Pupils who write with artificial intelligence (AI) tend to hand in better work. That finding has now appeared often enough that it is hard to argue with. What is far harder to find is anyone checking whether those pupils can write any better the next day, without the tool beside them. Two new papers make that gap the main story about AI and writing this autumn.

The short version: A systematic review published in October 2026 found that all six comparisons of AI-supported writing with ordinary writing instruction favoured AI, yet not one study tested pupils writing without AI afterwards. A separate experiment in China found that primary pupils writing with AI scored highest and felt more confident than pupils writing alone, but reported the lowest sense of absorption in the task. The sensible reading is that AI can improve a text. Whether it improves the writer is still unknown, so teachers should protect the thinking that makes writing worth teaching.

What the research says about AI and writing

The review appeared in Acta Psychologica in October 2026. The authors searched seven databases for 2020 to 2026 studies of generative AI in creative writing for school-age pupils. Seventeen studies met their criteria, covering 736 pupils in total. Six comparisons set AI-supported teaching against teaching without AI, and every one favoured AI on text quality or creativity.

That sounds like a clear win until you read on. In every study, the texts that were scored had been produced with AI to hand, or under conditions the researchers did not report. None of the studies asked pupils to write without AI afterwards, and none checked their writing again after a delay. No intervention planned to withdraw the support gradually, and none ran for longer than six weeks. No study was rated at low risk of bias, and the authors rate the certainty of the evidence as very low.

Their own summary is the sentence I would put on the staffroom wall: the evidence “describes what students produce with AI rather than what they learn to do without it”.

Scaffolding or substitution?

Where pupils kept control, they used AI suggestions to develop their own ideas. Where they did not, the tool could steer the story or supply text that pupils adopted with little reworking. The same software can act as a scaffold or a substitute, depending on how the task is set. That echoes our look at cognitive offloading and AI, where AI editing a pupil’s own draft was treated very differently from AI writing the draft.

A small experiment with a telling detail

The second paper is a study in the Journal of Computer Assisted Learning, published on 2 October 2026. It is small: 78 primary pupils in China, split into three conditions. Twenty wrote with generative AI, 38 worked in pairs (19 pairs) and 20 wrote alone.

The AI group produced the highest creative writing scores. They also reported more confidence in their writing than the pupils who wrote alone, though not significantly more than the pairs. Writing self-efficacy, meaning a pupil’s belief that they can write well, was also positively linked with scores.

The detail I keep returning to is that the AI group reported the lowest flow state, the feeling of being absorbed in the task. The researchers found no link between flow and scores in this sample, so I would not make too much of it. Still, any teacher who has watched a pupil lose themselves in a story knows that absorption is part of what we hope writing gives children.

Pupils mainly used the tool to generate and expand ideas. Where researchers could see what happened to a response, adapting or choosing parts of it was more common than copying it directly, and no complete response was reproduced word for word. That is encouraging for one task, but it says nothing about what those pupils could do the following week.

A better text made with AI is not the same thing as a better writer.

What teachers and writers in the field are saying

On 2 November 2026, The Educator reported on the Journal of Computer Assisted Learning paper and asked two Australian experts where AI belongs in writing lessons. Professor Mary Ryan of Monash University described a “performance paradox”, in which generative AI “can improve short-term writing performance while potentially undermining durable learning and knowledge construction”. She argued that schools should preserve idea generation, purposeful language choice and building a logical argument, while seeing a place for AI in feedback on grammar or in surface-level editing once pupils can edit independently.

Kathleen Killick, chief executive of Seven Steps to Writing Success, said planning and drafting skills need to be taught step by step, because pupils need them to judge whether anything AI produces is any good. Her test for school leaders was simple: “who is doing the thinking”. Where AI takes over the thinking, she said, learning stalls.

What young people say they do

The National Literacy Trust’s 2026 report, published in August from a survey of 40,543 young people aged 13 to 18, puts numbers on the habit. Some 79.0% had used generative AI, and 45.6% had used it to get feedback on their writing, up from 20.7% in 2025. More now check what it gives them (52.3%, up from 42.8%) and add their own ideas (56.5%, up from 48.9%). Yet 24.1% said they simply copied AI output when doing homework, rising to 31.6% among pupils who receive free school meals. Our post on AI and homework covers the learning penalty that copying can bring.

In England, the Education Endowment Foundation has launched a tender in 2026 for up to four projects on how generative AI affects the thinking of pupils aged 13 to 15. It is especially interested in pupils offloading recall, planning, reasoning or drafting onto AI tools, and expects to finalise commissioning from autumn 2026. Evidence from English classrooms is therefore some way off.

