On September 9, 2026, the Commons Education Select Committee took evidence on AI in schools. Pip Sanderson of the National Institute of Teaching told the committee what leaders say to her at the end of AI training sessions: "This is all great and really interesting, but what do I do?" The NASUWT's Darren Northcott described the Department for Education's guidance as "quite laissez-faire" and said schools want "much clearer rules of the road." Tes reported the session under the headline "Schools 'desperate' for clearer AI guidance."
I understand the frustration. I also think the diagnosis is wrong. Leaders in that room were not short of guidance. They were short of the small amount of understanding that would let them read the guidance and decide.
What should leaders know about AI?
You need to know five things, and only one of them is about how the technology works. The other four are about your own organization and your own sector: what it is doing when it answers, how it fails, what is already happening inside your walls, what is changing outside them, and what no guidance will ever decide for you. Everything else on the usual list (model names, prompt techniques, the difference between machine learning and deep learning) is interesting and does not change a single decision you will make this term.
That distinction matters because the two things get confused constantly. A leader who cannot explain a transformer is not unqualified to lead AI. A leader who cannot say what would make them stop using a tool is.
1. What is it doing when it answers?
A general-purpose AI assistant is producing the most plausible continuation of your words, not looking up a fact and reporting it. It has read an enormous amount of text and learned what tends to follow what. When you ask it for the attendance regulations, it is not opening the regulations. It is writing the kind of sentence that regulations are written in.
Sometimes it is genuinely retrieving, because the product has been wired to search, or to read a document you uploaded, or to query your own systems. Those are different machines wearing the same interface. The question a leader has to be able to ask a supplier, and get a straight answer to, is: when this thing answers my staff, is it reading something, or is it composing something?
That one question separates a tool you can put near a policy document from a tool you cannot.
2. How does it fail, and why is the failure quiet?
It fails by answering confidently when it should say it does not know, and the reason is structural rather than accidental. In Why Language Models Hallucinate, published on September 5, 2025, Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and colleagues argue that "standard training and evaluation procedures reward guessing over acknowledging uncertainty." The logic is the logic of a multiple-choice exam with no penalty for wrong answers. A guess might score. An honest "I don't know" scores nothing, every time.
Their own numbers show what that produces. On OpenAI's SimpleQA evaluation, o4-mini answered almost every question and got 75% of them wrong. GPT-5-thinking-mini declined to answer just over half, and its error rate came in at 26%. Almost the same accuracy. Opposite behavior when they did not know.
For a leader, the consequence is not "check the output." Everyone says that, and no one does it on the fortieth email of the day. The consequence is sharper: never put AI in a job where the correct output is sometimes nothing at all, and where nobody would notice a confident answer arriving instead. A machine built to produce an answer will produce one. If the missing answer was the safety feature, you have removed it.
3. What is already happening inside your organization?
More than you have authorized, and your staff are learning about it somewhere you are not. RAND's nationally representative survey work, published on September 30, 2025 by Christopher Doss and colleagues, found 53% of English language arts, math and science teachers using AI in 2025, up more than fifteen percentage points in a year, while 45% of principals reported having AI policies or guidance in place. More than 80% of students said no teacher had taught them how to use AI for schoolwork.
Then there is where the understanding comes from. Reporting on the IBM and Morning Consult study of 1,019 educators and 1,029 parents, Education Week's Jennifer Vilcarino noted on September 9, 2026 that 83% of educators feel confident they can teach about AI, and that the most common way they learn about AI in education is social media, at 39%, ahead of professional development at 30% and school or district communications at 28%.
Read those two sentences together. Your staff are confident, they are already using it, and your organization is the fourth voice in their education about it, behind an algorithmically ranked feed. That is not a training gap. It is a supply chain you do not control and have not tried to enter.
4. What is changing outside your organization?
The ground your decisions stand on is moving in three specific places, and you only need to track those three. Assessment, because what a piece of written work proves about a student has changed and the exam bodies are visibly reworking their position. Work, because software that takes actions rather than producing drafts changes what "a human checked it" is worth. And the rules, because the question of what schools must do is live in front of legislators right now: the Education Select Committee opened its AI and EdTech inquiry on February 26, 2026, and US districts have spent the year writing policies against state deadlines.
You do not need to follow any of this daily. I do follow it daily, for a newsletter of more than 44,000 people and a podcast I record every morning, and I can tell you that the daily layer is almost entirely noise for a leader. What you need is a quarterly answer to one question: has anything changed that alters what my organization is allowed to do, or what our students will be judged on? Two of those three have already moved this year. I have written separately about what happens to assessment when AI can produce excellent work and what AI agents mean for work, so I will not re-argue either here.
