I ask a room of forty adult learners to write me a prompt that summarises a sales report. Every hand goes up within two minutes. Then I ask the same room what they do when the summary comes back confidently wrong, right structure, right tone, one fabricated figure sitting quietly in paragraph three. Silence. And not the awkward kind where people know the answer and don't want to volunteer. The kind where nobody has ever been asked the question before.
I don't think this gets discussed enough, probably because it's less glamorous than whatever agentic thing is trending this month. But prompting has become genuinely easy, and that ease has done something subtle to how people relate to the output. Two years ago you had to work a bit to get a usable draft, and the friction kept you engaged with what came back. Now the draft arrives polished on the first try, and polished feels like correct. It very often isn't.
I'd love to lecture about this from a safe distance, except I do it too. Early in my teaching I'd run a pandas groupby in front of a class, get a clean-looking table, and sail on to the next slide without asking whether that aggregation even answered the question we'd posed. The table looked authoritative, so I treated it as authoritative. Nobody called me on it at the time; I caught it later reviewing my own materials, which was somehow worse. If I'm still capable of that after years of working with data, I really can't act surprised when someone does it after two hours with ChatGPT.
So when organisations come to me asking for prompt training, I've started gently suggesting they're about a year late to a problem that has mostly solved itself. What their staff are missing is something harder to name and much harder to teach: knowing what a wrong-but-plausible answer looks like in their own domain, and having the reflex to check before the thing goes into a client deck or a board paper. A finance analyst should be able to smell an implausible margin. An HR lead should notice when a policy summary has invented a clause. That kind of knowledge is domain-specific, which is exactly why generic AI training never builds it.
Exhibit · What generic AI training misses
Prompt training vs. verification training
Prompt training
Easy to demo
- Produces visible magic in ninety seconds
- Teaches tool use
- Gets applause
Verification training
Changes how people work
- Teaches spotting a plausible-but-wrong answer
- Domain-specific judgement
- Slower and less flattering to demo
I'll also admit why most one-day workshops skip this part, mine included for a while: it demos terribly. A live prompting exercise produces visible magic in ninety seconds and the room lights up. A verification exercise produces a slower, less flattering moment where someone realises they were about to sign off on something false. One of those gets applause. The other one changes how people work. For an embarrassingly long time I optimised for the applause.
These days I deliberately plant wrong-but-plausible outputs into exercises and let learners walk into them. It feels faintly unkind the first time you do it to a room of polite professionals. But the people who get caught once, safely, are the ones who develop the checking habit, and a few have told me so months later, which is more than I can say for most of what I ever taught from slides.
If you're scoping AI training for a team, one question is worth asking any provider: how much of the programme covers what to do when the model is wrong? If you get a shrug, you're paying for prompt literacy your staff would have picked up on their own by Christmas.
If you take one thing: don't budget for prompting. Budget for the unglamorous skill of catching plausible nonsense before it ships, that's the part nobody learns by accident.