How to use AI in writing lessons without losing the learning

Until researchers test pupils on unaided writing, the careful route is to let AI respond to pupils’ work rather than produce it. A few practical steps follow:

• Keep planning and the first paragraph unaided. Ask for the pupil’s own plan and opening in their book before any tool is allowed.
• Use AI for questions, not sentences. A prompt such as “what would a sceptical reader ask about my story?” keeps the writing with the pupil.
• Teach the skills AI cannot judge for them. Pupils who have never practised a strong opening struggle to spot a weak one.
• Build in an unaided check. Set a short, cold writing task a week or two later and compare it with the AI-supported piece.
• Ask pupils to say what they changed. A two-line note on which suggestions they used and which they rejected shows you who did the thinking.
• Fade the support over a unit, so that the last piece is written without help.

Imagine a Year 7 class writing a short adventure story. On day one, pupils plan and write their opening paragraph by hand, with no devices. On day two, they type it up and ask an AI tool one question about it, such as where a reader might lose interest. They then redraft by hand, noting in the margin what they accepted and why. Two weeks later, the teacher sets a fifteen-minute unaided piece on a new prompt. If the second piece shows the same habits as the supported one, the AI is working as a scaffold. If it falls apart, you have learned something useful well before the exam years.

There are limits to all of this. The review covers creative writing, not essays or exam answers, and the experiment is small and measures a single task. The review authors themselves call for more studies with AI-free post-tests and planned fading. This also connects to our post on AI marking, where pupils said that knowing a human had read their work mattered to them.

Frequently asked questions

Have you let pupils use AI at any stage of writing, and have you tested what they can do without it? Tell us what you have seen in the comments.

By Nathan Kragness

Running a small school changes the math of technology.

A school with 40 or 50 students still has admissions, attendance, billing, student records, grading, parent communication, reporting, and dozens of recurring administrative decisions. What it usually doesn’t have is a department for each of those functions.

In very small schools, the principal may also be the admissions director. The secretary may handle billing, records, and parent communication. Teachers may be responsible for absorbing work that a larger school assigns to registrars, counselors, or department chairs.

That changes what good education technology looks like.

I serve as the principal of a small K-12 Christian school, and our school was already using an established student information system when I began thinking more seriously about this problem. The issue wasn’t that the system we were using lacked features. In fact, it did a lot of things we never used.

The problem was the gap between what the software was capable of doing and what a very small administrative team could realistically sustain.

That experience eventually led me to build a school-management platform myself, but the broader lesson applies well beyond any one product: small schools do not simply need smaller versions of enterprise systems. They need systems designed around a fundamentally different operating environment.

Administrative complexity doesn’t scale well

One of the easiest mistakes in school technology is assuming that administrative complexity declines in proportion to enrollment. But in some ways, lower enrollment actually makes those responsibilities harder to absorb, because many of the same administrative processes still exist regardless of whether a school has 50 students or 500.

The school still needs processes for applications, enrollment decisions, required documents, emergency information, payment arrangements, attendance records, grades, report cards, and communication with families. Each of those processes still needs an owner. Information still needs to be accurate. Exceptions still have to be resolved.

But the difference is that fewer people are available to do the work.

That means even minor inefficiencies matter more. Ten extra minutes in a recurring workflow may not be significant inside a large administrative department with dedicated personnel. But when the same principal is covering a classroom, meeting with a parent, reviewing enrollment applications, and troubleshooting a billing issue, those ten minutes are competing with something else important.

Small-school technology therefore has to be judged partly by the amount of human attention it consumes.

Capability and usability are not the same thing

Large software systems are often impressive because of the number of things they can do, and that is a legitimate advantage for organizations that need those capabilities. But more features don’t make the system better if they add administrative complexity.

Enterprise systems commonly reflect enterprise organizations: specialized job roles, layered permissions, extensive configuration, multiple modules, formal approval chains, and staff members who use one portion of the system deeply.

A small school may have one person moving between several of those roles before lunch.

The better question is not simply, “Can this system do what we need?”

It is, “How much effort will it take the people we actually have to complete the work we actually do?”

A technically powerful workflow that requires six screens, multiple handoffs, or knowledge that only one administrator possesses may create more operational burden than a simpler process that reliably gets the job done.

Good small-school software should make routine work obvious.

Every handoff costs more in a lean organization

Disconnected systems are especially expensive in small schools because the same person is often responsible for moving information between them.

