Yes, with one line drawn through the middle of the plan. Guidance reported by Schools Week on May 28, 2026 lets SENCOs, teachers and teaching assistants use AI to draft initial support plans and staff guides for pupils with SEND, and says AI should not decide whether a child has a special educational need. A SENCO's skill is knowing which parts of a plan fall on which side.
My own way of thinking, not research: the part of a support plan a stranger could write is the part that was never doing the work.
What does the DfE guidance say about AI and SEND?
It says AI may draft, and people must decide. Schools Week reported on May 28, 2026 that the updated guidance, first issued in June 2025 and developed by Chiltern Learning Trust and the Chartered College of Teaching working with the Department for Education, now covers SEND. According to that report, staff may use AI to draft initial support plans, staff guides for specific needs and template letters.
The same report lists the limits. Staff should be "extremely cautious" about typing a child's SEN or additional needs into a tool, because information that looks anonymous can identify a pupil once it is combined with other details. AI should not write statutory documents such as contributions to an education, health and care (EHC) plan without significant professional review. It should not handle sensitive personal information without safeguards, and it should not provide social, emotional and mental health support. Existing tools such as voice typing, screen readers and text-to-speech should not be overlooked in favor of AI.
One caution about my own evidence. This article relies on Schools Week's account of the guidance, because the SEND material sits in slides and transcripts I could not open. Read the original before you quote it to governors.
The government's product safety standards for generative AI matter here too. Their January 19, 2026 update added standards on cognitive development, emotional and social development, mental health and manipulation. They are written for suppliers, but schools can use them to judge any tool a pupil with additional needs will meet.
Why does a support plan that a stranger could write fail the child?
Because a plan helps only as far as it records something true about one child. The SEND code of practice says settings "should know precisely where children and young people with SEN are in their learning" (paragraph 1.25, Department for Education and Department of Health and Social Care).
Take a hypothetical Year 8 boy who goes quiet in group work. A language model, given a rough profile, will return a sound plan in seconds: a quiet space, tasks in smaller steps, a visual timetable, notice before changes. Every line is reasonable. Every line would suit a hundred other children.
The plan only his SENCO can write says something else. He stops answering when an adult stands over his desk and starts again when she crouches beside it. He will read aloud to his Year 7 reading buddy but not to his own class. Thursday, with its supply teacher, is where it goes wrong.
The first plan is a document. The second is knowledge, and a machine cannot draft knowledge it was never given.
The danger is quiet. A SENCO reading the first plan finds nothing wrong, because nothing is wrong, and it goes in the file. What is missing is an absence, and absences are hard to see. Barristers Jim Hirschmann and Alice de Coverley of 3PB made the legal version of this point in a July 4, 2025 paper on AI and EHC plans. They expected public law challenges where plans are drafted by AI "without adequate human oversight", and noted that AI output is "highly susceptible to bias and misinformation". They were writing about local authorities, but the logic travels to a school's contribution.
What is the Stranger Test?
The Stranger Test is three questions I ask before handing any part of a SEND plan to AI: Would a stranger who found this text be unable to tell which child it describes? Could a stranger who has never met the child do this task well? And will someone who knows the child read every line before it goes anywhere? A task that earns three yeses can go to the machine. A task that earns a no stays out of the tool, stays with you, or stays in the drawer, depending on which question said no.
It is my suggested way of thinking about the line, not a validated instrument. The questions work in order.
- The identity question. A no means the text stays out of the tool. Removing a name is not enough. A year group, a diagnosis, a school and one incident can point at a single child, which is the risk the guidance names. Asking a tool whether a pupil's writing suggests dyslexia fails here for a second reason: it asks the machine to infer health information. The Information Commissioner's Office says that if you infer special category data, or treat someone differently on the strength of an inference, you are processing it however confident you are that the inference is right (ICO guidance on special category data). That is also the decision the guidance says AI should not make.
- The knowledge question. A no means the task stays with you. If the work depends on having sat beside the child, no prompt can supply it. Typing more detail into the tool does not fix this. It fails question one instead.
- The checking question. A no means it does not go anywhere. The person who reads the output must know the child well enough to notice what is missing. A new teaching assistant or a supply teacher cannot do that, so the check becomes a signature on a page nobody has tested.
Which SEND tasks pass the Stranger Test?
