents that tell you exactly where the catastrophic failure is coming from.

What Aviation Figured Out

After decades of automation-related accidents, aviation developed a framework that I think applies directly to how we should be using AI.

Modern airline pilots must maintain manual flying proficiency through regular practice, even though we rarely need to hand-fly in normal operations. We intentionally practice hand flying during flights and spend time in simulators specifically practicing what to do when automation fails. Not because we expect it to fail often, but because when it does, the stakes don’t allow for a learning curve.

Four principles came out of this:

Recognize when automation is wrong. We learn the subtle signs that autopilot is doing something unexpected. Not just obvious failures, but subtle deviations from normal behavior. You develop a gut feel for when something’s off, and that only comes from deeply understanding what “normal” looks like. With AI, this means knowing your domain well enough that a wrong answer doesn’t just look plausible. It bothers you.

Understand how the system thinks. You can’t evaluate automation if you don’t understand how it makes decisions. We learn the logic trees of every automated system in the cockpit. What inputs does it use? What assumptions does it make? Where are its blind spots? The same applies to AI. If you don’t understand what an LLM is actually doing when it generates a response, you can’t meaningfully evaluate what it gives you.

Maintain your manual skills. Even in highly automated aircraft, we hand-fly regularly. During training, during portions of flights, during practice approaches. The skills have to stay sharp because you never know when you’ll need them. For AI users, this means deliberately doing the cognitive work yourself on a regular basis. That means writing without AI, analyzing without AI, solving problems without AI.

Train for degraded conditions. Simulator training deliberately puts us in scenarios where automation has failed. You have to make decisions with incomplete information and tools that aren’t helping. This is trained, not hoped for. With AI, you should be stress-testing yourself: Can I still think through this problem from scratch? Can I still write a coherent argument without a first draft from Claude?

Using AI Like a Pilot Uses Autopilot

Here’s the framework I keep coming back to in my own daily use:

Let it handle the routine. Autopilot excels at holding altitude and heading. AI excels at first drafts, research synthesis, and idea generation. Let it do those things. That’s what it’s good at and that’s where it saves you real time.

But stay in command. I know where the airplane is going. I’m ready to take over instantly. With AI, that means knowing what you’re trying to accomplish, verifying what it gives you, and being ready to do the work manually.

Build your skills deliberately. We practice manual flying even when we rarely need it. You should practice writing, coding, analyzing, and thinking. Even when AI could do it for you. Especially when AI could do it for you. The practice isn’t wasted. It’s insurance.

Know the limitations cold. Every autopilot has conditions where it doesn’t work well. Every AI has domains where it fails predictably. If you don’t know where the edges are, you won’t see the cliff until you’re over it.

I ask myself regularly: is AI making me better at this, or is it making me need AI to do this?

The Choice

Aviation had this same decision point in the 1930s: embrace automation blindly, or develop a framework for using it safely. Commercial aviation is now the safest form of transportation ever created, not because we rejected automation, but because we learned to pair it with human expertise.

You have the same choice with AI right now.

You can outsource the hard cognitive work and get faster in the short term. Or you can use AI as an amplifier while maintaining the skills and judgment that make the AI output actually useful.

The difference isn’t whether you use AI. It’s whether you maintain the expertise to know when it’s wrong.

Aviation learned this the hard way. We don’t have to.

The AI applications discussed here are for personal knowledge retention and study acceleration only. All aviation learning must be verified against official FAA sources. These tools supplement (never replace) human instruction and judgment. I maintain strict compliance with my employer’s AI policies, using these techniques exclusively for off-duty personal development.

As aviation professionals, we hold a higher standard of responsibility and that extends to how we use AI tools.

This isn’t theoretical. Every conversation with ChatGPT, Claude, or Gemini creates a data trail. Your prompts get stored, analyzed, and potentially used for model training. If you’re copying internal procedures, flight operations details, or anything from company systems, you’re creating real privacy and security exposure.

Think of AI like talking to a stranger in an airport terminal. You’d discuss general aviation topics, regulations, or flight training theory, but you wouldn’t show them your company iPad or share internal SOPs.

The same rule applies to AI: discuss only public knowledge, verify everything, and keep company information completely separate.

Here’s how to use AI safely without compromising what you’re trusted to protect.

Subscribe now

The Core Rule: Never Share Sensitive Material

Don’t upload or discuss:

• Company standard operating procedures (SOPs)

• Airport security procedures or diagrams

• Crew scheduling information or internal communications

• Specific tail numbers or aircraft registration details

• Proprietary training materials from your airline or flight school

• Personal logbook entries with identifying information

• Any data that could violate company policy or NDAs

Even if an AI provider claims your data is private, most company policies explicitly forbid sharing proprietary information with third-party AI tools.

Use general examples instead. If you want to practice a procedure, describe it generically: “a twin-engine turboprop” instead of your actual aircraft type. “A Class C airport in the Midwest” instead of your home base.

Pictures/Screenshots

Another common mistake is accidentally uploading a screenshot or photo with sensitive material into an AI tool. Here is a list of images you should never send to an AI app:

• Company iPads, Jeppesen charts, or proprietary approach plates

• Your logbook (even cropped, metadata and formatting can identify you)

• ForeFlight screenshots showing your route, airport, or aircraft type

• Cockpit instruments or displays from your actual aircraft

• Any physical documents from your airline or flight school

You might remember not to type sensitive information, but a quick photo of a company approach plate to ask a question feels harmless. It’s not. Image uploads often contain metadata (location, device type, timestamp) and the content itself becomes training data.

Instead, try describing what you’re seeing: “On an ILS approach plate, what does the lightning bolt symbol next to the minimums mean?” You could also use publicly available examples from faa.gov.

Turn Off Data Sharing Immediately

Many pilots don’t realize that AI platforms use your conversations to improve their models by default. That means your questions about company procedures, specific aircraft tail numbers, or airport operations could end up in training data that anyone might access through future AI queries.

Most major AI platforms allow you to opt out of having your conversations used for model training. Do this before your first aviation query.