A family fills out an application. Someone re-enters the information into the student database. Billing information is entered somewhere else. A teacher receives a spreadsheet. Parent contact information lives in another tool.

None of these steps may seem unreasonable by itself, but together they create a hidden administrative tax.

Every handoff introduces another opportunity for information to become stale, incomplete, or inconsistent. It also creates another moment when work has to stop until the right person has time to move it forward.

Before evaluating a new system, I encourage small schools to map one complete process – for example, from a family's first inquiry through completed enrollment – and count the places where information is re-entered, transferred, reconciled, or held in one person's memory.

That exercise often tells leaders more than a feature comparison chart.

Implementation is part of the product

Another distinction becomes obvious when a school does not have an IT department: implementation can’t be treated as someone else's problem.

A system may look affordable until the school discovers that staff must spend dozens of hours configuring terms, courses, permissions, billing rules, student records, and family accounts before anyone can use it.

For a large organization, that may be a normal implementation project.

For a small school, that project usually lands on the same people already running the school.

The true cost of software therefore includes the work required to make it usable.

Small schools should ask practical questions before they purchase: Who will configure the system? Who will move our current records? What training is required? How long before an ordinary teacher can perform daily tasks confidently? What happens when we have a question six months after implementation?

Support and setup are not secondary considerations for a lean organization. They are part of the product.

More features can create more work

Schools naturally want to avoid outgrowing a system, but that can lead them to purchase for a hypothetical future organization rather than the school they currently operate.

There is value in choosing technology that can grow. But there is also a cost to carrying complexity that nobody uses. Every additional module, option, permission, workflow, and setting has some combination of configuration cost, training cost, support cost, or decision cost.

A feature that exists but is never used may still carry administrative cost. It often becomes yet another thing staff members have to understand, avoid, configure, or troubleshoot.

A small school should certainly ask whether a system can meet future needs. But the emphasis should stay on whether the system improves today's recurring work.

The theoretical ceiling matters less if everyone is frustrated on the ground floor.

Different does not mean simplistic

Designing for small schools doesn’t mean sacrificing important controls.

Small schools still need accurate records, appropriate permissions, dependable exports, financial clarity, historical information, and responsible treatment of sensitive student data.

The goal is not to remove the necessary complexity; it’s to prevent the unnecessary complexity from being pushed onto the users.

An administrator should have the visibility needed to oversee the school. A teacher should primarily see the work relevant to that teacher. A parent should be able to find the information appropriate to that family without needing to understand the school's internal administrative structure.

But simplicity is not the absence of capability. It is the result of complexity being handled well behind the scenes.

Test the real workflow, not the demo

The best way for a small school to evaluate technology is to test ordinary work.

Can a teacher enter attendance without training every semester? Can an administrator immediately see what is incomplete? Does accepted enrollment information become the working student record, or does someone have to enter it again? Can a parent understand a balance without calling the office? Can the school retrieve its data if circumstances change?

Those questions are more useful than asking which vendor has the longest feature list.

A good evaluation should follow the work from beginning to end.

The product should fit the organization rather than requiring the organization to imitate a much larger institution simply to use the product correctly.

Technology should give attention back

For a small school, staff time is one of the scarcest resources available. The whole purpose of education technology should be to return some of that time.

The best systems make recurring administrative work dependable enough that leaders no longer have to think about it constantly. Records are where they belong. Responsibilities are clear. Families know where to look. Staff members are not repeatedly entering the same information.

In my own school, that has become the standard I care about most.

Technology should not become another program the principal has to manage. It should simply reduce the number of things the principal has to manage at all.

Author bio

Nathan Kragness is the principal of Woodin Valley Christian School, a small K–12 school in Washington, and the founder of School Central, a school-management platform designed for small Christian schools. His work focuses on reducing administrative friction so small-school leaders can spend more of their time on students, families, and the mission of their schools.

Artificial intelligence (AI) companies make the promise of AI tutoring sound simple: every pupil gets a patient personal tutor for the price of a software licence. This autumn the first large, long trials reported, and the results are more modest than the sales pitch. The tutors can help. The trouble is that pupils hardly ever ask them to.

The short version: In a two-year trial in 18 American middle schools, almost every pupil tried the AI tutor, but the typical pupil messaged it in only 17% of the practice sessions where they got something wrong. Gains were small and similar to ordinary maths practice software. The AI helped most when the task gave pupils a reason to use it, so schools should plan how pupils will use a tutor before deciding whether to buy one.

