AI may draft a decision, but it may not make one. A machine can write the first version of a support plan, a report comment or a letter home, because a person reads it, changes it and answers for it. It may not decide who qualifies, what grade stands or who is suspended, because then nobody is left who can explain the reasoning.
On September 1, 2026, the District of Columbia's Office of the State Superintendent of Education published a model AI policy that sorts staff use of AI into colors. Red, prohibited: student discipline decisions, evaluations of teacher performance, and determining eligibility for individualized education programs or Section 504 accommodations. Yellow, permitted with enhanced oversight: drafting IEP language, and reviewing and grading student work.
Read those two lists again. Drafting IEP language is yellow. Determining who gets an IEP is red. The subject is identical. The verb is not.
Four thousand miles away, England's Department for Education updated its guidance for schools on May 28, 2026 and drew the line in the same place. Staff may use AI to draft initial versions of support plans and template letters home. Staff may not use it to decide whether a student has a special educational need, and statutory documents such as EHCP contributions need significant professional review before they go anywhere. Neither document announces the rule it is applying. Both apply the same one.
That matters this term, because a lot of leaders are writing a policy against a deadline. Ohio districts had to adopt one by July 1, 2026 under House Bill 96. Maryland's SB 720 gave districts 120 days from the state department's guidance, which put its deadline around September 2026, according to the Maryland Association of Counties' Conduit Street on September 10, 2026. California's Department of Education released a model policy this summer. Copy the lists and you inherit this year's answer to a question that changes every time a vendor ships a feature. Take the rule and you can write next year's answer yourself.
Where is the line between drafting a decision and making one?
The defensible line runs between producing the words of a decision and producing the decision itself.
A machine can write the first version of a support plan, a report comment, a governors' paper or a letter to a family, because a person still reads it, changes it and stands behind it. A machine may not determine eligibility, settle a grade or decide a suspension, because at that moment nobody is left holding the reasoning.
This has little to do with how good the model is, and everything to do with what happens when the person on the receiving end asks why. A parent whose child has been refused an assessment is owed an explanation from someone who can give one. So is a teacher who has been marked down, and so is a student who has been sent home. If the only available answer is that the system flagged it, the decision was never the school's to begin with.
Nobody is harmed by a machine writing a sentence. People are harmed by a decision nobody can explain.
Why do the model policies look like they disagree?
The apparent contradiction between Washington and Westminster dissolves the moment you read for verbs instead of tasks.
Washington's red list is a list of determinations: who qualifies, what grade, what consequence, how a teacher performed. Its yellow list is a list of drafts and reviews, each of which lands in front of a human being who has to accept or reject it. England's guidance does the same thing in prose rather than color: draft the plan, do not decide the need.
Once you see that, the disagreement between jurisdictions turns out to be smaller than it looks. What varies is where each system places a particular task, following local law, local risk appetite and local capacity. What does not vary is the principle underneath.
Leaders keep asking me which state or which national guidance to follow. It is usually the wrong question. The lists differ. The rule does not.
Which tasks do DC and England place where?
The table below sets the two documents side by side, with my reading of the verb each task falls under. The DC column comes from the model policy's release notice, opened on October 12, 2026. England's column rests on Schools Week's May 28, 2026 reporting of the Department for Education's guidance, not on the guidance itself, so read the original before you quote it to a governing body.
| Task | DC model policy (September 1, 2026) | England, as reported May 28, 2026 | My reading |
|---|---|---|---|
| Determining IEP or Section 504 eligibility; deciding whether a pupil has a special educational need | Red, prohibited | Should not be handed to AI | Decide: refuse |
| Student discipline decisions | Red, prohibited | Not covered in the reporting I read | Decide: refuse |
| Evaluating teacher performance | Red, prohibited | Not covered in the reporting I read | Decide: refuse |
| Drafting IEP language or SEND support plans | Yellow, enhanced oversight | First drafts allowed; staff told to be "extremely cautious" entering SEN information | Draft |
| EHCP contributions and other statutory documents | Not covered in the release | "Significant professional review" required | Draft, with heavy review |
| Reviewing and grading student work | Yellow, enhanced oversight | Not covered in the reporting I read | Inform |
| Lesson planning, school communications, data analysis | Green, permitted when staff stay informed and a person reviews the output | Not covered in the reporting I read | Draft |
The empty cells matter as much as the full ones. Where a document is silent, your policy has to speak, and the next two sections give you a way to do it.
