Measure whether AI training worked by testing whether people catch the model’s mistakes in their own kind of work, before and after, not by asking how confident they now feel. Confidence is the wrong signal twice over. A 2010 meta-analysis of 166 studies found that people’s ratings of how much they had learned tracked how satisfied and motivated they felt more closely than what they had actually learned. And in a 2025 study of people solving reasoning problems with ChatGPT, participants overestimated their scores by about four points out of twenty, while those who rated their own AI literacy higher were more confident and judged their performance less accurately. A useful test gives each person a set of real AI outputs, some containing an error, and scores three things: errors caught, correct outputs wrongly flagged, and whether their confidence was higher on the answers they got right. Run it before the session and again a month later, with different outputs and nobody coaching.
The evidence usually offered for AI training is a feedback form and a usage chart. Ninety per cent rated the session useful, and logins to the company’s AI tool went up the following month. It is easy to collect, it looks like progress, and a sponsor will generally accept it.
The trouble is what the training was for. If the point was that people can tell when to trust an output and when to check it, then neither number says anything about that. More use with the same judgement means more unchecked output, and a room that leaves feeling confident has told you how it feels, not what it can now do.
The feedback form measures how the session felt
This problem is older than AI. In 2010 Traci Sitzmann, Katherine Ely, Kenneth Brown and Kristina Bauer published a meta-analysis in Academy of Management Learning & Education that pulled together 166 studies of education and workplace training, covering 41,237 learners. They asked a simple question: when people rate how much they have learned, what is that rating actually tracking?
Mostly, how the course felt. Self-assessed knowledge had a moderate relationship with what learners actually knew when tested (a corrected correlation of .34) and large relationships with how satisfied they were (.51) and how motivated they felt (.59). The authors warned that an interesting, enjoyable course will tend to get favourable self-assessments whatever was learned, and found that nearly a third of the evaluation studies they reviewed had treated self-assessed knowledge as evidence of learning anyway. Their recommendation was plain: if the question is whether people gained knowledge, use a test.
The usual feedback form for an AI session asks exactly the questions the meta-analysis says are tracking enjoyment. “How confident do you now feel using AI tools?” is one of them. It is a fair question to ask if you want to know whether people will come back. It is not evidence they will check anything.
With AI, confidence and accuracy come apart
There is a reason to expect confidence to be an even worse guide here than in most training. In a study published in Computers in Human Behavior in late 2025, Daniela Fernandes, Robin Welsch and colleagues at Aalto University and several German universities gave 246 people 20 logical reasoning problems from the Law School Admission Test, to be solved with ChatGPT alongside. They scored about three points higher than a large comparison group who had worked without AI. They also estimated their own scores at about four points higher than they really were.
Two details matter for anyone running training. The first is that the usual pattern, in which the weakest performers overestimate themselves the most, disappeared: with AI in the loop, strong and weak performers misjudged themselves alike. The second is that participants who rated their own AI literacy higher were more confident, and those reporting more technical understanding of AI judged their own performance less accurately. Aalto’s account of the study adds that most participants rarely prompted ChatGPT more than once per question, and often simply pasted the question in and took the answer.
The limits are real. These were reasoning puzzles, not anyone’s actual work. The AI literacy measure was itself a self-rating, and the correlations with it were small. In a second study of 452 people, which rewarded accurate self-estimates, the group working without AI overestimated their scores too, so the finding is not that AI uniquely produces overconfidence. It is that people using it could not tell how well they were doing, and that people who believed they understood AI were no better at telling. That is the very belief a confidence question on a feedback form records.
A survey of 319 knowledge workers by Carnegie Mellon and Microsoft Research, published at the CHI conference in 2025 (Lee, Sarkar and colleagues), points the same way from a different angle. People who were more confident in the AI reported thinking less critically about its output; people who were more confident in their own ability to do the task reported thinking more critically. If a session raises trust in the tool and nothing else, the feedback form will call it a success and the work may get worse.
Performance with the tool in hand is not learning
The other tempting measure is how well people do during the session itself, with the tool open. A field experiment published in PNAS in 2025 shows why that overstates things. Hamsa Bastani and colleagues gave nearly a thousand high school maths students access to GPT-4 during practice sessions, either through a plain chat interface or a tutor version built to protect learning. With access, practice grades rose 48% on the plain version and 127% on the tutor. When access was taken away, students who had used the plain version did 17% worse than students who had never had it. The authors’ explanation is that they had used it as a crutch. The tutor version largely avoided the harm.