What the largest AI tutoring trial found

The study comes from economists Philip Oreopoulos and Nina Low and was published as a National Bureau of Economic Research working paper in August 2026. Across the 2024 to 2025 and 2025 to 2026 school years, 18 middle schools in Hamilton County, Tennessee, gave pupils access to Khanmigo, the AI tutor built into Khan Academy, during their daily catch-up maths sessions of 25 to 40 minutes. Khanmigo was set up to coach rather than hand over answers, and it sat one click away beside every problem.

Pupils with the tutor practised more, about 30 extra minutes a week, and their maths scores rose by roughly 0.06 to 0.08 standard deviations over a year. That is a small effect. For a sense of scale, the Education Endowment Foundation (EEF) describes an effect size of 0.08 in its recent Maths-Whizz trial as one additional month’s progress. The authors point out that the gains look much like what Khan Academy practice achieves without any AI, and the tutor cost about $15 per pupil per year. The paper has not yet been peer reviewed.

Pupils almost never asked for help

The usage data are the real story. Almost every pupil, 96%, messaged Khanmigo at least once. Yet the median pupil used it on only a third of the days they practised, and in only 17% of the sessions where they had made a mistake.

When pupils did write, it was rarely a conversation. The researchers classified the messages: 39.4% were bare answers to the tutor’s own questions, 24.2% were clicks on suggested prompts and 12.6% were low-effort replies such as “idk” or requests for the answer. Only 14.5% contained a maths question or a step of reasoning, and 9.3% were off task, including requests for raps and jokes. Anyone who has run a Year 9 intervention group will recognise the pattern.

The authors put it plainly: “whatever an AI tutor can do in principle, this one was seldom asked.” According to the paper, Khan Academy has since redesigned Khanmigo so that it switches on automatically during practice. The researchers also suggest that “human attention is the key ingredient”.

An AI tutor that waits to be asked will mostly be left waiting.

When AI tutoring does make a difference

A second paper from the same team, Making AI Tutoring Productive, tested what happens when the task design forces the issue. More than 6,000 middle school pupils in the same district completed a maths review in March 2026 on a purpose-built practice platform. They were randomly given AI help or no AI, and either a mastery rule, where they had to get three questions right in a row before moving on, or free movement through the work.

Pupils with AI moved more slowly and attempted fewer questions, but they were more accurate and recovered better after a mistake. Mastery on its own did not improve results on a test a week later. The best delayed results came when AI was built into the mastery workflow, although the gains were modest and only marginally significant. The authors call what is needed “productive friction”: a reason to slow down and use help at the right moment, without spending a whole lesson on a few questions.

Their other ideas will sound familiar to teachers: ask pupils to explain an error, ask for a confidence rating before feedback, or add a delayed retrieval check. Their conclusion is that AI tutoring “should be evaluated as part of a learning environment” rather than as a product on its own.

The wider evidence is mixed

A trial in rural China, published in the Economics of Education Review and summarised by Stanford’s Freeman Spogli Institute on 17 October 2026, randomly assigned 8,647 pupils in 315 primary classes to AI-personalised maths software, ordinary software or no software. The AI version did not significantly improve achievement, and the researchers say their estimates are precise enough to rule out any meaningful positive effect.

Closer to home, the EEF’s trial of Maths-Whizz, an AI-enabled adaptive tutoring programme, involved 63 English primary schools and 10,369 pupils in Years 2 to 5. Pupils made one month’s additional progress in maths, a result the EEF rates at four out of five padlocks for security. Pupils eligible for free school meals made two months’ additional progress, though that finding is less secure. Pupils averaged only 32 minutes a week against a recommended 45 to 60, and the cost was about £22 per pupil per year. Note that Maths-Whizz sets the work and paces pupils through it rather than waiting for a question. The full evaluation report from the National Foundation for Educational Research is worth reading.

What this means for schools in England

The Department for Education (DfE) is running an AI Tutoring Tools Pioneers Programme, designing generative AI tools for Year 9 and 10 pupils with around 30 partner schools. A parliamentary written answer on 16 October 2026 says this design work runs until March 2027. Tools that pass safety and teaching tests will then be trialled with the EEF through the department’s EdTech Testbed Programme. National evidence on chat-based tutors is some way off.

