A leadership team can usually tell you how many staff attended AI training last term. Far fewer can tell you what those staff can now do that they could not do before. Almost none can tell you what would still be standing if the tool were switched off on Monday morning.
That last question is the one that matters, and it is the one nobody asks.
The short answer: measure what survives the tool, not what people can do with it
AI training builds capability when the skill it leaves behind outlives the product it was taught on, and it builds dependence when it does not. That is the whole distinction, and it is testable. If the tool disappeared and your staff went back to working exactly as they did eighteen months ago, you did not build capability. You rented it, and the lease has just expired.
This is not an argument against training. It is an argument about what to ask for. Most AI training on the market teaches people to operate a product. Very little of it teaches people to judge the product's output, to decide when not to use it, or to pass what they have learned to the colleague at the next desk. Those three things are what capability is made of, and they are the three things a single session almost never produces.
Training went up sharply. Depth did not
The volume of AI training in schools has risen fast, and its shape has barely changed. Education Week reported in May 2026 that the share of teachers receiving no AI training at all fell from 60 percent in October 2024 to 42 percent by winter 2026, which is real progress. The more revealing figure sits beside it: only 9 percent were receiving ongoing training, and 22 percent had attended more than one session. Elizabeth Heubeck's reporting also found that schools concentrate on efficiency uses, with one principal admitting that moving staff beyond AI as an efficiency tool toward something they could build into teaching and learning had been the harder job.
Microsoft's 2026 AI in Education Report, published on June 24, 2026, puts the gap in sharper terms. It found that 88 percent of educators have already used AI for school-related purposes, while 53 percent say they have received no formal AI training. Two thirds of educators in that survey wanted their institution to provide training monthly or quarterly. Usage has run ahead of instruction, and the people doing the using have noticed.
The pattern underneath both findings was described a year earlier. RAND's survey of district leaders, published on April 8, 2025 by Melissa Kay Diliberti, Robin J. Lake and Steven R. Weiner, found that 48 percent of districts had provided AI training by fall 2024, up from 23 percent the year before. It also found what that training consisted of. District leaders prioritized addressing teachers' concerns, confusion and fears about the technology ahead of practical application. Nearly all of it was optional. Most of it was built in house, because leaders struggled to find established experts who understood AI in an educational setting, and most of it was one-off rather than sustained.
Read those three findings together and the story is not that organizations are failing to train people. It is that they are succeeding at the easiest version of training and calling it the finished job.
Confidence is not capability, and the two come apart faster than leaders expect
Here is the distinction almost every training plan gets wrong. Confidence is a feeling about a tool. Capability is a change in what somebody can do. Training produces the first reliably and the second only by design, and the first is far easier to see in a feedback form.
I understand why the feeling gets measured. It is the thing people report on the way out of the room, and a room full of staff who are less frightened of AI than they were that morning is a genuine achievement, particularly if the previous position was refusal. But comfort with a product is a poor proxy for judgment about the work. A person who can get a good answer out of one AI product has learned a product. A person who can tell a good answer from a plausible one has learned the work. Only the second survives the next version, the next procurement decision, and the next member of staff who asks how to do it.
The risk on the other side of this is now being named openly. Harvard's Liz Mineo reported in May 2026 on what researchers are calling never-skilling: not the erosion of a skill somebody once had, which is deskilling, but the failure to build it in the first place. Stephanie Smith Budhai of the University of Delaware put it plainly in that piece: students do not know how to write a topic sentence because they are asking AI for the topic sentence. That research is about students, and I want to be careful not to overclaim it. My own reading is that the mechanism does not care about age. An adult who has only ever produced a scheme of work with a model's help, and who has never had to argue for one, has the same gap. Nobody has taken a skill away from them. It was simply never built.
The Dependence Test
The Dependence Test is three questions I ask of any AI training before an organization commissions more of it: If the tool were switched off on Monday, what would this person be able to do that they could not do before? Who here can now teach this to somebody else, without the person who ran the session? And when the output is wrong, who notices, and what do they do about it? Training that survives all three has built capability, which stays in the organization. Training that fails any of them has built dependence, which leaves with the tool.
This is my suggested way of reading a training plan, not a validated instrument. It is deliberately hard to pass, because the failure it is designed to catch looks like success on every other measure.
If the tool were switched off on Monday, what could this person still do?
The first question separates a skill from a subscription. Answer it about a named person and a named task, not about a cohort. If the honest answer is that they would go back to what they did before, the training transferred fluency in a product and nothing else. A good answer sounds like this: they would still write the first draft badly and fix it, faster than they used to, because they now know what a weak draft looks like and where it usually goes wrong. That is judgment, and it was not in the product.
Who here can teach this to somebody else?