ChatGPT (OpenAI)

1. Click your profile icon (bottom-left corner)

2. Select “Settings”

3. Go to “Data Controls”

4. Toggle off “Improve the model for everyone”

Claude (Anthropic)

1. Click your profile icon (bottom-left corner)

2. Select “Settings”

3. Go to “Privacy”

4. Verify conversations are not being used for training

Gemini (Google):

1. Click on settings in the bottom left corner

2. Select “Activity”

3. Go to “Gemini Apps Activity”

4. Toggle off “Save Gemini Apps Activity”

5. You can also delete past activity from this page

Important: Even with these settings disabled, treat every AI conversation as potentially visible to others. These settings reduce risk but don’t eliminate it.

Mobile App Privacy Settings

The settings are different on mobile and that’s where most pilots will actually use AI.

Between flights, during commutes, or in the crew lounge, you’re likely using your phone. Here’s what changes:

ChatGPT (iOS/Android):

• Tap your profile icon → Settings → Data Controls

• Toggle off “Improve the model for everyone”

Mobile-specific risk: The app may sync conversations across devices. If you use ChatGPT for work on desktop, keep aviation study in temporary chat mode on mobile to prevent mixing contexts.

Claude (iOS/Android):

• Tap profile → Settings → Privacy

• Verify training opt-out status (same as web)

Gemini (iOS/Android):

• Menu → Settings → Gemini Apps Activity

• Toggle off activity saving

Warning: Google’s mobile apps are more aggressive about data collection. Consider using Gemini in a mobile browser with temporary chat instead of the app.

Use Incognito/Temporary Chat Modes

Most AI platforms offer temporary chat modes that don’t save your conversation history:

• ChatGPT/Gemini: Enable “Temporary chat” before starting your conversation

• Claude: Turn on “Incognito mode” in the top right corner

These modes are useful when you’re working with information you don’t want stored long-term, though you should still follow the data sensitivity guidelines above.

ChatGPT (OpenAI):

Claude (Anthropic):

Gemini (Google):

Voice Mode Warning

Many GenAI applications offer voice mode. These have separate privacy implications.

Voice data may be processed differently than text and conversations might be stored for quality/safety monitoring, even with data sharing disabled. It’s easier to accidentally mention specific details (tail numbers, airports, names) when speaking naturally.

Only use voice mode in a temporary/incognito chat. Treat it like you’re on a recorded line and stick to theoretical/general questions, never operational specifics.

Voice is great for brainstorming or exploring concepts, but type out your actual study questions. Typing makes you more deliberate about what you’re sharing.

Create Dedicated Accounts

Keep your aviation AI study completely separate from work:

Use a personal email for your AI accounts. Never your company email address. This maintains clear separation between personal learning and professional duties.

Don’t mix contexts. If you use ChatGPT for multiple projects, create a separate account for aviation study. This prevents accidental crossover and makes it obvious that your aviation AI use is personal education, not operational decision-making.

Keep it clearly labeled. Name your Claude Projects things like “Personal Flight Training Study” so there’s zero ambiguity about what you’re doing.

Browser-level separation matters too. Use separate browser profiles for aviation AI study. Cookies and logins can cross-contaminate even with separate accounts. Firefox Container Tabs or Chrome profiles work well for keeping aviation study completely isolated.

What Your Company Policy Probably Says

Most airlines and flight schools have policies that either explicitly prohibit using AI tools for operational purposes or forbid sharing information with third-party services.

This is why framing your AI use as “personal study on your own time using publicly available FAA materials” is critical. You’re not using AI to perform your job. You’re using it to accelerate your personal understanding of publicly available aviation knowledge.

If You Made a Mistake

Don’t panic. Take these steps immediately:

1. Delete the conversation

• ChatGPT: Click the conversation → Delete

• Claude: Click “...” menu → Delete conversation

• Gemini: Settings → Gemini Apps Activity → Delete activity

2. Contact the provider (if truly sensitive)

• Use their support/privacy contact forms

• Request permanent deletion of specific conversation IDs

• Document your request with screenshots

3. Notify your company if required

• Check your company’s data breach/security policies

• If you uploaded genuinely proprietary info, you may be required to report it

• Better to self-report than be discovered later

4. Change your approach going forward

• Review this checklist again

• Set a phone reminder to verify incognito mode before AI use

• Consider this a learning moment, not a career-ender

Most beginner mistakes are relatively low-risk (asking about your specific airplane type, mentioning your home airport). Deletion plus changed habits is usually sufficient. The goal is preventing patterns of exposure, not achieving perfect operational security.

Final Checklist Before Using AI for Aviation

✅ Data sharing is turned off on all AI platforms (web and mobile)

✅ You’re using a personal email, not a company email

✅ You understand what information is off-limits

✅ You know never to upload screenshots of company materials or logbooks

✅ You’re using temporary/incognito mode for sensitive study topics

✅ You have separate browser profiles or accounts for aviation study

✅ You’re only using publicly available materials as sources

✅ You’ve confirmed (or assume) your company prohibits AI for operational use

If you can check all these boxes, you’re ready to start building your AI-enhanced study system safely.

Questions about a specific privacy scenario? Drop a comment or message me. I want to make sure you’re protected before you start experimenting with AI study tools.

The AI applications discussed here are for personal knowledge retention and study acceleration only. All aviation learning must be verified against official FAA sources. These tools supplement (never replace) human instruction and judgment. I maintain strict compliance with my employer's AI policies, using these techniques exclusively for off-duty personal development.

I recently did some digging on Reddit to get a feel for the general sentiment towards AI use in aviation. What I found was that pilots are overwhelmingly against it.

The most common argument I found had to do with hallucinations. These can be a real threat when dealing with specialized knowledge and where accuracy is paramount.

Everyone deserves to have their own opinion and I can definitely come up with some arguments in support of their stance. Professional pilots have worked incredibly hard to get to where they are, put in thousands of hours, and deserve to be proud of what they have accomplished.

Change can be difficult to accept when we have been doing things a certain way for so long. Understanding how Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) work can seem complex, but knowing what the limitations are when using AI tools is a crucial step towards using them safely and responsibly.