What is the Answerability Test?
The Answerability Test is three questions I ask before letting AI near any decision. A decision that survives all three can be drafted by a machine. A decision that fails any of them has to be made by a person, and that person has to be named.
- Who does this decision land on? Name them. If you cannot, stop here.
- Could a named person explain it to them, in their own words, without mentioning the tool?
- If it turned out to be wrong, who puts it right, and by when?
This is my suggested way of thinking about the line, not a validated instrument, and it is deliberately blunt so that it survives contact with a busy leadership team.
The second question does the real work. "Without mentioning the tool" is the part people find uncomfortable, and it should be. If a deputy head can justify a decision on its merits, the AI that helped produce it was a drafting aid. If the justification collapses into a description of the software, the school has quietly transferred a judgment it is still legally and morally holding.
The third question is the one most policies forget. Every automated process eventually produces a wrong answer about a real child, and the organizations that come out of it well are the ones that already knew who fixes it.
How do you sort decisions into draft, inform and decide?
Sort every AI-touched decision in your organization into one of three verbs, and treat the third as a refusal.
Draft. AI produces the words. A person owns the judgment and could defend it unaided. A support plan drafted by a model and rewritten by the teacher who knows the child sits here, along with most report writing and nearly all correspondence.
Inform. AI supplies evidence into a decision a person makes, and that person knows how the evidence was produced and where it fails. An attendance model that flags a student as at risk belongs here, but only if the pastoral lead can say what the flag is counting and which students it systematically misses.
Decide. The output becomes the outcome, and nobody is in a position to answer for it. Automatic grade release, automated eligibility screening and any rule that fires without a human in the path belong here. This is the category to refuse, in writing, in your policy, by name.
The first two are work. The third is an abdication wearing the clothes of efficiency.
How do you run the Decision Sort on your own decisions?
Run the Decision Sort by listing the ten decisions your organization makes most often about individual people and putting each through the Answerability Test and the three verbs, in one table. The Decision Sort is my suggested way of combining the two, not a validated method. Copy the layout below. The four filled rows are hypothetical, there to show the shape.
| Decision | Lands on | Explainable without the tool? | Who fixes it, and by when | Verb | Named owner |
|---|---|---|---|---|---|
| Year 9 report comments (hypothetical) | Students and families | Yes, the teacher wrote the final wording | Head of year, within a week | Draft | Head of year |
| Attendance-risk flag (hypothetical) | Families, pastoral team | Only if the lead can say what the flag counts | Pastoral lead, within a week | Inform | Pastoral lead |
| Predicted grade released automatically to parents (hypothetical) | Students and families | No | Nobody named | Decide | None, so refuse |
| Referral for SEN assessment (hypothetical) | The pupil | Must be yes | SENCO, within a term | Person decides | SENCO |
| Your decision 1 to 10 |
Then read the table back with these six checks:
- Any "No" in the explainable column. Move the decision to a person, or stop it. A decision nobody can explain without the tool is the tool's decision.
- Any blank in the owner column. Name one person. "The office" and "the team" are not owners.
- Any "Decide" row. Refuse it in writing in your policy, with the owner's name beside it. The occasional rule that fires without a person, such as an automatic overdue-book notice, is rarely worth the argument. The rows that matter are the ones that land on a child in a way a parent would want explained.
- The data going in. Schools Week reports that staff are told to be "extremely cautious" about entering information on a child's special needs or disabilities, because details that look anonymous can identify a pupil when combined with other data. A draft is safe only if what went into it was. The rule on student data and public AI tools sits upstream of every Draft row.