The second question asks whether the capability is actually in the organization or only in an individual. A capability that has never been transferred once inside the building is not an organizational capability, whatever the attendance register says. It is a single point of failure with a person's name on it. This is the question that exposes the champion model most quickly: one enthusiastic colleague who can do remarkable things and cannot explain any of it is a dependency, not a capability, and the organization will discover which when they change jobs.
When the output is wrong, who notices, and what do they do?
The third question is the one that decides whether any of this is safe. AI will produce a plausible wrong answer eventually, and the test of a capable organization is not that this never happens but that somebody catches it and knows the route. If the only person who would spot an error is the person who generated it, and there is no route for them to raise it, you have built a system that works until it does not. This connects directly to the question of who is accountable when an AI tool gets something wrong, which is a governance decision and needs making before the training, not after it.
The Mistake I See Most Often
Across the leadership teams I work with, the mistake I see most often is commissioning the second training day before anyone has decided what the first one was supposed to change. The day gets booked, it goes well, the feedback is warm, and the request that arrives a term later is for more of the same, at greater depth, for a wider group. Nobody has asked what people stopped doing, or who can now teach it, or what happens when the output is wrong. The organization has a training habit and no capability strategy, and the two feel identical from the inside.
I have run enough of these sessions to know how flattering the feedback is. Having delivered training to more than 150,000 educators across more than thirty countries, I can say that the warmest rooms are not reliably the ones where the most changed. A session that leaves people enthusiastic and unchanged produces better evaluation scores than one that leaves them unsettled about how they have been working. When I taught, and later when I was responsible for digital strategy across a further education organization, the professional development that actually moved practice was almost never the day everybody enjoyed. It was the follow-up nobody had asked for.
What capability-building training does differently
Training that passes The Dependence Test is designed differently in three specific ways, and none of them is about content.
It teaches the judgment, not the interface. The session works on real material the organization is stuck with, and the hardest part of it is evaluating what comes back rather than producing it. If the agenda is a tour of features, you are buying a product demonstration. This is the same argument that applies to what happens after the initial AI training: a skill only becomes a change in the work when somebody makes a decision in between.
It builds a second teacher on purpose. Somebody in the room is identified in advance as the person who will run the next version, and their preparation is part of the design rather than an afterthought. This is the cheapest structural change available and it is the one most often skipped, because it makes the first session longer and the second one unnecessary.
It names the error path before it needs one. Staff leave knowing what a wrong output looks like in their own job, who they tell, and what happens next. A hypothetical example, to show the shape of it: a department agrees that any AI-generated report comment which names an assessment the student did not sit goes to the head of department the same day, and the head of department logs it. Invented detail, real mechanism. Without something like it, the third question of the test has no answer.
Is some dependence acceptable?
Yes, and pretending otherwise would be dishonest. Nobody expects a leadership team to be able to run payroll by hand if the system fails, and nobody thinks less of a school for depending on its electricity supply. Dependence on a tool is a reasonable position to hold deliberately.
The distinction is whether it was chosen. Deliberate dependence comes with a named owner, a known failure mode and a plan for the day the supplier changes the terms or withdraws the product. Accidental dependence has none of those, and it is usually discovered rather than decided. The question to put to a leadership team is not whether you depend on anything, but whether you could say out loud which capabilities you have decided not to hold, and who agreed to that.
Where I would not accept it is anywhere the work is also how somebody learns the job. That is the argument I have made about what happens to entry-level work, and it applies inside a staff team as much as to a labor market. If the task you have automated was the one that produced the next person capable of judging it, the dependence compounds, and the organization gets quietly worse at something it no longer notices doing.
What to do next
Take the next AI training your organization has planned and put the three questions to it before it runs, not after. If you cannot answer the first about a named person and a named task, the session is a product demonstration and should be rewritten. If you cannot answer the second, add the second teacher now, while it is still a scheduling decision. If you cannot answer the third, stop and make the governance decision first, because training people to use a tool nobody can correct is the one version of this that makes an organization less safe rather than more capable.
If your leadership team is working out what its staff actually need to be able to do, rather than how many sessions to run, that is the kind of work I support through AI training and capability programs for schools and organizations. The related question of how to raise confidence across a whole workforce without stopping there is covered in how leaders build AI confidence across a workforce. I also write about this weekly for leaders who are making these decisions now, and you can join that list here.
Sources and further reading
- Elizabeth Heubeck, "More Schools Are Providing AI Training for Teachers. Is It Any Good?", Education Week, May 18, 2026.
- Microsoft, "2026 AI in Education Report", June 24, 2026.
- Melissa Kay Diliberti, Robin J. Lake and Steven R. Weiner, "More Districts Are Training Teachers on Artificial Intelligence: Findings from the American School District Panel", RAND Corporation, April 8, 2025.
- Liz Mineo, "'Deskilling' is bad. This is worse.", The Harvard Gazette, May 11, 2026.
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and the author of bestselling books on AI in education. He works with schools, trusts and organizations on AI strategy, readiness and leadership capability. More about Dan.