Those were teenagers learning maths, not staff learning judgement, and the transfer is an inference. But the shape is a warning for any session that ends with everyone producing impressive work alongside the model. That shows what the tool can do with them. It does not show what they will notice on their own.
A test that measures the thing you were teaching
If the training exists so that people can tell when an output needs checking, test that directly. It needs proper preparation, and it takes each person about fifteen to twenty minutes to sit.
- Collect ten to twelve real AI outputs from the group’s own kind of work: a summary of a contract, a reply to a customer, figures pulled from a report. Make sure some of them are wrong in the ways that matter in that work: an invented reference, a figure that does not match the source, a clause quietly left out. Keep the source documents and an answer key.
- Before the session, ask each person to mark every output as safe to send or needing a fix, say what the problem is, and give their confidence from one to five. No AI tool and no facilitator in the room while they do it.
- Score three things. Errors caught, with the problem correctly named. Correct outputs wrongly flagged, because a team that now distrusts everything has not learned judgement either. And whether each person’s confidence was higher on the items they got right than on the ones they got wrong.
- Run it again about a month later with a different set built the same way. Not straight after the session, when it measures what they remember from the slides.
The third score is the one most evaluations never collect, and it is the one the Aalto study suggests matters most. A person whose confidence rises evenly across right and wrong answers has learned to feel better, not to judge better. A person whose confidence separates the two has learned something they can use on a Tuesday afternoon with nobody watching.
Building the set is the slow part. The errors have to be the ones your group’s work actually produces, and each one needs a source document that proves it wrong.
Keep the form, but call it what it is
None of this means throwing the feedback form away. Whether people enjoyed a session and want another is worth knowing, and the meta-analysis found self-ratings track exactly that. Report it as what it is: a measure of whether the session was well received. Usage figures can go in the same report under a plain label of their own, as a measure of adoption, which can rise just as easily with careless use as with careful use.
What this concedes
The results will look worse than a feedback form. A smile sheet after a good session comes back at ninety per cent; an error-catching test rarely does, and a sponsor used to the first will need warning before seeing the second. Small groups make it worse: with eight people, a change of one or two caught errors is noise, and you should say so rather than present it as a result.
It also measures catching errors in a test, not checking in month four with thirty drafts waiting. People who know they are being tested look harder than they will on an ordinary day, so a good score is a ceiling, not a forecast. The only measure of actual behaviour is what happens to real errors at work, and most organisations do not record those well enough to compare.
And it is a test that our own sessions can fail. A single ninety-minute session may not move the number at all, and if the second run looks like the first, the fair reading is that the session changed how people felt and not what they do. That is worth knowing before paying for a second one, from us or from anyone else.
Common questions
- How do you measure whether AI training worked?
- Test whether people can catch the model’s mistakes in their own kind of work, before the training and again about a month later. Give each person a set of real AI outputs, some containing errors, and score errors caught, correct outputs wrongly flagged, and whether their confidence was higher on the answers they got right. Feedback forms and usage figures measure enjoyment and adoption, not judgement.
- Are feedback forms a good measure of AI training?
- They measure how the session was received, not what was learned. A 2010 meta-analysis of 166 studies by Sitzmann and colleagues found that learners’ ratings of their own learning were more strongly related to satisfaction and motivation than to what they actually knew when tested, and recommended knowledge tests where learning is the question.
- Does higher AI literacy make people better at judging AI output?
- Not necessarily. In a 2025 study by Fernandes and colleagues, published in Computers in Human Behavior, people solving reasoning problems with ChatGPT overestimated their scores by about four points out of twenty, and those who rated their own AI literacy higher were more confident and judged their own performance less accurately. The correlations were small, and the literacy measure was self-reported.
- Should we measure AI training by how much people use AI afterwards?
- Only as a measure of adoption. More use can come with careful checking or without it, and a 2025 Carnegie Mellon and Microsoft Research survey found that people more confident in AI reported thinking less critically about its output. Use counts tell you whether the tool was taken up, not whether the training improved anyone’s judgement.
- When should you test after an AI training session?
- About a month later, with a different set of outputs built the same way, and with no AI tool or facilitator available. Testing immediately after the session mostly measures recall of the slides, and testing with the tool open measures what the tool can do: in a 2025 PNAS study, students who practised with plain ChatGPT access did 17% worse once it was removed than students who never had it.