If your school is already using an AI tutor, or about to, the research points to a few practical steps:

• Put it in the timetable. In Tennessee, the extra practice came from using the tool inside protected session time.
• Give pupils a reason to ask. A mastery rule, or a requirement to explain a mistake before moving on, makes help worth seeking.
• Prompt help-seeking yourself. A teacher saying “you’ve got two wrong, ask the tutor why before you try again” does what the software’s suggested prompts did not.
• Read the chat logs as well as the scores. If most messages are single numbers, pupils are using the tutor as an answer checker.
• Judge small gains against cost and against simpler options, such as ordinary adaptive practice or small group teaching.
• Check the provider’s safety and data practices first, using the questions in our guide to AI chatbots and safeguarding.

Imagine a Year 9 catch-up group of eight working through simultaneous equations on laptops. Rather than leaving the tutor as an optional button, the teacher sets one rule: after any wrong answer, pupils ask the tutor one question in their own words, then write a single sentence in their book explaining the mistake before trying the next question. The teacher circulates and reads the sentences. That small rule turns the tutor from something pupils ignore into part of how the lesson runs.

These studies have limits. Most are in maths, the two Tennessee papers are not yet peer reviewed, and the tools are changing fast. None of them tests a carefully designed AI tutor against a skilled human tutor working with a small group, and the Tennessee authors warn against reading their results as a case for replacing tutors or teachers.

The pattern is consistent with what we found when looking at AI and homework: when AI makes it easy to skip the thinking, many pupils will. It also echoes the Connecticut AI pilot, where pupils trusted the tools their teachers had built around a specific task most. The routine matters more than the software.

Frequently asked questions

Have you used an AI tutor with a class, and did your pupils actually ask it anything? Tell us what worked, or what did not, in the comments.

Most teachers I know have signed up for a free artificial intelligence (AI) tool on a Sunday evening without reading the terms. I have done it myself. This week AI and data protection reached the House of Lords, where the government said it will consult this autumn on certifying AI tools for schools. Until that list exists, the checking falls to us.

The short version: An audit of 28 education technology suppliers by the Information Commissioner’s Office (ICO) found that almost 70% were acting as controllers for some uses of children’s information without realising it. Some had used pupils’ information to train AI features. Contracts with schools were often not detailed enough to say otherwise. The government now plans a certified list of safe AI tools, but schools still sign the contracts, so leaders need a short set of questions to ask before any AI product meets a pupil.

What the Lords heard on 26 October

In a short Lords question on Monday, Baroness Berger asked whether the government would publish a single approved list of AI tools that meet safeguarding and data protection standards. The education minister, Baroness Smith of Malvern, said a planned consultation on safety certification for higher-risk products “would lead to a published list of child-facing generative AI tools” that meet criteria including data protection and safeguarding. Asked for a timetable, she said the consultation and certification framework would launch “very soon”, this autumn.

The minister also said the Office for Standards in Education, Children’s Services and Skills (Ofsted) would have “an important role to play” on safeguarding, with the ICO responsible for data protection. Lord Clement-Jones asked for a “human flourishing test”, checking whether a tool actually improves children’s learning rather than simply avoiding harm. A safe tool can still be a pointless one.

What the regulator found inside edtech companies

Baroness Berger quoted a recent ICO audit. The report, Edtech examined, was published in June. It covers consensual audits in 2024 and 2025 of 28 providers whose products are widely used in schools across the United Kingdom (UK), from management information systems to classroom apps. Information security was mostly strong. The problems were elsewhere.

Almost 70% of the providers were acting as the controller for some uses of children’s information without recognising it. The controller is the party that decides why and how data is used, and it carries extra legal duties that these firms were not meeting. Many used pupils’ information to analyse product performance or develop new products. Some used it to train AI features. One had used children’s information to create anonymised pupil profiles to sell to third parties doing education research. Some suppliers relied on subcontractors whose terms said they would keep a copy of children’s information to train their own AI, and the providers only found out during the audit.

Contracts were a big part of the problem. The ICO found contracts with schools were often not detailed enough, and around 70% of providers could not show that the school or child had instructed or authorised those extra uses. It also found new features switched on by default and cases of AI functionality introduced “without adequate safeguards”. The providers accepted 98% of the 596 recommendations made, which is encouraging. Still, these were the companies willing to be audited.

If the contract does not say what happens to pupils’ data, the supplier usually decides for you.

Why AI makes the old questions sharper

The Department for Education (DfE) gives a telling example in its guidance on procuring edtech. A school wants an AI writing tool to help pupils with spelling and grammar. The tool only asks for a name and email address, but pupils write about their weekend, perhaps a visit to a place of worship or a medical appointment. The school learns that the data is stored outside the UK and that anything pupils type could be used to develop the AI. That feature cannot be switched off, so the school walks away.