- The review date. Put one on every row. Vendors change what their products do, so a sort done in October is a statement about October.
- The decide column goes to the board. It is the column your governors or trustees will be asked about.
What I Tell Leadership Teams
I tell them the list in their policy will be out of date before the year is out, and to write the rule down next to it.
Across the leadership teams I work with, the pattern is consistent: they arrive with a document that names tools and leave with one that names decisions. That change is small on paper and large in practice, because a policy built around tools has to be rewritten every time procurement signs something new, and a policy built around decisions only has to be rewritten when the school changes its mind about who is accountable.
The second thing I tell them is to run the test on decisions they already make with software, not just the ones involving generative AI. Timetabling, behavior points, reading-age scores and safeguarding filters have been shaping outcomes for years with far less scrutiny than a chatbot now attracts. Some of those, examined honestly, are already sitting in the decide column.
Having worked with government bodies on AI guidance, including the Department for Education, KHDA in Dubai and the Ministry of Education in Kazakhstan, I would add one caution about model policies generally. They are written to be adoptable by every institution in a jurisdiction, which means they describe a floor. A floor is a useful thing to stand on. It is not a strategy, and it is not the same as knowing where your own line falls.
What does good look like?
A school where this is settled can produce, on request, a short list of decisions AI may not make and a named person against each one.
That is the observable test, and it is a low bar that very few clear. The RAND American Educator Panels, reporting in September 2025, found 45 percent of principals reported having school or district policies or guidance on AI use, against 53 percent of English language arts, math and science teachers who said they used AI for school. Guidance has been running behind practice for a while, and the deadlines arriving this fall are the system's attempt to close the gap.
Good also looks like a decision that has been reversed. When a leadership team can tell you about the time an AI-assisted judgment was wrong, who noticed, and what happened next, the accountability is real rather than declared. Until then, you have a document.
Where should a leadership team start this term?
Fill in the Decision Sort in one meeting, with the people who make those decisions in the room, because the person doing the drafting knows exactly where the judgment currently sits.
Then take the decide column to your board or governors, because that column is the one they will be asked about, and the questions a leadership team should be able to answer about AI start with who owns the answer. When a decision on your list is really a question of who answers when AI gets it wrong, settle that first.
This work sits alongside two decisions many schools have already made this year: whether and how to restrict students' own use of AI, which is a separate question about learning rather than accountability, and what responsible AI adoption looks like in practice, which supplies the surrounding behaviors. If you are unsure whether the rest of your organization could hold a decision like this, the seven dimensions of AI readiness will tell you where the weakness is.
The deadline in your calendar asks for a policy. The question underneath it asks who you are willing to name. If your leadership team or your board is working through that question this term, it is the kind of work I support through AI strategy sessions with school and trust leadership teams. I write each week about the decisions leaders are keeping for themselves in the newsletter.
Sources and further reading
Sources last checked October 12, 2026.
- OSSE Releases AI Model Policy to Guide Responsible Staff Use in Schools, Office of the State Superintendent of Education, District of Columbia, September 1, 2026.
- DfE guidance suggests schools use AI to draft SEND support plans, Schools Week, May 28, 2026, reporting the Department for Education's updated guidance for schools and colleges.
- AI Model Policy for Ohio Districts and Schools, Ohio Department of Education and Workforce, adoption required by July 1, 2026 under House Bill 96.
- Deadline for AI Policies in Local Schools Approaches in Maryland, Conduit Street, Maryland Association of Counties, September 10, 2026, on SB 720.
- Release of Model Policy: Artificial Intelligence in Education, California Department of Education, July 6, 2026.
- AI Use in Schools Is Quickly Increasing but Guidance Lags Behind, Doss and colleagues, RAND, September 30, 2025.
Dan Fitzpatrick is The AI Educator: a Forbes contributor, keynote speaker and author who helps leaders and organizations understand, plan for and lead the changes caused by AI.