I understand the resistance. But blanket rejection isn't the answer. Instead of dismissing AI entirely, let's address four questions that can help pilots use this technology safely:

1. How do LLMs and RAG actually work?

2. How can we leverage this new technology in a safe environment?

3. What systems can we put in place to ensure accuracy in the output we are getting from AI?

4. What safety practices can we adopt to ensure that we are using AI safely?

How do LLMs and RAG actually work?

Before we talk about safe AI use in aviation, it helps to understand what you’re actually working with. Most of the resistance I’ve seen comes from misunderstanding what these tools are and aren’t.

Large Language Models (LLMs) are pattern recognition systems, not databases.

When you ask ChatGPT, Claude, or Gemini a question, it’s not looking up the answer in some authoritative aviation reference. It’s predicting what words should come next based on patterns it learned from millions of documents during training.

Think of it like this: An LLM has read thousands of pilot forums, FAA publications, flight training materials, and aviation websites. It’s incredibly good at recognizing patterns in how aviation concepts are explained. But it doesn’t “know” anything. It’s generating text that statistically resembles how humans write about aviation.

This is why hallucinations happen.

RAG (Retrieval-Augmented Generation) is how we fix this problem.

RAG stands for Retrieval-Augmented Generation. Instead of relying purely on the LLM’s training, RAG systems first search through documents you’ve provided, find relevant passages, and then use those passages to generate an answer.

Here’s the difference:

Without RAG (standard LLM):

• You: “What are Class D airspace weather minimums?”

• LLM: generates answer based on pattern recognition from training data

• Risk: Might mix up Class D with Class E, might cite old regulations, might confidently make something up

With RAG:

• You: “What are Class D airspace weather minimums?”

• System: searches through FAR/AIM you’ve uploaded, finds 14 CFR 91.155(c)

• System: reads the actual regulation text

• LLM: generates answer based on that specific passage

• Result: “Per 14 CFR 91.155(c), Class D airspace requires 3 SM visibility, 500 feet below, 1,000 feet above, 2,000 feet horizontal from clouds.”

The answer is grounded in the actual document, not just the model’s training.

The limitation you need to understand:

Even with RAG, LLMs can still mess up. They might pull the right regulation but misunderstand a clause. They might cite a section accurately but miss an exception two paragraphs later. They’re really good at finding relevant information quickly, but they don’t have judgment about what matters.

RAG dramatically improves accuracy, but it doesn’t make AI infallible.

How can we leverage this new technology in a safe environment?

The aviation industry has historically been slow to implement new technology (and for good reason).

GenAI is still in its infancy stage. The first LLM as we know it wasn’t released until 2022. AI still makes comical mistakes and we have the right to be skeptical.

With this in mind, AI has some incredible use cases for aviation that can be implemented right now, in low-stakes environments.

I’m not talking about typing “how do you fly an airplane” into ChatGPT.

I’m talking about building systems that aid in the study process and meet the following requirements:

1. Only provide answers referencing the exact source so that they are easily verifiable

2. Do not interfere with the development of aeronautical decision-making

Here’s what that looks like in practice. I use a Claude Project with a custom prompt that requires every answer to cite specific FAR/AIM sections. When I ask about VFR weather minimums in Class D airspace, Claude responds: “Per 14 CFR 91.155(c), you need 3 statute miles visibility, 500 feet below clouds, 1,000 feet above clouds, and 2,000 feet horizontal from clouds.”

The two-step process is simple: AI generates the answer with source → I open the FAR/AIM and verify both the regulation number and the content. I’m not just checking if Claude is right. I’m building the habit of knowing where to find answers and understanding why they matter.

Beyond flight training, there are other low-stakes applications worth exploring:

• Weather briefing summarization for learning: Have AI condense a lengthy TAF or area forecast into plain language as a study tool to help you understand weather products better. This is for building your skills in interpreting weather, not for actual flight planning.

• Logbook analytics: Use AI to track personal currency requirements or identify trends in your flight experience

• NOTAM translation practice: Convert dense NOTAM language into readable summaries as a learning exercise to improve your ability to parse NOTAMs quickly

The key is these are all knowledge-building exercises, not real-time operational decisions.

What AI Should NOT Replace

Let me be clear about where AI has no business being used:

• Real-time operational decisions: Any decision made in the context of an actual flight. Go/no-go calls, fuel planning minimums, whether to continue an approach

• Weather assessment for actual flights: If you can’t analyze METARs, TAFs, and area forecasts without AI assistance, you’re not ready to make flight decisions. Use AI to learn how to read weather products, then do the real analysis yourself.

• Emergency procedures: These must be muscle memory, not something you’re looking up

• In-cockpit use: AI has no place in the aircraft during flight operations. Use it for practice scenarios on the ground only.

What systems can we put in place to ensure accuracy in the output we are getting from AI?

Verification systems are what separate safe AI use from dangerous shortcuts.

The core verification protocol:

1. AI suggests an answer with source citation

2. You check that source in the official publication (FAR/AIM, ACS, POH)

3. You understand why the regulation or procedure exists

Example: Claude tells me Class B airspace typically extends from the surface to 10,000 feet MSL. I don’t just accept that. I pull up the sectional chart for my local Bravo and verify the actual altitude rings. Sometimes it’s to 8,000 feet, sometimes 12,500 feet. The exercise taught me that Bravo altitudes vary by location. Something I wouldn’t have internalized if I’d just trusted the AI.

Red flags that demand immediate verification:

• Any regulation citation (14 CFR, AIM references)

• Performance numbers (V-speeds, weight and balance calculations)

• Weather minimums for different airspace classes

• Radio phraseology or ATC procedures

Additional verification methods:

• Cross-reference multiple AI outputs: Ask the same question to Claude and ChatGPT, see if answers align

• Use AI with web search enabled: Tools like Claude can search current FAA publications in real-time

• Build custom knowledge bases: Upload authoritative documents (FAA handbooks, your aircraft’s POH) to create AI systems grounded in verified sources

• Instructor spot-checks: Have your CFI review your AI-generated study materials periodically

What safety practices can we adopt to ensure that we are using AI safely?