The tasks that pass are the ones about a kind of need or a kind of meeting, and the tasks that fail are the ones about a child. The table sorts hypothetical examples. A yes means the task passes that question.
| Task (hypothetical) | Identity | Knowledge | Checking | Verdict |
|---|---|---|---|---|
| Template letter inviting parents to a review meeting | Yes | Yes | Yes | Draft with AI |
| Staff guide to what ADHD can look like in a lesson, written about no particular child | Yes | Yes | Yes, the SENCO reads it | Draft with AI |
| Blank plan layout: headings, review dates, prompts for each section | Yes | Yes | Yes | Draft with AI |
| Strategies section for one named pupil, typed with details | No | No | Yes | SENCO writes |
| Summary of a conversation with a pupil about what helps | No | No | Yes | SENCO writes |
| Message to parents about a concern | Yes | No | Yes | Person writes |
| Contribution to an EHC needs assessment | No | No | Needs significant professional review | SENCO writes |
| "Does this pupil's writing suggest dyslexia?" | No | No | No | Do not use AI |
The sixth row follows the report on the guidance, which says AI should not be used for messages involving concern, reassurance or sensitive context. The pattern across the table is the same as the pattern in the Year 8 example. AI can build the frame. The SENCO supplies the child.
What I See in Practice
The pressure on SEND paperwork is usually workload, and the first place people want to save time is the paperwork that carries the most professional judgment. That is the pattern I see across the leadership teams I work with: AI arrives where the hours are tightest, not where it is safest.
Be fair to the pressure. A SENCO who gets hours back each week from letters, layouts and meeting notes has gained real capacity, and the Stranger Test releases those hours. The mistake is spending the saving on the one section of the plan that cannot be delegated.
As a school trustee, the question I would want a board to ask is not whether the school uses AI for SEND. It is which parts of a plan a person still writes, and who signs them. A school that can answer that in two sentences is governing the tool. A school that cannot is hoping.
What should a headteacher and SENCO settle before anyone drafts with AI?
Five decisions, agreed in writing, before the first plan goes near a tool.
- Name the tool. Use one that has been through a data protection impact assessment, not whatever a member of staff already has open. Wiltshire Council's public page on AI in EHC plan drafting, which carries no date, shows what a local authority tells families: the tool is "a support tool only, not as a decision-maker", a human officer "remains fully responsible", and the data is used only for statutory SEND functions. A school's tool deserves a comparable statement. Our article on what to require before approving an AI tool sets out the checks.
- Write the no column. Put half a page of tasks that never go to AI where every member of staff can see it, starting with the table above.
- Start with what the pupil already uses. Voice typing, screen readers and text-to-speech come first, as the guidance says. They put the tool in the child's hands without sending the child's details anywhere.
- Settle who answers. When a plan is wrong, one named person owns the correction. Who is accountable when AI gets it wrong covers how to name that person before the error arrives.
- Set the rule on student data. The line between a prompt that is safe and one that is not must be one every adult can state. Can staff put student data into ChatGPT gives the rule to write.
Where does the Stranger Test stop working?
It stops where the risk moves from the plan to the pupil. The test governs what staff type into a tool. It says nothing about a child with additional needs who talks to a chatbot directly. The guidance, as reported, warns that vulnerable children need help telling AI from a human because of the risk of emotional attachment. That belongs with safeguarding, covered in what AI means for safeguarding in schools.
It is also England-shaped. The statutory anchors above, the SEND code of practice and the EHC plan, are English law and practice. Leaders elsewhere, including teams that write IEPs in the United States, can use the three questions, but should check their own legal duties and state or national guidance rather than lean on this article for them.
Finally, the guidance will keep moving. Treat the table as a draft to revisit each term, and date it.
The Leadership Question
The question is not whether your SENCO should use AI. It is what you want the hours back for.
A SENCO with the letters and layouts handled has more time to sit beside the pupil who goes quiet on Thursdays. That is a decision for the leadership team to make deliberately, and to protect, because the time that comes back will be spent on something. If your leadership team is working through where AI belongs in SEND, this is the kind of work I support through AI strategy sessions and advisory work with school leaders.
Sources and further reading
- Schools Week, "DfE guidance suggests schools use AI to draft SEND support plans", May 28, 2026: schoolsweek.co.uk. Read at source October 13, 2026.
- Department for Education, "Generative AI: product safety standards", first published January 22, 2025, updated January 19, 2026: gov.uk.
- Department for Education and Department of Health and Social Care, "SEND code of practice: 0 to 25 years", published June 11, 2014, updated September 12, 2024 (paragraph 1.25 quoted): gov.uk.
- Information Commissioner's Office, "What is special category data?": ico.org.uk.
- Jim Hirschmann and Alice de Coverley, 3PB Barristers, "Artificial Intelligence in Education and EHC Plans", July 4, 2025: 3pb.co.uk.
- Wiltshire Council, "AI in EHCP drafting" (undated page): wiltshire.gov.uk.
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor, an international keynote speaker and a school trustee, and a former secondary teacher and assistant headteacher.
