A principal can tell that AI training worked when a colleague does a real piece of work differently afterward and someone other than that colleague has seen it. Attendance sheets, satisfaction forms and "I feel more confident" answers would look much the same after a poor session, so they cannot settle the question. Look at the work.
Why can't attendance and confidence scores tell you whether AI training worked?
Each of them would read the same whether the training changed anything or not. A sign-in sheet proves a room was booked. A satisfaction form records the afternoon, not the following month. A confidence rating records a feeling, and a fluent chatbot is very good at producing that feeling in people who have stopped checking what it says.
Most of the public numbers have this shape. The EdWeek Research Center found that 42 percent of the US teachers it surveyed reported no AI training in winter 2025-26, down from almost 60 percent in October 2024, while 22 percent had attended multiple sessions and 9 percent had ongoing training, according to Education Week's May 2026 report. Those figures count what schools offered. The Center on Reinventing Public Education (CRPE) found that 86 percent of the 79 districts in its early-adopter database offered teacher professional development in 2025-26, up from 63 percent, in an analysis published by The 74 in November 2025. That is a count of offers too, and the database covers early adopters, not districts in general.
Then there are the feelings. New Zealand's Education Review Office (ERO) drew on almost 4,000 survey responses for its July 2026 report, "Ready or not". It found that close to three-quarters of teachers are confident using AI, and that more than half report concerns that current, unplanned use is hurting students' critical thinking. Teachers were confident and worried at once, and every figure is self-reported. The report gives no figure for how many staff had received training, so it cannot tell you whether training produced either feeling.
What does the research say about whether AI training changes practice?
Not much yet, and the people who looked say so. Ofsted's June 27, 2025 summary of research commissioned by the Department for Education drew on 21 interviews with leaders at schools, colleges and trusts that had used AI for at least 12 months. Ofsted said robust evidence of AI's impact on educational outcomes is limited, and that most of what exists is "explorative, short-term and in limited domains." The leaders' main reason for adopting AI was to reduce workload.
A better answer is coming. A pre-registered randomized trial in Greece, registered in February 2026, plans to follow 1,200 teachers through an eight-week program and compare training in overseeing AI with training in non-AI digital tools, measuring both productivity and how far teachers rely on AI recommendations. The registry lists no results, and I would not cite it as evidence of anything until it does. It is worth watching because it asks the question schools need answered: does the training change what a teacher does with the tool?
Until that evidence arrives, the best evidence a school leader will get is the evidence they collect about their own staff, their own tasks and their own students. That is less of a burden than it sounds, because you do not need a trial. You need to look.
What is the Seen Test?
The Seen Test is my suggested way of deciding whether a training session earned the label "successful." It is not a validated instrument.
The Seen Test is three questions I ask before a leadership team calls any AI training a success: What did someone see a colleague do afterward that they could not do before? Would we have the same evidence if the training had been poor? And what will we do with the answer? Training that passes all three has been evaluated. Training that fails any of them has only been reported.
The first question separates three kinds of evidence, and they are not equal. In order of strength:
- Observed. A person other than the trainee saw the work happen, or checked a record of it: a colleague sat in on a planning session, or a line manager read the revision history of a document.
- Produced. The work itself exists and can be compared with earlier work: a lesson plan, a parent email or a report comment from before the training and from after it.
- Asked. Staff said what they do now, in a survey or a conversation. This is the kind most often presented alone, and the weakest. It is still useful, and my piece on how to survey staff about AI covers how to get honest answers.
The second question is the one that does the damage. Ask of each piece of evidence whether it would have looked the same after a poor session. Attendance would. A four-out-of-five satisfaction score would. A teacher saying "I use it more now" would, because more use is not better use. A lesson plan that now records what the teacher corrected in the AI's draft would not, because only a session that taught checking produces that habit.
The third question keeps the exercise honest. If no outcome would change what you do next, you are collecting evidence for a file. Decide in advance what a good result and a poor result would each lead to: repeat the session, change it, stop it, or move the effort somewhere else.
What should you collect, and when?