Treat any form of AI like a study partner who occasionally makes things up. Here’s how to stay safe:

1. Never provide an AI tool with sensitive material

Don’t upload internal company documents, airport security procedures, standard operating procedures (SOPs), crew scheduling information, or any personal information that could violate privacy or security protocols.

Why this matters: AI training data practices vary by provider. Some tools use your inputs to improve their models. Even if they claim not to, company policies often explicitly forbid sharing proprietary information with third-party AI tools.

What to do instead: Use general examples or create sanitized versions. If you want to practice company procedures, describe the scenario without using actual company SOPs or airport-specific security details.

2. Verify every AI output with official resources

This isn’t optional. Every single answer needs verification against authoritative sources.

The verification workflow:

• AI provides an answer with source citation

• You open the actual FAR/AIM, ACS, POH, or chart

• You confirm both the citation and the content are correct

• You understand the reasoning, not just the answer

3. Use AI as a study accelerator, not an authority

Your CFI is the authority. The FAA publications are the authority. The manufacturer’s POH is the authority.

AI is a smart study partner who needs fact-checking. As a CFI who trained many students, I’ve seen how personalized explanation accelerates understanding. AI can scale that approach, but it can’t replace the human element of knowing when a student truly gets it versus when they’re just repeating words.

Think of AI like a senior student pilot who’s really helpful but sometimes mixes up details. You’d listen to their explanation, appreciate the help, but you’d still verify everything before your checkride.

4. Document your learning process

Keep notes on what you verify and how you verified it. This builds good ADM habits for when stakes are higher.

Sample workflow:

• Question: “What are the VFR cruising altitudes?”

• AI answer: “Per 14 CFR 91.159, eastbound flights use odd thousands plus 500 feet, westbound use even thousands plus 500 feet, when more than 3,000 feet AGL”

• Your verification note: “Confirmed in FAR 91.159. Key detail: only applies above 3,000 AGL. Below that, any altitude is legal (though still need to comply with 91.119 minimum safe altitudes). Checked on sectional, terrain in my practice area means I’m usually above 3,000 AGL anyway.”

This documentation proves you’re not just copying AI answers. You’re building understanding.

Building Your Own Safe AI Study System

I’ve walked through the principles and safety protocols, but theory only gets you so far. Next week, I’ll show you exactly how I’ve built a safe AI study system for personal aviation learning.

We’ll cover:

• Setting up Claude Projects with custom prompts that enforce source citations

• Creating a RAG system with your own FAR/AIM and aviation documents

• Real examples of catching hallucinations through verification

• The exact workflow I use from question to verified answer

• Common mistakes I’ve made and how to avoid them

Conclusion

The aviation community’s skepticism about AI isn’t wrong. The technology is new, makes mistakes, and operates in a field where mistakes kill people.

But refusing to explore it safely means missing opportunities to accelerate learning in low-stakes environments. AI won’t replace good training. It won’t replace the judgment you build from experience and mistakes. It won’t replace your CFI or the hours you need to log.

What it can do is make the knowledge acquisition phase faster and more personalized, as long as you treat it with the same healthy skepticism you’d apply to any new tool.

The question isn’t whether AI is perfect (it’s not). The question is whether we can build systems that let us benefit from its strengths while protecting against its weaknesses.

I think we can.

Have you caught an AI making a dangerous mistake in aviation content? What was it? Drop a comment or send me a message. I want to hear what concerns I haven’t addressed.

The AI applications discussed here are for personal knowledge retention and study acceleration only. All aviation learning must be verified against official FAA sources. These tools supplement (never replace) human instruction and judgment. I maintain strict compliance with my employer’s AI policies, using these techniques exclusively for off-duty personal development.

I remember my days as a CFI, sitting in a ground school classroom, watching a student fumble through a paper FAR/AIM trying to find a single regulation while the clock ticked down during an expensive ground school lesson. Back then, AI felt like something out of a sci-fi movie.

I flight instructed for over two years. The ones who quit rarely failed because flying was hard. They quit because the knowledge acquisition phase killed their momentum before they ever developed the judgment that makes a good pilot. That friction, that gap between "I want to fly" and "I need to memorize where regulations live in a 1,000-page book," lost me more students than actual flight training ever did.

But I've been deep in AI experimentation for years now, and 2026 feels like the inflection point. AI isn't a 'nice-to-have' anymore. It's the baseline.

Why This Matters Now

According to Boeing’s Pilot and Technician Outlook, 660,000 new pilots will be needed to fly the global commercial aviation fleet over the next 20 years. The traditional training pipeline has obviously worked up until now, but the question is how can we leverage this new technology to make training more efficient and let student pilots spend more time focused on building aeronautical decision-making skills rather than having their head buried in a book for endless hours.

What Actually Changed in 2026

AI is shaping the way we learn and now provides an expedited path to mastering the ins and outs of the dense documents and manuals required during initial training.

Tools like US Aviation Academy’s AI Knowledge System demolished the friction that discouraged many students from progressing as fast as they would have liked. Instead of hunting through a dense manual for one regulation, students get instant, plain-language explanations that actually make sense.

Sporty’s ChatFAR, ChatCFI, and ChatDPE aren’t just answering questions, they’re teaching students to think like pilots instead of becoming world-class memorizers who freeze when the examiner asks “now what?”

The data backs this up: Platforms like AviatorPro are using AI-driven micro-learning matched to how human brains actually retain information, cutting ground training time by up to 40%. That’s not marketing fluff. That’s students soloing faster because they’re not burning through their motivation on regulatory scavenger hunts.

CFIs Became Mentors Again

Here’s what the skeptics miss:

When AI handles the data entry, maneuver recognition, and regulatory lookup, the human instructor can focus on what actually matters: leadership, aeronautical decision-making, and the behavioral traits that separate future captains from pilots who don’t make it far.

The students who succeed and make good airline captains don’t need to have perfect steep turns right from the beginning, they need good judgment. AI can’t teach that. But it can free up the 60% of an instructor’s time currently wasted on administrative paperwork so they can focus on cultivating it.