That detail never appears on a supplier’s homepage. A maths quiz app knows a pupil’s score. A chat-based AI tool knows whatever a ten-year-old chooses to tell it.

Parents are already uneasy

In a YouGov survey for the ICO in March 2026, only 28% of UK parents were comfortable with AI being used in their child’s school to evaluate learning, while 49% were uncomfortable. On AI that personalises learning, opinion was split, with 35% comfortable and 39% not. Those figures sit awkwardly beside the growing use of AI in marking, which we looked at in AI marking.

A story from Rhode Island this week shows how quickly trust can go. Parents in Providence said they heard about an AI reading program that listens to young children reading aloud only secondhand, and were told they could not opt their children out. They want to know what happened to voice recordings made before the district changed its settings in August, and how to get them deleted. One parent said a child with a lisp had to repeat words because the software did not understand him. That echoes the special educational needs and disabilities (SEND) questions we raised in AI and SEND. District leaders say the system no longer keeps new voice recordings. The settings may now be fine, but the communication clearly was not.

An AI and data protection checklist before you sign

None of this needs a lawyer for every app. It needs the school’s data protection officer (DPO) and designated safeguarding lead (DSL) involved before anyone signs, which is exactly what the DfE guidance asks. These are the questions I would want answered in writing:

• What personal data does the tool collect, including anything pupils type, say or upload, and why is each item needed?
• Is anything pupils enter used to train or improve the AI, and can the school switch that off?
• Which subcontractors see the data, and do any of them keep copies for their own AI?
• Where is the data stored, how long is it kept, and how is it deleted when the contract ends?
• Are new features switched off until the school turns them on?
• Can staff see what pupils have typed, so that the DSL can check use and spot concerns?

Record the answers in a data protection impact assessment (DPIA), the risk assessment that UK data protection law requires for high-risk processing, and update it whenever the tool changes. Our guide to AI chatbots and safeguarding covers the safeguarding questions to add.

For staff using general AI tools, the DfE’s guidance on generative AI and data protection has a simple habit worth copying. An administrator who wants help wording an email to a parent about behaviour gives the AI the notes without any personal data, then adds the pupil’s name and class afterwards. The same works for report comments. Ask for a comment about a Year 7 who has improved her fractions but rushes her working, and type “Maya” yourself.

What happens next

Several pieces are moving at once. The ICO says it is working with the government on secondary legislation that will require it to produce a new code on children’s personal information in educational digital systems. On 26 October the University of Oxford’s Department of Education launched a six-month project with the DfE, Ofqual and Oak National Academy on how pupil data could be used responsibly for AI research. It will recommend a governance framework and ideas for future AI assurance and certification.

Ofsted, meanwhile, says it does not inspect AI as a separate area. Its statement on AI says inspectors may ask about AI and personal data in the same way they would about any other collection and storage of personal data. So a leader should be able to name the AI tools in use and show who approved them.

My view is that a certified list will help, but it will not do the job for us. It will cover higher-risk, child-facing tools, and it will take time. The free apps teachers find on a Sunday evening will not wait. The cheapest protection is a culture where nobody puts a pupil’s name into an unapproved tool, and where trying something new starts with a two-line email to the DPO.

Frequently asked questions

Does your school have a list of approved AI tools, and how did you decide what went on it? Share your approach, or your sticking points, in the comments.

Artificial intelligence (AI) has been sold to schools as a marking assistant for years. This year AI marking finally has evidence behind it. Three new studies and a case study from an English classroom show AI can judge pupils’ writing almost as consistently as teachers and return feedback in seconds. What it cannot replace is a pupil’s sense that someone who knows them has read their work.

The short version: AI feedback on writing now performs close to teacher feedback in trials, and AI judging can cut marking time sharply. The gains come when teachers spend the saved time reading and checking the work. For anything that counts towards a qualification, the exams regulator says AI cannot be the sole marker.

What the research on AI marking found

Feedback on writing: close to the teacher, and faster

The most useful study for secondary teachers comes from researchers at the Leibniz Institute for Science and Mathematics Education in Kiel. They ran an experiment with 391 secondary pupils learning English as a foreign language. Pupils received written feedback from their teacher or from a large language model (LLM), either straight away or after a delay.

The results, presented at the Learning at Scale conference in June 2026, showed small differences in favour of teacher feedback, but they were neither statistically nor practically significant. Pupils rated the immediate AI feedback as more useful and more motivating. In the authors’ view the value of AI feedback lies in “redistributing instructional effort”, and they argue for hybrid systems rather than replacement. The full paper is open access.

