Most schools and organisations have just run the year's first AI training session, and most of the people who sat through it are no more confident than they were before. That is not a guess. Bett's YouGov survey of 1,033 UK teachers, The State of AI in Education 2026 (fieldwork June 2026, published 27 August), found that 80% now use AI for work and weekly use has doubled in a year, while 41% have had no AI training from their school and the share expecting to use AI more next year has fallen from 68% to 47%. The Royal Society's Teacher Tapp survey of more than 9,000 teachers in England, submitted to the Education Committee in May 2026 and reported widely at the start of this term, found that only 14% are confident across the full range of AI literacy.
Use is going up. Confidence is not following it. And the standard response, another training day, is the smallest lever a leader has.
Here is the answer, then the reasoning. Leaders build AI confidence across a workforce by creating three conditions that a training session cannot create on its own: permission (people know what they may use AI for), practice (they have used it on their own work, repeatedly, alongside someone who has done it before), and protection (it is safe to be seen using it and safe to say when it went wrong). Training is the third thing you do, not the first, and it only works once the other two exist.
Why AI confidence is not the same as AI use
AI confidence is the state in which a member of staff can use AI on real work without asking anyone whether they are allowed to, and would tell someone if it got something wrong; usage figures tell you nothing about whether that state exists. The Bett survey makes the gap visible on a single page: 80% of teachers use AI for work, 48% say their school has no written AI policy, and only 18% understand the policy their school does have. A person using AI in an organisation with no rules is not confident. They are unsupervised. From the outside the two look identical.
Confidence is not enthusiasm either. Enthusiasm is what your early adopters have, and they will have it whatever you do. Confidence is what your sceptics have when they can use the tool on their own work without asking permission and without hiding it. Do not count the enthusiasts. Ask the sceptics.
The Royal Society data draws one more line worth keeping. A third of teachers were confident in the practical use of the tools but not in the rest of AI literacy: understanding what the tools are doing and judging when to trust them. Fluency is a skill, and you can teach it in an afternoon. Confidence is a condition, created or withheld by the organisation around the person. That distinction is the whole argument of this article.
What the research says actually builds confidence
The strongest evidence available this year says confidence follows structured support combined with clear whole-school guidance, and does not follow training on its own. The Royal Society's written evidence puts it in one sentence: teachers in schools with structured professional development and whole-school guidance are "roughly three times more likely to be confident across all dimensions of AI literacy than those receiving none." Only 18% had received structured professional development, 15% had clear whole-school guidance, and 23% had received no support of any kind. The lever is the combination. Guidance without practice produces people who know the rules and never use the tool. Practice without guidance produces people who use the tool and hope nobody asks.
Two more findings sharpen the picture. David Marshall of Auburn University and Tim Pressley of Christopher Newport University surveyed more than 400 US teachers and reported in Education Week on 24 April 2026 that "AI didn't directly reduce teachers' workload"; what mattered was confidence, and their advice to leaders was "repeated, practice-based support", clear rules and protected time. And training is rising while its depth is not: the EdWeek Research Center's surveys, reported on 18 May 2026, show the share of US teachers with no AI training falling from 60% in October 2024 to 42% by early 2026, while only 9% describe their training as ongoing. Training is up, ongoing training is rare, and confidence is flat. Whatever is missing, it is not a shortage of sessions.
The Confidence Test
The Confidence Test is three questions I ask a member of staff, not their leader, to find out whether AI confidence exists in an organisation: Do you know what you are allowed to use it for? Have you used it on your own work this week? If it got something wrong, would you tell someone? Confidence is when all three answers are yes from the people least likely to give them.
I ask staff rather than leaders for a reason. Leaders describe the organisation they intended to build. Staff describe the one they work in. The Royal Society survey found headteachers among the least confident groups of all, according to EdTech Innovation Hub's coverage of the results on 2 September 2026, which is one more reason the view from the top is not the measure.
The three questions map onto the three conditions: permission, practice, protection. When one answer is no, you have found the condition that is missing, and that is the thing to fix before anyone books another trainer. This is my suggested way of reading an organisation, not a validated instrument.
Permission: people need to know what they may do, not only what they may not
The first condition of confidence is a permission that a member of staff can state in one sentence without looking it up. Most AI policies fail this test because they were written to prevent harm rather than to direct use, so they list prohibitions and leave the permissions implied. Staff are not confused about what is banned. They are confused about what is allowed, and careful people, given no answer, assume the answer is nothing.
The second question in the Readiness Test is "What may I do with it?", and the Bett finding that only 18% of teachers understand their school's policy well tells you how rarely that question has an answer people can carry in their heads. Weak permission reads: "Staff may use AI tools appropriately and in line with data protection requirements." Strong permission reads: "You may use the two approved tools to draft resources, letters and first versions of reports. You may not paste in a pupil's name, a medical detail or anything from a safeguarding record. If you are unsure, ask your head of department, who has been asked to say yes wherever they can."
Practice: confidence comes from using it on your own work, more than once
The second condition is repeated use on the person's real work, alongside someone who has done it before. This is where the training day fails, and it fails structurally rather than through anyone's fault. A training day teaches the tool on somebody else's example, in a room with no marking in it, and then sends people back to a week with no time in it. RAND's study of US districts (8 April 2025) found that nearly all district AI training was voluntary. Voluntary training reaches the enthusiasts. The sceptics were not in the room.