When a student freezes during an engine failure simulation, that’s a judgment problem, not a knowledge problem. They know the checklist. What they don’t know yet is how to manage task saturation, prioritize when everything feels urgent, and make a decision with incomplete information while the altimeter is unwinding.

With more time during initial training focused on building these judgment skills, students can move up the timeline and start building the foundation of what will become the most important element in their aviation career: Airmanship.

What Changed

Tools like Redbird’s GIFT system handle the grunt work autonomously. The “Virtual CFI” objectively scores maneuvers in real-time. No more instructor sitting in the sim writing notes while the student sits there wondering what they did wrong. The AXIS AI Debriefing Station auto-recognizes steep turns, slow flight, stalls, and spits out analysis instantly.

So instead of spending 20 minutes post-flight filling out forms, the CFI walks in with the data already processed and asks the question that actually matters:

The student who answers “I was just trying to make it look smooth” needs different coaching than the student who says “I saw we were high and fast, so I wanted more time to get stable before committing.”

Same maneuver. Same result on paper. Completely different decision-making process. That’s what the CFI is there to develop.

What This Looks Like in Practice

Before AI: Student asks, “What’s the difference between Part 91 and Part 135 weather minimums?” CFI stops what they’re doing, pulls up the regs, explains for the 12th time this week, student nods, forgets by next lesson.

After AI: Student asks their Claude Project, gets a clear explanation with examples, comes to the CFI and says, “So if I’m flying cargo under Part 135, I need to think about weather differently than when I was training. How does that change my go/no-go decision-making?”

Now we’re actually teaching.

The student did the knowledge work. The CFI develops the judgment that turns knowledge into airmanship.

The Resistance (And Why It’s Wrong)

Some instructors see AI as a threat. “If students can just ask ChatGPT, why do they need me?”

Because ChatGPT can’t read the micro-expressions when a student says “I’m comfortable with crosswinds” but their body language screams otherwise.

Because AI can’t recognize when a student is developing hazardous attitudes that will harm their progress long after they pass the checkride.

Because teaching someone to think like a pilot requires pattern recognition built from watching hundreds of students make the same mistakes you made, and knowing exactly which mental model will click for this person at this stage of training.

The instructors who resist AI will compete with YouTube and Sporty’s courses on knowledge delivery and lose.

The instructors who use AI to handle the knowledge layer will become what students actually need: mentors who develop judgment, not human Google replacements.

What Students Actually Remember

I guarantee you: Ten years into an airline career, no pilot remembers the ground lesson where their CFI explained cloud clearance requirements.

They remember the CFI who talked them through their first real IMC encounter. Who debriefed the cross-country where they made a bad fuel decision but landed safely and learned from it. Who helped them develop the self-awareness to recognize when fatigue was affecting their decision-making.

I can recall the exact moment I first flew into a cloud, as well as the instructor I was with and the immediate disorientation I felt, as if it was yesterday.

That’s the job AI is handing back to us.

AI handles the “what.”

Instructors develop the “now what.”

And for the first time in decades, we have the bandwidth to actually do it right.

The Bottom Line

If you’re a CFI and this scares you, I get it. Change is uncomfortable, especially in an industry built on tradition.

But ask yourself: Did you become an instructor to fill out paperwork and repeat FAR/AIM explanations, or to shape the next generation of pilots who’ll make good decisions when it matters?

AI doesn’t replace you. It finally lets you do the job you signed up for.

The students who get instructors who understand that difference will become better pilots, faster.

And the instructors who embrace it will stop burning out and start remembering why they loved teaching in the first place.

The New Training Reality

The 2026 ecosystem is producing outcome-based pilots who are actually ready for a technologically demanding sky. We’ve stopped just logging time and started certifying true competence.

For students entering training today, this means:

• Personalized learning paths that adapt to your weaknesses instead of forcing everyone through the same curriculum

• Agentic AI assistants that function like digital interns - managing schedules, researching regs, and organizing study materials while you focus on flying

• Immediate feedback loops that compress the learning cycle from days to minutes

• Knowledge validation that's actually honest - AI asks follow-up questions until you can explain the regulation in your own words, catches contradictions in your reasoning, and forces you to articulate "why" instead of just reciting "what" from the FAR/AIM.

For CFIs, it means reclaiming the parts of instruction that actually matter while offloading the bureaucratic overhead that burns people out of this career.

For the industry, it means we might actually train 660,000 pilots without compromising safety or producing “paper pilots” who can pass tests but can’t make decisions.

What This Means for You

If you’re training now: Leverage AI for knowledge acquisition ruthlessly. Build custom Claude Projects with the FAR/AIM, POH for your training aircraft, and your local procedures. Let AI compress the memorization phase so you have more mental bandwidth for scenario-based decision-making during actual flight time.

If you’re a CFI: Your value just went up. You’re no longer competing with YouTube videos and Sporty’s courses. You’re the human who develops judgment which AI fundamentally cannot teach. Embrace AI as the tool that frees you to focus on higher-order instruction.

If you're running a flight school: Schools that ban AI will watch students use it anyway. Secretly, inconsistently, without guidance. Smart schools are building official AI policies that standardize which tools students use, how CFIs verify AI-generated answers against FAA sources, and where the hard line sits between knowledge acquisition (AI-friendly) and judgment development (human-only). The schools that figure this out first will market faster completion times and better-prepared graduates.

If you’re considering aviation from tech: This is your moment. The industry needs people who understand both domains. The gap between “aviation people who don’t trust AI” and “AI people who don’t understand aviation regulations” is massive (and valuable).

If you’re considering aviation as a career and think this sounds overwhelming, the opposite is true. This is the first time in aviation history where the technology is actually aligned with how humans learn instead of fighting against it.

The barrier to entry just got lower. The quality of training just got higher. And the students who embrace these tools will get to airline cockpits faster and better prepared than any generation before them.

The question isn’t whether AI will transform flight training. In 2026, that question is already answered.

The question is whether you’ll use it to your advantage or let it pass you by.