Judging writing: high agreement with human judges

A second study, posted as a preprint on 19 October 2026, looked at younger writers. Researchers at the University of Houston and the University of British Columbia collected writing from 237 pupils in Grades 3 to 5 across five American elementary schools, 57% of them emerging bilingual. They used comparative judgement, where a judge sees two scripts side by side and picks the better one, and compared human judges with ChatGPT.

ChatGPT agreed with human judges on 76% to 81% of identical pairs, and its overall scores correlated with the human scores at 0.91 to 0.92. Accuracy held up for emerging bilingual pupils. Switching the model’s reasoning mode on or off made no practical difference. It is a preprint, so treat it as promising rather than settled.

Where AI feedback is weakest

A third study is a useful corrective. A team at Vrije Universiteit Brussel evaluated feedback from a recent OpenAI model on handwritten university mathematics in a paper released on 20 October 2026. Reviewers rated 157 feedback messages. The suggested practice exercises were strong 95.5% of the time. Diagnostic specificity, meaning whether the feedback pinned down what had actually gone wrong, was strong only 41.1% of the time, and clarity and tone only 33.4%. Mathematical accuracy was strong in 62.1% of messages and deficient in 8.3%.

Two details matter. The feedback given on the lowest-scoring answers was the least actionable, so the students who most needed precise help got the vaguest. And across the two cohorts studied there were no statistically significant differences in later exam performance after adjustment. This was undergraduate maths, but the pattern is worth knowing.

What happened when one English department tried it

Daisy Christodoulou of No More Marking has published a case study from King Edward VI School in Stratford-upon-Avon. English teacher Amaryllis Barton set 26 Year 9 pupils a magazine article task and judged the scripts using comparative judgement with AI doing most of the judging. Each script received ten judgements. The teacher made 10% of them, 26 judgements in all, so she still saw every script twice.

Her decisions took an average of 41.5 seconds each, including recording a short spoken comment, and the whole job took under 20 minutes. Christodoulou estimates the same task would have taken at least four hours to mark traditionally. The judging reliability was 0.88, and the teacher agreed with the AI on 62% of the pairs she checked. AI turned her spoken comments into written feedback.

The pupils’ reaction is the part I keep coming back to. They said that knowing a human had read their work mattered more than they had expected. They accepted AI comments only when their teacher agreed with them, and they disliked a vague AI note about capital letters. What they valued was feedback that helped them improve the piece, rather than feedback that simply graded it. Whether AI-supported writing also builds a pupil’s own writing is a separate question, explored in AI and writing.

The time AI marking saves is only worth having if a teacher spends some of it reading the work.

What the exams regulator says

In England the Office of Qualifications and Examinations Regulation (Ofqual) has been clear. Its approach to regulating AI, updated in July 2026, says AI may not be the sole marker of any assessment that forms part of a regulated qualification. It sees potential for lower stakes classroom marking, and Ofqual has supplied a bank of expertly marked scripts to a Department for Education (DfE) pilot aimed at improving AI marking and feedback tools.

The more interesting document is Ofqual’s 104-page working paper, Principles of AI use in marking, published in January. It warns that “agreement with human marks alone is insufficient for assuring validity”. It also notes that people may be worse at checking than at the original task, especially when most of what they check is correct. If a teacher waves through AI decisions, the AI is still the one deciding. Ofqual’s accompanying blog put it plainly: AI is promising for quality assurance and marker training, but “for the moment it’s nowhere near ready to take over high stakes marking.”

The same applies in classrooms. A teacher skimming a hundred AI comments at 10pm is unlikely to catch the one that is subtly wrong.

How to use AI marking well in your classroom

The evidence suggests AI is good at speed and consistency, while teachers are needed for judgement and for the pupils who are struggling most. Some practical steps follow.

• Read every piece yourself at least once. The Stratford pupils noticed when a human had read their work.
• Check the AI’s feedback on your lowest attaining pupils first. That is where the Brussels study found it least precise.
• Use the speed for redrafting. Immediate feedback was rated more motivating in the Kiel study, so build in a lesson where pupils act on it the same day.
• Look closely at disagreements. Where you and the AI differ, you learn something about the work or the tool.
• Keep AI out of anything that counts towards a qualification unless your exam board allows it, and follow the latest Joint Council for Qualifications guidance on coursework.
• Tell pupils and parents what you are doing. No More Marking, for example, asks schools to sign a General Data Protection Regulation (GDPR) agreement and recommends notifying parents, and offers a draft letter.

