Nobody Learns to Fly from a Manual

Why the way we train pilots is the way experienced professionals should learn AI

In summary:

• Most AI training fails experienced professionals because the format — webinars, prompt lists, demos — gives them nothing to anchor their judgement to.
• The evidence is compelling: over 200 studies with more than 10,000 participants confirm that simulation-based learning outperforms conventional instruction, especially in problem-solving and complex situations.
• In my doctoral research, people who practised in a simulation made roughly 7 times fewer errors and completed real tasks in less than half the time of those who only received lessons.
• The key finding: inside a well-designed simulation, anxiety stopped translating into errors. It doesn't disappear; it stops costing you.
• Your professional experience isn't an obstacle to learning AI — it's the raw material that makes the simulation, and AI itself, genuinely work.

Before I ever set foot in a classroom as a teacher, I trained as an aerospace engineer. And if there is one thing aviation understood decades before the rest of us, it is this: when the stakes are high and the situations are messy, you don't learn by reading about it. You learn by doing it somewhere safe first.

No airline hands a new pilot a manual, shows them a few videos, and sends them off with two hundred passengers. Pilots spend hours in simulators — environments built to feel consequential without being catastrophic. They fly into storms that aren't there. They lose engines that were never running. They make the mistakes, feel the mistakes, and correct the mistakes, long before any of it matters.

Now compare that with how most professionals are being asked to learn AI: a webinar, a list of prompts, perhaps a lunchtime demo of a chatbot. Then back to your desk, where you're somehow expected to know when to trust these tools, when to question them, and how to bring them into work you've spent twenty years learning to do well.

It's the equivalent of handing someone the flight manual and pointing at the runway. And when it doesn't work — when people feel more confused and more anxious after the training than before — we blame the people, not the method.

The problem isn't you. It's the format.

Here is something I've noticed over years of teaching AI to non-technical professionals: the people who struggle most with typical AI training are often the most experienced people in the room. Not because they can't learn — their careers prove otherwise — but because typical AI training gives them nothing to attach their experience to.

A prompt list is generic by design. A tool demo shows you what the software does, not what you should do with it. And so the person with two decades of judgement in banking, or retail, or operations sits through the session and comes out with the same uneasy feeling: I understand the words, but I don't know what this means for my work.

That feeling is worth taking seriously, because the research says something interesting about it. Working with AI is what learning scientists call an ill-structured problem: there's rarely one right answer, the situation shifts, and success depends on judgement, not memorised procedure. I made this argument in my own academic work years ago — traditional, linear course formats are simply not built to convey that kind of complexity. They're excellent at transmitting facts. They're poor at building the situational judgement that professionals actually need.

Simulations are built for exactly that gap.

What the evidence says

This isn't just an engineer's fondness for flight simulators talking. Large-scale reviews of simulation-based learning — pooling hundreds of studies across medicine, management, teacher education, and engineering — consistently find substantial learning gains compared with conventional instruction, with the strongest effects precisely where AI fluency lives: problem-solving, diagnosis, and managing unfolding situations (Chernikova et al., 2020; 2024).

And here's a detail I find particularly encouraging: the recent research suggests that what makes a simulation work is not visual realism or expensive technology. It's what researchers call functional correspondence — whether the simulation preserves the decisions, trade-offs, and judgement calls that matter in the real task. A well-designed scenario on your laptop can outperform an elaborate virtual-reality rig, if the decisions are the right ones.

I also have my own data to draw on. In my doctoral research at Concordia University, I compared how people learned a piece of unfamiliar software under different conditions: some received only lessons, some practised in a straightforward practice environment, and some practised inside a game-like simulation. When it came time to perform real tasks in the actual interface, the differences were striking. People who had only received lessons averaged over three errors per exercise and took nearly two minutes to finish. People who had trained in the simulation averaged close to zero errors and finished in under forty seconds.

One nuance from that study has stayed with me, because it's honest and it matters. The simulation group did not score best on the multiple-choice questions — the trivia, if you like. They scored best where it counted: doing the task, quickly and accurately. Simulation doesn't make you better at talking about a skill. It makes you better at the skill.

The anxiety question

There's one more finding from that research I want to share carefully, because I think it speaks directly to anyone who feels a knot in their stomach when AI comes up at work.

Across all participants in my study, higher anxiety went hand in hand with more errors. No surprise there — anxiety taxes performance; we've known this for a long time. But when I returned to my data recently with fresh eyes, something stood out: within the simulation environment, that penalty effectively disappeared. Anxious learners performed on par with calmer ones.