5. What will no guidance ever decide for you?
Which of the available things matters most in your school, and what you will give up to do it. This is the part leaders are really asking about when they ask for clearer rules of the road, and it is the part no regulator can supply. Guidance draws the edges of the field. It cannot tell you where to stand in it.
Even the clearest possible national guidance would tell you what you may not do with a child's data, what you must record, who must be informed. It would not tell you whether your priority this year is giving teachers their evenings back, or rebuilding how coursework is assessed, or making sure the twelve students who cannot get help at home can get it from a machine at 9pm. Those are the same choices you were making before AI existed, about time, attention and who gets the benefit of a scarce resource.
I sit on a school board, and the thing I notice from that side of the table is how rarely the paper in front of us contains a choice. It contains compliance. Compliance is the floor. Nobody ever got a better school by standing on the floor.
The Mistake I See Most Often
Leaders try to learn AI as a subject, when the thing they actually need is enough understanding to interrogate one decision. Across the leadership teams I work with, this is the pattern that wastes the most time: a head or a superintendent quietly reading around the technology for months, feeling steadily less qualified as they read, while the decisions in front of them sit unmade.
The reading is not the problem. The order is. Start with a decision that is already on your desk, and learn only what that decision requires. Someone wants to use an AI tool to draft support plans: what do you need to know? Whether it is reading the child's record or composing plausible sentences about a child like that one. Who checks it, and whether they would notice an error without being told to look. Whether a parent could be told plainly what happened. Three answers, all obtainable in a week, none of which require you to understand a neural network.
The leaders who move fastest on AI are not the ones who studied it first. They are the ones who picked a live decision and let it tell them what they needed to know.
The Enough Test
You know enough to decide when you can answer three questions about the tool in front of you.
The Enough Test is three questions I ask a leader who wants to know how much they need to understand about AI before deciding: Can you say in one sentence what the tool is doing when it answers? Can you name one thing it will get wrong this term? And can you say what would make you stop using it? A leader who can answer all three knows enough to decide. A leader who cannot is not behind on training. They are deciding on someone else's understanding.
This is my suggested way of thinking about the question, not a validated instrument. Across the leadership teams I work with, it is rare to find one that can answer all three about a tool their staff already rely on. The value is not in passing. It is in discovering which of the three you cannot answer.
Take the AI use your organization now depends on most. Not the pilot. The one that would cause complaints if it stopped on Monday. Run the three questions on that.
If you cannot answer the first, the supplier can, and should, in a sentence. If you cannot answer the second, the people using it daily can, and the fact that they have not told you is itself the finding. If you cannot answer the third, that is the one to write down this week, because a use with no stopping condition is not a decision your organization has made. It is a habit it has acquired.
Where this leaves the guidance question
Clearer guidance would help, and it is reasonable to want it. It will not arrive in time to settle what you do this term, and when it arrives it will draw edges rather than set priorities. The five things above do not depend on it. They depend on knowing what your tools are doing, how they break, what your people are already doing with them, what is shifting underneath you, and which of the available goods you are actually choosing.
That is a short list. It is also, in my experience, the difference between a leadership team that decides and a leadership team that waits. Waiting has its own risks, and they are currently being distributed across your staff one private experiment at a time.
If you are the person who has to make the AI call and answer for it, that is exactly the work I do in one-to-one coaching with school leaders: not a curriculum about AI, but structured conversations about the decision in front of you. For teams working through this together, what good AI leadership looks like sets out the visible signs, and the questions every leadership team should be asking about AI is the agenda I would put in front of them.
Sources and further reading
- "Schools 'desperate' for clearer AI guidance," Tes, September 9, 2026, reporting oral evidence to the Commons Education Select Committee. tes.com
- Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and colleagues, "Why Language Models Hallucinate," OpenAI, September 5, 2025. openai.com
- Christopher J. Doss et al., "AI Use in Schools Is Quickly Increasing but Guidance Lags Behind: Findings from the RAND Survey Panels," RAND, September 30, 2025. rand.org
- Jennifer Vilcarino, "Educators Feel Confident They Can Teach About AI. What Do Parents Think?," Education Week, September 9, 2026, reporting the IBM and Morning Consult survey of 1,019 educators and 1,029 parents. edweek.org
- "AI and EdTech: MPs launch new inquiry examining technology's role in education," Education Committee, UK Parliament, February 26, 2026. committees.parliament.uk
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and a bestselling author on AI in education, and works with school, trust and district leadership teams on AI strategy, readiness and adoption. More about Dan.