Subscribe now

What a start to 2026! While many were celebrating the New Year, the aviation industry was busy proving that its future is inseparable from artificial intelligence. This week saw a fascinating convergence where aircraft engines are being repurposed to power AI data centers, while carriers like American Airlines are using generative AI to turn rigid travel filters into natural conversations.

Here is the breakdown of the key AI advances that reshaped the skies this week.

The Conversational Traveler & Private Marketplaces

The first week of 2026 confirmed that the traveler journey is moving away from manual searching toward agentic, conversational assistance.

• Natural Language Travel Search: On January 2, 2026, American Airlines’ vice president of digital customer experience, Sam Liyanage, highlighted how the airline is shifting away from “rigid filters” toward natural, conversational intent. Their GenAI-powered search tool allows travelers to use highly specific, complex prompts—such as “somewhere warm under $800”—blending budget, loyalty status, and timing into a single choice.

• AI-Driven Connection Protection: According to the sources, American Airlines has expanded its use of Connect Assist, an AI-driven flight hold system. The technology uses a complex algorithm to decide whether to hold a departing flight for late-arriving passengers, aiming to reduce missed connections while protecting the overall network. On average, these holds last about 10 minutes and are communicated to passengers via automated text messages.

• Private Jet Aggregation: On January 2, 2026, Mach2 launched an AI-powered Empty Leg Marketplace. The platform uses AI to automatically process and publish empty leg flight lists from operators, creating an intelligent, searchable marketplace for brokers and end-users without requiring manual data entry.

Operational Intelligence: From Weather to Power Generation

Aviation technology is no longer staying within the airframe; it is now providing the critical infrastructure needed to sustain the AI industry itself.

• Next-Gen AI Weather Models: On January 2, 2026, NOAA deployed a groundbreaking suite of AI-driven global weather prediction models. These include the AIGFS (Artificial Intelligence Global Forecast System), which uses 99.7% less computing power than traditional models, and the HGEFS (Hybrid-GEFS), which combines AI and physics-based systems to consistently outperform traditional ensembles.

• Repurposing Jet Engines for AI Data Centers: In a striking trend, the sources reveal that aviation companies are pivoting to solve the global energy crunch. On December 31, 2025, FTAI Aviation launched “FTAI Power” to convert surplus CFM56 engines into gas turbines for data centers. Similarly, on January 2, 2026, Boom Supersonic secured a $300 million funding round and a deal with Crusoe Energy to deploy 29 “Superpower” turbines—natural gas-fired units derived from jet engine technology—to provide off-grid power for AI workloads.

• African Digital Transformation: On January 2, 2026, Ethiopian Airlines Group partnered with Nucore Technologies to integrate advanced AI-driven back-office automation and analytics into their Agency Portal, supporting their strategy to become one of the world’s most competitive aviation groups by 2035.

Defense, Autonomy, and Safety Reckonings

The military sector continues to lead in mission-critical AI, even as regulators and manufacturers face new hardware challenges.

• Mid-Flight Mission Replanning: On January 1, 2026, Lockheed Martin’s Skunk Works demonstrated AI-driven contingency management software. During the live demo, an AI agent detected simulated fuel issues on a Stalker XE drone, generated updated mission options in seconds, and automatically reassigned tasks to a second drone once the human operator approved the new route.

• AI-Enhanced Maintenance Partnerships: On December 29, 2025, Lockheed Martin and MANTECH formalized a strategic agreement to upgrade U.S. military aircraft sustainment. The partnership will utilize AI for real-time performance monitoring and predictive maintenance to reduce downtime and extend the operational life of combat aircraft.

• Safety Alerts and Directives: On January 2, 2026, researchers from Honeywell and Harmon Eyes unveiled a Pilot State Monitoring system designed to track sleepiness or drowsiness in the flight deck and issue proactive alerts. Simultaneously, the FAA issued a critical airworthiness directive for GE90 engines (powering Boeing 777s) due to iron contamination in powder metal components, highlighting the industry’s focus on proactive failure prevention.

This week cemented a new reality: aviation is not just a consumer of AI, but a vital provider of the hardware and data logic that fuels the wider technological landscape. As the Global AI in Aviation Market heads toward a projected US$ 6.47 billion by 2033, the integration of these systems is moving from experimental pilots to foundational infrastructure.

Subscribe now

The AI applications discussed here are for personal knowledge retention and study acceleration only. All aviation learning must be verified against official FAA sources. These tools supplement (never replace) human instruction and judgment. I maintain strict compliance with my employer’s AI policies, using these techniques exclusively for off-duty personal development.

I was a flight instructor for two years, taught hundreds of students, and spent thousands of hours in aircraft and classrooms teaching students how to fly an airplane.

As I reflect back on that experience, I can’t help but imagine how it might have been different if I had AI at my disposal back then.

I kept all my old teaching materials: lesson plans, ground instruction outlines, briefing templates. They sat in a folder on my hard drive for years, untouched.

Then I got curious.

What if I fed all of this to AI? Not to replace instruction (I’m not teaching anymore) but to see how modern tools could have transformed the way I taught ground school.

Could AI solve the problems I struggled with as a CFI? The time crunch, the one-size-fits-all curriculum, the students who memorized but didn’t understand?

In this post I’m going to talk about how I would structure my teaching if I were still a flight instructor.

AI in Flight Training: What’s Actually Allowed?

The short answer: There’s no FAA regulation prohibiting AI use in ground instruction because the FAA regulates outcomes, not teaching tools.

The agency’s 2024 “Roadmap for Artificial Intelligence Safety Assurance” explicitly lists “training and training simulators” as approved AI application areas, but stops short of issuing CFI-specific guidance.

In practice, this means instructors can use AI tools (chatbots, adaptive courseware, question generators) as long as students still meet Part 61/141 knowledge standards, receive proper endorsements, and pass their checkrides.

The unwritten rule? AI can tutor, but the human CFI remains responsible for verifying accuracy, correcting AI hallucinations, and ensuring every required aeronautical knowledge area gets covered before signing off a student.

The Three-Question Test:

1. Does it align with current FAA regulations? (I keep AIM/FARs open)

2. Does it match my aircraft’s POH?

3. Would I sign off on this as a CFI?

How to Think About AI in Ground Instruction

Don’t think of AI as your replacement, think of it as your teaching multiplier.