Collect a small sample of everyday work before the training, again four weeks after it, and once more at the end of the term. Three snapshots of real work will tell you more than any form, and they take a few hours in total.
| When | What to collect | Kind of evidence | Good sign | Warning sign |
|---|---|---|---|---|
| Before the session | Five or six pieces of ordinary work from volunteers (a lesson plan, a parent email, a report comment), plus a one-minute written answer to "What do you check in an AI draft before you use it?" | Produced, asked | You have a baseline to compare against | The answers are blank or "nothing" |
| Four weeks after | The same kinds of work from the same people, plus one short observation or conversation about a task they did with AI | Produced, observed | Visible checking: corrections, rewrites, a note on what was changed | Work looks the same, or looks the same but faster |
| End of term | One decision or routine that changed, with a date and an owner | Observed | A named task is now done differently and would be done that way next month | Nothing stopped, and nothing replaced it |
Tell staff in advance what you will look at and why, and use work they already produce. The aim is to learn about the training, not to grade the teacher. If people suspect otherwise, you will get their best work and learn nothing.
What does this look like in a school?
Here is a hypothetical example, labeled as one. A secondary school runs a two-hour session on using AI for lesson planning. Attendance is full, and the feedback is warm. Before the session, the deputy head collects five lesson plans from volunteers and asks each of them what they check in an AI draft.
Four weeks later she collects the next lesson plan from the same five teachers. In three of them nothing visible has changed. In one, the plan now includes a short note on what the teacher corrected in the AI's first draft. One teacher has stopped using AI for planning, because the first draft misjudged her class and she found rewriting it slower than starting fresh.
The last result looks like a failure and may be the best of the five. That teacher exercised judgment about when the tool earns its place. The summary line that would normally have reported the session, "full attendance, positive feedback," would have hidden all three outcomes. What the deputy actually learned is that the session produced one change in practice, one considered refusal and three unchanged teachers, and she knows which part of the session to rework.
What I see in practice
I have trained more than 150,000 educators across more than 30 countries. The feedback form at the end of one of those sessions is the easiest evidence to collect and the least informative about what happens the following Monday. I do not say that as a criticism of the people who send them out. A form is what a busy school has time for.
The mistake I would expect most leadership teams to make is to treat the training day as the finish line and the evaluation as paperwork afterward. The better order runs the other way. Decide what you want to see in people's work, then design the session to produce it. My piece on what happens after the initial AI training sets out the Release Question, which asks what people will stop doing when training works. It pairs well with the Seen Test: one tells you what should change, the other tells you whether it did.
Where does this approach run out?
Looking at work can feel like surveillance, and a small sample can mislead. Both objections are fair, and neither is a reason to fall back on attendance.
On surveillance, keep the sample to volunteers, use work that already exists, and report patterns across the group, never a verdict on an individual. On the small sample, treat five pieces of work as a prompt for questions, not a measurement. If three of five look unchanged, your next step is a conversation about why, not a conclusion about the whole staff.
There is also a harder limit. Some benefits show up later or not on paper: a teacher who feels able to say "I don't trust that output" may not leave any trace in a lesson plan. A conversation can catch some of it. Some will stay invisible, and a leader should say so plainly in the report, not claim a certainty the evidence cannot support. For the wider question of what to show a board, see how to prove AI is having an impact, and for the difference between use and real change, how to measure AI maturity.
What should you do before the next training day?
Write down, before the session, the one piece of work you expect to look different afterward. Collect a baseline of that work from a handful of volunteers. Put a date on the calendar for the four-week check and name the person who will do it. Those three steps take less time than the session itself, and they are the difference between a training day and a training result.
Two related pieces help you plan what to look for: what AI skills staff need beyond basic training and whether your training builds capability or dependence. To see how a trained skill becomes a changed routine, what AI guidance teachers need, task by task shows what good guidance looks like for each job.
If your leadership team is working out what its AI training should change, this is the kind of work I support through AI training for schools. I also speak to school and system leaders and conference audiences on this theme, through my education keynotes.
Sources and further reading
- Heubeck, E. "More Schools Are Providing AI Training for Teachers. Is It Any Good?" Education Week, May 18, 2026, drawing on EdWeek Research Center surveys. Article.
- Dusseault, B., and Hurwitz, J. "New Data Shows More Districts Are Adopting AI but Still Need a Coherent Strategy." The 74, November 12, 2025, drawing on the Center on Reinventing Public Education's early-adopter district database. Article.
- Education Review Office (New Zealand). "Ready or not: How are schools responding to Artificial Intelligence?" July 2026. Report.
- Ofsted. "AI in education: how schools and further education colleges are making it work." Press release on research commissioned by the Department for Education, June 27, 2025. Release.
- Goulas, S., Megalokonomou, R., and Sotirakopoulos, P. "Does AI training increase teacher productivity?" AEA RCT Registry, trial AEARCTR-0017831, registered February 2026; no results listed. Registry entry.
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and an international keynote speaker who has trained more than 150,000 educators. More about Dan.