I want to be precise about what this does and doesn't mean. It was a single study, with around seventy participants, measuring general test anxiety rather than anxiety about AI specifically. It does not show that simulations make anxiety vanish — the feeling was presumably still there. What it suggests is subtler and, I think, more useful: a well-designed simulation removes the cost of anxiety. When mistakes carry no real consequences, the nervousness that normally degrades performance loses its grip on the outcome.

For experienced professionals who feel they're being watched, judged, or measured every time they touch an AI tool at work, that's not a small thing. It might be the whole thing.

Why this works even better when you bring experience

Here's where the argument turns in favour of exactly the people who feel most left behind.

A simulation is only as rich as the judgement you bring into it. Put a novice in a flight simulator and they'll learn the controls. Put an experienced pilot in one and they'll stress-test their entire decision-making repertoire. The environment is the same; the depth of learning is not.

The same is true when the simulation is a business scenario and the tools are AI. In my courses, participants step into the leadership team of Orvaux — a fictional family business of craftsmen whose beautiful handmade goods have earned devoted customers, but whose shops are quietly struggling: fewer visitors, rising costs, and a family that reads the same worrying numbers through completely different eyes. Their task isn't to memorise what machine learning is. It's to work out which of Orvaux's problems data can actually see, what a sensible AI solution would look like, and — crucially — where AI shouldn't be trusted to decide at all.

Notice what happens in that scenario. The retail manager draws on every difficult store conversation she's ever had. The finance professional reads the balance sheet the way only someone who has lived with one can. Their experience isn't an obstacle to learning AI — it's the raw material the whole exercise runs on. This is what I call Bring Your Own Life: the recognition that your professional and personal history is precisely what makes AI genuinely useful in your hands, and a simulation is the environment where that history gets to do its work.

My students tell me the same thing in their own words. What they remember and value, year after year, is the real-life scenarios — the moments when an abstract concept suddenly attached itself to a situation they recognised. And their most common request, by a wide margin? More hands-on work, not less.

An honest caveat

I'd be doing you a disservice if I pretended simulations were magic. They're not — they're engineering, and engineering can be done badly. In my own research, a small group of participants found the game format genuinely stressful, largely because of design choices (a countdown timer, forced restarts) that added pressure without adding learning. That taught me as much as the successes did: the safety of the environment isn't a by-product, it's a design requirement. A simulation for experienced professionals should feel like a cockpit with an instructor beside you, not an arcade machine with a leaderboard.

That is the standard I hold my own scenarios to. Consequential enough to matter. Safe enough to experiment. Rich enough that your experience has somewhere to go.

The runway is right there

If you've sat through AI training that left you cold, the problem was almost certainly not your age, your background, or your aptitude. It was that someone handed you the manual and called it flying.

There's a better way, and aviation has been quietly proving it for the better part of a century: give capable people a realistic situation, real tools, real decisions, and zero real consequences — then watch what twenty years of professional judgement can do.

That's the environment I build. And if you're curious what it feels like to fly one, I'd be glad to show you.

Juan Carlos Sanchez-Lozano is the founder of Scenaptic.AI. He holds a degree in aerospace engineering and a PhD in educational technology, and has trained professionals from organisations including Amazon, McKinsey, PwC, and Pfizer.

References

Chernikova, O., Heitzmann, N., Stadler, M., Holzberger, D., Seidel, T., & Fischer, F. (2020). Simulation-based learning in higher education: A meta-analysis. Review of Educational Research, 90(4), 499–541. https://doi.org/10.3102/0034654320933544

Chernikova, O., Holzberger, D., Heitzmann, N., Stadler, M., Seidel, T., & Fischer, F. (2024). Where salience goes beyond authenticity: A meta-analysis on simulation-based learning in higher education. Zeitschrift für Pädagogische Psychologie, 38(1–2), 15–25. https://doi.org/10.1024/1010-0652/a000357

Sánchez-Lozano, J. C. (2010). Distributed information resources and embodied cognition in software application training: Interaction patterns in online environments and digital games [Doctoral dissertation, Concordia University].

Sanchez-Lozano, J. C. (n.d.). Advanced skills acquisition in professional settings: Computer games and simulations as parallel support systems in e-learning [Unpublished manuscript]. Concordia University.

Sanchez-Lozano, J. C., & Schmid, R. (2007, April 9–13). Application of structural analysis to game-based learning [Paper presentation]. American Educational Research Association Annual Meeting, Chicago, IL, United States.