What builds confidence is smaller and more frequent. A hypothetical example: a head of science gives the last twenty minutes of a department meeting to one task, drafting feedback comments for the mock exam everyone is marking that week, using the approved tool, and does it alongside them on her own class set. Two weeks later they do it again with a different task. Nobody has attended a course. Everybody has used the tool on their own work, twice, next to someone who could answer the question they were embarrassed to ask. That is the part of structured support most organisations leave out.
Protection: it has to be safe to be seen using it and safe to be wrong
The third condition is that nobody is punished for trying and nobody is punished for reporting a mistake. The fourth question in the Sponsorship Test is "Have we made it safe to be seen using it?", and for most organisations the honest answer is that they have never said. Silence is read as risk. Where senior leaders have not been seen using AI, a teacher who uses it to draft reports is taking a small professional gamble every time, and confidence does not grow in people who feel they are gambling.
The mistake question matters even more. AI will produce something wrong in your organisation this term. The only question is whether you hear about it. The people who would tell you are exactly the people you need, because they are the ones checking, and the first reported error should be called good news in public. An organisation that treats the first AI mistake as a disciplinary matter has just guaranteed it will never hear about the second.
Why the time saved decides whether confidence lasts
Confidence collapses when the time AI saves is absorbed by the organisation without anyone deciding what it was for. This is the finding in the Bett survey that leaders should sit with longest. Among teachers using AI, 51% said it had eased their workload, but only 35% were working fewer hours and 55% said their hours were unchanged. Bett's Duncan Verry put it plainly: teachers "tell us it saves them time on particular jobs, but most say they are not working fewer hours", because something else fills it. The drop from 68% to 47% in those expecting to use AI more next year is what cooling enthusiasm looks like in a survey, and I read it as the sound of a promise not being kept.
The promise was less work. Nobody decided what the saved time was for, so the organisation took it back. I wrote in Why AI Pilots So Often Go Nowhere that "a time saving that no one redirects is not an outcome. It is a rounding error." The same rule applies to confidence. Say publicly what the hour it saves is for: earlier finishes, fewer written reports, more time with the pupils who need it, anything, as long as it is named and then protected. A saving that vanishes teaches people the tool changed nothing. A saving they can see teaches them it was worth learning.
What I Have Learned From Working With Organisations
I have sat on both sides of the training day, as the assistant headteacher responsible for planning it and now as the person invited in to deliver it, and the lesson from both sides is the same. The session is never the thing that changes behaviour. What changes behaviour is what the leadership team does in the fortnight afterwards. When I write about this for Forbes, the replies are rarely from people asking how to run a better session. They are from leaders who ran a good one and cannot understand why nothing moved.
Across the organisations I work with, the pattern is consistent enough that I can usually predict it before I arrive. The enthusiasts have been using AI for a year. The middle tried it once, got something mediocre, and stopped. The sceptics have decided it is a fad and are waiting to be proved right. A training day makes the enthusiasts happy, gives the middle a second mediocre experience, and confirms the sceptics' suspicion. The moment the room changes is not when the trainer shows something impressive. It is when someone senior says what they used it for last week and what it got wrong. Permission and protection arrive in the same sentence, and the sceptics start listening. The mistakes I see most often all follow from the same misreading: training the enthusiasts, teaching the tool instead of the task, a policy that only says no, and measuring attendance, which tells you who was in the room and nothing about who changed.
What good looks like, and what to do this term
An organisation with AI confidence is one where the least enthusiastic member of staff can pass the Confidence Test without preparing for it: a permission short enough to remember, practice built into meetings that already exist, senior leaders who have been heard describing what AI got wrong for them, and a mistake that was reported, thanked and fixed within the last month. If nothing has gone wrong yet, either nobody is using it or nobody is telling you.
So run the Confidence Test with the five people least likely to volunteer for anything about AI, and fix whichever condition fails before you spend another day on training. If permission fails, write the one-sentence permission this week and have it said aloud by the most senior person available. If practice fails, put twenty minutes of real work into the next three team meetings and ask team leaders to do the task alongside their people. If protection fails, have a senior leader describe, in public, something AI got wrong for them and what they did about it. Then decide what the saved time is for, say it, and defend it.
Do that, and the training you run next will land on an organisation that is ready for it. Skip it, and the training will do what training done first always does: please the people who were already convinced, and leave everyone else where they were.
If your leadership team is working out what comes after the training day, this is the kind of work I support through leadership development and organisation-wide AI training, built around the conditions above rather than around a single session.
Sources and further reading
- The State of AI in Education 2026 (Working It Out), Bett with YouGov, survey of 1,033 UK teachers, fieldwork June 2026, published 27 August 2026. uk.bettshow.com; coverage in Tes and Schools Week, 27 August 2026.
- Written evidence from the Royal Society to the House of Commons Education Committee (AIE0215), including the Teacher Tapp survey of more than 9,000 teachers in England, fieldwork March 2026, submitted May 2026. committees.parliament.uk; project page at royalsociety.org; coverage in EdTech Innovation Hub, 2 September 2026.
- Marshall, D. T. and Pressley, T., We Studied How AI Shapes Teachers' Well-Being. Here's What We Found, Education Week (opinion), 24 April 2026. edweek.org
- Heubeck, E., More Schools Are Providing AI Training for Teachers. Is It Any Good?, Education Week, 18 May 2026, reporting EdWeek Research Center survey data. edweek.org
- Diliberti, M. K., Lake, R. J. and Weiner, S. R., More Districts Are Training Teachers on Artificial Intelligence: Findings from the American School District Panel, RAND, 8 April 2025. rand.org
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor, a former assistant headteacher and Director of Digital Strategy, and works with schools and organisations on AI strategy, leadership and capability. About Dan.
