What does the DfE's 2026 report actually say about ChatGPT in schools?
Teachers are using ChatGPT and Claude widely — but for planning and admin, not for marking. That is the quiet but important finding in the Department for Education's June 2026 report, Assessment of the education technology market in England. GenAI (generative AI — the family of large language models that produce text on demand) has spread fast through staffrooms, and the report documents where it is actually landing. EdTech here means the use of technology to support and enhance education, and within that market the DfE traces a clear split between general-purpose tools and the specialised systems schools buy for specific jobs.
The report situates this against a market worth £6.5bn in turnover across 1,123 UK EdTech companies, growing at an average of 8.8% a year (DfE, June 2026, p.7). Classroom and assessment tools, the DfE says, represent the second-highest area of engagement (p.9). But the headline behaviour is general-purpose adoption arriving ahead of specialised tools — teachers reaching for whatever is to hand.
The phrasing is deliberate. The DfE names the use cases — lesson planning and administrative tasks — and stops there. Marking is conspicuously absent from that sentence, and the report is candid that "reliable evidence on how EdTech is used in practice in schools and colleges remains limited" (p.8). The behaviour the DfE can observe is teachers using general-purpose AI to draft, summarise and organise — the low-risk, high-frequency tasks where a generic model genuinely saves time.
Where does general-purpose AI work well — and where does it break down?
The distinction the DfE draws maps neatly onto a simple test: does the task depend on a specific UK mark scheme, or not? General-purpose tools shine on open-ended drafting and fall over the moment exam-board criteria are involved.
| Task | General-purpose AI (ChatGPT, Claude) | Specialised, curriculum-aligned AI |
|---|---|---|
| Drafting a lesson plan | Works well — fast, flexible, no mark scheme needed | Overkill — not the tool's purpose |
| Writing a parent email or report comment | Works well — rephrasing and tone are general tasks | Not designed for this |
| Generating a model answer to study | Useful — but verify against the spec yourself | Grounded in the actual assessment objectives |
| Marking against an AQA / Edexcel level descriptor | Breaks down — no mark scheme inside the model; inconsistent | Built for this — applies the same band descriptors each time |
| Awarding a grade on identifiable pupil work | Breaks down — data-protection and hallucination risk | Designed for pupil work, with human moderation in the loop |
| Consistent feedback across a class set | Breaks down — answers vary query to query | Standardised against fixed criteria |
The pattern is clear: general-purpose AI is excellent at the tasks the DfE actually observed teachers using it for, and unreliable at the one task it is most often imagined doing — marking. That is not a flaw in ChatGPT; it is a question of fit, which is also the heart of the broader specialised versus general-purpose AI debate in education.
Why aren't teachers using ChatGPT for marking at scale?
Marking is the obvious place to want AI help — it is the workload that pushes teachers towards these tools in the first place. So why does the DfE see general-purpose uptake everywhere except marking? Three reasons stand out:
- Hallucination risk. A general-purpose model will confidently invent a mark, misquote an assessment objective, or praise an answer that misses the criteria entirely. For drafting that is recoverable; for grading it is not.
- No mark scheme alignment. There is no AQA, Edexcel, OCR or WJEC mark scheme inside a general LLM. It is reasoning from general patterns in language, not applying the specific band descriptors a UK examiner uses.
- Inconsistency across queries. Ask the same model to mark the same essay twice and you can get two different levels. Reliable assessment demands the opposite — the same standard applied to every script.
None of this means AI cannot mark. It means the marking job needs a tool built for it. The pull factor is real — the workload behind it is documented in our look at how many hours UK teachers spend marking each week — but the answer is alignment, not a generic chatbot.
What is "specialised AI" and why does the DfE highlight it?
Specialised AI is built for one job and grounded in the criteria that job requires — in assessment, that means a specific exam-board mark scheme rather than the model's general knowledge. The DfE flags this as the direction of travel for the whole market, not a niche.
That reading is consistent with the assessment vertical being the fastest-growing part of the market — 71 companies, £470m turnover and 18.1% annual growth, the highest of any vertical (DfE, June 2026, p.8). Curriculum-aligned marking tools sit squarely in that space. ReMarkAble AI is one example of the trend the DfE describes: a UK-built tool that marks student work against AQA, Edexcel, OCR and WJEC mark schemes across KS1 (Key Stage 1, ages 5–7) through A-Level, returning examiner-style feedback rather than the open-ended commentary a general chatbot produces. It is one instance among others — the point is the category, not any single product. How well such tools hold up against examiner standards is the subject of our piece on examiner-quality AI marking and moderation.