AI can handle the content generation, the personalization, pattern identification, and adaptive sequencing.

You can focus on validation, safety judgment, reading the student, and real-time adaptation.

You focus on what humans do best (teaching, connecting, mentoring), AI handles repetitive cognitive work.

Here are seven ways that I would adapt my curriculum if I were still teaching, along with the exact prompts CFIs can use today.

The Complete System: 7 Teaching Methods

Method 1: Socratic Question Generation

The Problem: Creating questions that build understanding (not just test recall) is hard and time-consuming.

I wanted students to discover principles through guided questioning, but crafting good Socratic dialogues took 45 minutes per concept. I’d fall back on “What’s the regulation?” instead of “Why does the regulation exist?”

The Prompt:

Create a Socratic dialogue to help a student discover [concept]. 

Requirements:
- Start with observable facts or their existing knowledge
- Build through 5-7 questions that progressively deepen understanding
- Each answer should lead logically to the next question
- End with practical application
- Include 3 different question paths depending on whether the student initially focuses on [aspect A], [aspect B], or [aspect C]

Do not provide answers—only the questions that guide discovery.

Example - Density Altitude (Engine-Focused Path):

Q1: "You notice the engine seems 'lazy' on a hot day at high elevation. What's physically different about the air?"

Q2: "Right—less dense air. Your engine mixes fuel and air at a specific ratio. If there are fewer air molecules entering the cylinders, what happens to power output?"

Q3: "Exactly. Now think about your propeller—it generates thrust by accelerating air backwards. How does that less-dense air affect thrust?"

Q4: "So you've got reduced engine power AND reduced propeller efficiency. What does this mean for your takeoff roll?"

Q5: "Given this understanding, why do you think the POH has different takeoff distance charts for different density altitudes?"

Why This Works: Students who discover concepts retain them 3x longer than those who are told. What would have taken me 45 minutes to create would now take 5 minutes to generate and 5 minutes to validate.

Method 2: Adaptive Learning Pathways

The Problem: Every student moves through the syllabus at the same pace, regardless of mastery.

Student A aces aerodynamics but we still spend two sessions on it because that’s the syllabus. Meanwhile, Student A struggles with airspace but we have to move on to stay on schedule.

The Prompt:

Based on this student's quiz results [paste scores by topic], create a personalized 2-week study plan. 

For topics scoring >90%: Provide only a brief refresher (10 minutes max)
For topics scoring 70-89%: Generate 3-5 additional practice problems
For topics scoring <70%: Create a targeted 20-30 minute mini-lesson addressing the root misunderstanding, then generate 5 practice scenarios

Identify any prerequisite knowledge gaps that might be causing struggles in lower-scoring topics.

Why This Works: Students waste hours on mastered content. This approach accelerates through strengths and drills weaknesses.

Method 3: Scenario Complexity Scaling

The Problem: Scenarios are either too easy (student bored) or too hard (student overwhelmed).

I’d create one cross-country scenario and reuse it. My advanced student found it trivial while my struggling student felt overwhelmed. Creating five different versions manually? Hours I didn’t have.

The Prompt:

Take this cross-country planning scenario: [paste your base scenario]

Generate 5 versions at different difficulty levels:

Level 1 (Foundational): Perfect VFR conditions, simple route, all services available. Focus: basic planning skills.

Level 2 (Single Complication): Add ONE weather or operational consideration requiring basic decision-making.

Level 3 (Intermediate): Add TWO complications requiring integrated thinking (weather + fuel consideration).

Level 4 (Advanced): Add THREE complications requiring prioritization and contingency planning.

Level 5 (Commercial-Level): Multiple layers: weather, equipment, operational pressure, real-time decisions.

For each level, specify what decision-making skills the student is developing.

Example - KPTK to KLSE:

Level 1: CAVOK weather, light winds, direct route. Learning: Basic nav planning, weight & balance.

Level 3: 25kt headwind, destination MVFR with lowering ceilings. Must calculate groundspeed impact, select alternate, plan fuel for alternate. Learning: Go/no-go decisions, alternate selection.

Level 5: All Level 3 conditions PLUS VOR failure, passenger with medical appointment (schedule pressure), uncertain fuel availability. Must balance safety vs. external pressure, plan for uncertainties. Learning: ADM under pressure, professional decision-making.

Why This Works: Every student gets appropriately challenged. You can also generate lateral variations (same difficulty, different situation) so students can’t memorize scenarios.

Method 4: Misconception Identification

The Problem: Students repeat mistakes because you’re treating symptoms, not root causes.

A student makes an error. I explain. They nod. Next lesson: same mistake. I was addressing surface errors without identifying underlying misconceptions.

The Prompt:

This student explained [concept] as follows: [paste student's exact words]

Analyze this explanation for:
1. What they understand correctly
2. The specific misconception (not just "wrong" but WHY they think this)
3. The root cause (what prerequisite understanding is missing)

Then create:
1. A 5-minute teaching intervention that corrects without being condescending
2. An analogy that specifically addresses this misconception
3. A practice scenario that tests whether they've corrected their understanding

Example - Adverse Yaw:

Student: “Adverse yaw happens because the down aileron creates more drag, so the nose yaws the wrong direction.”

AI Analysis:

• ✓ Correct: Ailerons deflect asymmetrically, creates drag

• ✗ Misconception: Thinks it’s aileron deflection drag, missing induced drag from increased angle of attack

• Root Cause: Doesn’t understand lift/induced drag relationship

AI Correction: “You’re right about drag, but not from the aileron sticking into the wind. The down aileron increases that wing’s angle of attack. More AOA = more lift. But more lift ALWAYS creates more induced drag. It’s the increased lift creating drag at the wingtip—not the aileron itself.”

Analogy: "Imagine you are walking on pavement. You want to turn left. Your left foot is on clean concrete, but your right foot steps into deep, sticky mud (Induced Drag). Even though you are trying to turn left, that right leg gets pulled back by the mud, twisting your body to the right."

Why This Works: Surgically fixes the actual gap instead of re-explaining the entire concept.