What should schools do about staff GenAI policy?
The most useful thing the DfE's data does for school leaders is hand them a ready-made distinction. A staff GenAI policy does not need to ban or bless AI wholesale; it needs to separate the two jobs the report already separates:
- General-purpose AI for drafting and admin. Approve ChatGPT or Claude for lesson planning, rephrasing, summarising and idea generation — the uses the DfE actually observed — with a firm rule that no identifiable pupil data goes in.
- Specialised, curriculum-aligned AI for grading-adjacent tasks. Where AI touches marks or pupil work, require a tool built around the relevant mark scheme, and keep a human in the loop to sign off anything that influences a grade.
- Name an owner and review cycle. The DfE lists strategic capability gaps and sustained-adoption measurement among the four system-level constraints schools face (p.10) — a named owner who reviews tools and outcomes addresses both.
Keeping the policy short and task-based is more sustainable than chasing every new tool. The line that does the work is simple: general AI for drafting, specialised AI for anything curriculum-aligned, and human judgement on every grade.
See what curriculum-aligned marking looks like
ReMarkAble AI marks student work against AQA, Edexcel, OCR and WJEC mark schemes from KS1 through A-Level, returning examiner-style feedback with strengths, areas to improve and next steps. It is the specialised end of the split the DfE describes — and it is free to try, no card required.
Frequently Asked Questions
Can teachers use ChatGPT to mark student work?
They can, but the DfE's June 2026 EdTech market assessment found that schools are using general-purpose tools such as ChatGPT and Claude mainly for lesson planning and administrative tasks — not for marking at scale. The reason is structural: a general-purpose large language model has no AQA, Edexcel, OCR or WJEC mark scheme inside it, so it cannot reliably apply level descriptors the way an examiner does. ChatGPT is genuinely useful for drafting a model answer or rephrasing feedback, but for awarding marks against exam-board criteria it tends to be inconsistent. For summative or moderation-grade marking, teachers should treat it as a drafting aid, not an examiner.
Is ChatGPT accurate for GCSE English?
ChatGPT can produce plausible-sounding commentary on a GCSE English essay, but accuracy against the actual mark scheme is the problem. General-purpose models do not hold the specific band descriptors for AQA, Edexcel, OCR or WJEC English specifications, and they can hallucinate marks, misapply assessment objectives, or grade the same answer differently on repeat queries. That inconsistency is fine for brainstorming but unreliable for grading. Curriculum-aligned tools built around the relevant mark scheme are designed to close that gap, which is why the DfE highlights a shift towards specialised AI.
What's the difference between ChatGPT and a specialised AI marker?
ChatGPT is a general-purpose GenAI platform built to handle almost any text task, from emails to lesson plans. A specialised AI marker is built for one job — assessing student work against a specific exam-board mark scheme — and is grounded in those criteria rather than relying on the model's general knowledge. The practical difference is consistency and curriculum alignment: a specialised tool applies the same level descriptors every time and references the assessment objectives a UK examiner would use. The DfE's 2026 report frames this as investment moving towards specialised AI solutions across the market.
Does the DfE have guidance on AI in schools?
The DfE's June 2026 report, Assessment of the education technology market in England, maps how schools are adopting AI and documents the rapid uptake of general-purpose GenAI for planning and admin. It is a market assessment rather than a classroom rulebook, and it explicitly notes that reliable evidence on how EdTech is used in practice remains limited. It also sets out four system-level constraints schools face — strategic capability gaps, resource and prioritisation pressures, procurement and infrastructure constraints, and sustained adoption and impact measurement. Schools should read it alongside their own data-protection and safeguarding duties when writing local policy.
How should schools write an AI policy for staff?
Start by distinguishing the two jobs the DfE's data already separates: general-purpose AI for drafting and admin, and specialised, curriculum-aligned AI for tasks that touch grading and pupil data. A workable staff policy names which tools are approved for which task, sets a clear rule that no identifiable pupil data goes into a general-purpose tool, and requires human sign-off on anything that influences a grade. It should also address data protection, age-appropriate use, and a named owner for reviewing tools. Keeping the policy short and task-based is more sustainable than trying to ban or approve AI wholesale.