Method 5: Multi-Modal Content Transformation

The Problem: Your verbal explanation doesn’t work for every learning style.

I’d explain verbally. Student confused. Draw on whiteboard. Still confused. Try an analogy. Finally clicks! But now they can’t review it because the whiteboard’s erased.

The Prompt:

I need to teach [concept]. Create a complete learning package:

1. TEXT: 100-150 word explanation at 10th grade reading level
2. VISUAL: Detailed description for a diagram I can sketch
3. ANALOGY: Non-aviation comparison anyone can understand
4. PROCEDURE: Step-by-step process or calculation method
5. WORKED EXAMPLE: Numerical problem with complete solution
6. PRACTICE: Three problems (easy, medium, hard) with answers
7. COMMON MISTAKES: What students get wrong and why

Example - Load Factor (Excerpts):

ANALOGY: “Imagine swinging a bucket of water in a circle. Tighter circle = harder the water presses against the bucket. That’s load factor. Your wings are the bucket, you/the aircraft are the water. Steeper bank = tighter circle = more load factor.”

WORKED EXAMPLE: “C172 weighs 2,300 lbs, stalls at 48 knots level. You enter 60° bank. Load factor = 1/cos(60°) = 2G. Stall speed increases by √2 = 1.41. New stall speed = 48 × 1.41 = 68 knots.”

Why This Works: Different students need different formats. Having all formats ready means no student is left behind.

Method 6: Adaptive Oral Exam Preparation

The Problem: Students can’t practice realistic oral exams 24/7 with you.

Traditional prep: static question banks. Students memorized answers but couldn’t handle follow-ups. I couldn’t be available at 10 PM before their checkride.

The System Prompt:

You are a designated pilot examiner conducting a Private Pilot oral exam following current ACS standards.

Your approach:
- When student answers correctly but superficially, ask "Can you explain why that matters?" or "What would happen if...?"
- When incorrect, don't say "wrong"—ask Socratic questions to guide them
- When complete/accurate, acknowledge and move to related or more complex topic
- Maintain professional but encouraging demeanor

After 10-12 questions, provide detailed debrief:
- Topics showing mastery
- Topics needing review
- Overall readiness assessment
- Specific study areas before checkride

Begin by asking about [specific topic or let student choose].

Why This Works: Students practice unlimited times, AI adjusts to their level, identifies gaps you might miss in a single session.

Advanced Variation:

This student struggles with airspace. Conduct a 15-minute focused oral only on airspace. Start easy (Class G basics) and progressively increase difficulty until you find their ceiling. Provide targeted feedback on exactly where understanding breaks down.

Method 7: Just-in-Time Knowledge Delivery

The Problem: Front-loaded ground school leads to massive forgetting.

You teach weather theory in Week 2. Student flies their first XC in Week 8. They’ve forgotten 60% of what you taught.

The Prompt:

Student flying lesson [number] tomorrow: [topic/maneuvers]. Generate:

1. PRE-FLIGHT STUDY GUIDE (15-20 min): Key concepts, what to observe during demo, common mistakes, aircraft-specific notes

2. IN-FLIGHT OBSERVATION CHECKLIST: What to watch, questions to think about, connections to previous lessons

3. POST-FLIGHT REFLECTION: Questions connecting experience to theory, "Now that you've felt it, explain why..." questions, practice problems, preview of next lesson

4. REAL-WORLD CONTEXT: One scenario where this skill matters

Example - Short Field Takeoffs (Excerpts):

PRE-FLIGHT: “Key concept: obstacle clearance angle vs. ground roll. C172: 10° flaps, rotate at 51 KIAS (vs. 55 normal). Why? 51 gives you Vx immediately. Observe: MORE right rudder needed at slower speed. Pitch attitude feels aggressive—trust the airspeed.”

POST-FLIGHT: “You rotated at 51 vs. normal 55. What did you notice about control responsiveness? Why ‘mushy’ at that speed? [Tests: slower speed = less airflow = less control effectiveness]

Scenario: 2,200’ mountain strip, trees at departure end, DA 8,500’. Using short field technique, you reach 51 KIAS but airplane feels mushy and isn’t accelerating. What’s your decision point?”

Why This Works: Students learn right before using (maximum relevance), practice in flight (experiential), reflect immediately after (consolidation).

What You Actually Need

Let’s cut through the complexity:

Required:

• Your existing materials (lesson plans, notes, scenarios)

• AI subscription ($20/month—Claude or ChatGPT)

• Willingness to validate output

• 90 days to fully implement

Not Required:

• Technical background

• Coding skills

• Expensive software

What Doesn’t Change:

You’re still teaching flying. You’re still making safety decisions. You’re still reading students, adapting in real-time, providing mentorship.

AI handles the stuff that drained your energy: creating five versions of the same scenario, being available at 10 PM for questions, identifying why the student who “gets it” in ground school freezes in the cockpit.

Start Here

Don’t implement all seven methods at once. You’ll get overwhelmed and quit.

Do this instead:

Tomorrow morning, before your first lesson, pick ONE concept you’re teaching this week. Could be weather, airspace, weight & balance—doesn’t matter.

Take 10 minutes. Use Method 1 (Socratic Questions). Generate the question sequence. Read through it.

Ask yourself: “Would I actually use these questions?”

If yes: Use them in your lesson. See what happens.

If no: Try Method 5 (Multi-Modal Content) instead.

That’s it. That’s the test.

One method. One concept. One lesson.

If it doesn’t save you at least 30 minutes or produce noticeably better student understanding, comment below. Tell me what went wrong.

If it works? Add a second method next week.

The Bottom Line

The instructors who teach more effectively, serve more students, and maintain quality will thrive.

AI won’t replace flight instructors. But flight instructors who use AI will replace those who don’t.

I can’t go back and re-teach my students from 2014-2016. But you’re still teaching. You still have that student who’s calling you at 9 PM three days before their checkride, confused about something you’ve explained six times.

The tools are here. The prompts work. You’ve got the validation protocol.

The only question is: what are you going to try tomorrow morning?

Subscribe now