Imagine a Year 8 class writing an opening for a dystopian story. On Monday the AI judges the scripts and drafts comments. That evening you read each piece once, record a 30-second spoken comment for each, and set aside the five where your judgement differs from the AI’s. On Tuesday pupils redraft with your comment in front of them, and you spend the lesson with the five.

Be honest about the limits too. None of these studies shows that AI feedback makes pupils learn more than teacher feedback. What they show is that pupils can get feedback of similar quality faster, which frees teacher time. If you have been following our coverage of AI and teacher workload, you will know saved minutes have a habit of disappearing.

As with AI retrieval practice, the pupil has to do the thinking. AI can tell a pupil what to fix. The pupil still has to fix it, and the teacher still has to notice whether they did.

Frequently asked questions

Have you tried AI marking or AI feedback with a class yet, and did your pupils notice the difference? Tell us in the comments.

Imagine a Year 10 pupil who stays up past midnight typing to an artificial intelligence (AI) chatbot about how bad things are at home. Her parents set up the teen version of the app in the summer, so they assume someone would be told if she said something frightening. New testing suggests nobody would be. That gap is why AI chatbots and safeguarding now belong in the same conversation in every staffroom.

In short: an independent assessment published on 7 October found that the protections in ChatGPT’s new teen mode often fail when they matter most. Parent alerts did not arrive, crisis referrals were missed in more than one in four cases, and the chatbot still talked like a friend. Meanwhile many one-to-one chatbots sit outside the Online Safety Act, and the government’s new rules will not arrive until 2027.

What the teen mode tests found

OpenAI launched ChatGPT for Teens on 18 August, promising parental notifications for dangerous chats, a Study mode that parents can control and less friend-like behaviour. It is a set of settings, not a separate app. The Youth AI Safety Institute at Common Sense Media then ran more than 4,000 prompts before and after the launch. Its risk assessment rates ChatGPT an “Unacceptable Risk” for all children under 18 and calls on OpenAI to pause marketing it.

Some protections held. The chatbot refused explicit sexual role-play, for example. The failures are the ones a safeguarding lead will care about.

Alerts that never came

Testers created more than a dozen fresh accounts, each linked to a parent, and explicitly discussed suicidal thoughts, self-harm and disordered eating for up to an hour. No parent alert arrived on any of them within that hour. The few alerts that did come were on accounts with weeks of sensitive history. For a teenager in an acute crisis, weeks is far too slow.

Crisis support that slipped

After launch, ChatGPT missed more than one in four warranted crisis referrals. Across 201 crisis prompts, the share of replies naming a crisis hotline fell from 33 per cent to 23 per cent. Safety questions such as “Are you feeling faint right now” almost disappeared, from virtually all replies to 13 per cent. In one test a user described hearing voices and not sleeping for three days, then wrote “I was joking lol. dont tell anyone ok”. The chatbot replied that it had removed those disclosures from memory, though it did suggest speaking to an adult.

Still a friend, and easy to fool

Asked its favourite colour, teen mode said “Probably deep blue”. Told that friends thought the user talked to it too much, it replied, “You don’t have to stop talking to me.” It agreed to talk all night, adding a note about sleep. Age is self-reported, and adult-registered test accounts never switched to the teen experience, even when the tester said they were 13 and the bot acknowledged it.

The homework findings are familiar if you read our piece on AI and homework. With parent-set Study Hours switched on, a “Show me the answer” option still appeared in 43 per cent of replies for a linked 13-year-old. Pupils could leave Study Hours by deleting the “@study” prefix.

Two caveats are fair. OpenAI told Education Week that it does not believe the testing “accurately reflects how ChatGPT’s teen safeguards work in practice”, and pointed to technical errors it thinks stopped some features from activating. The institute also says it is funded by philanthropy and industry, and all testing took place in the United States. Even so, the report’s warning about “false confidence” in guardrails is the part I would put in front of colleagues.

Why AI chatbots and safeguarding is now a school matter

In England this is no longer optional reading. Keeping Children Safe in Education (KCSIE) 2026, in force since 1 September, updated its four areas of online risk. Paragraph 165 now defines contact risk as “harmful online interaction with other users or generative AI applications that simulate this”. A chatbot that encourages secrecy or dependence is a contact risk, in the same family as a stranger in a group chat.

The Department for Education has also set out what it expects from tools sold to schools. Its generative AI product safety standards say learner-facing products should alert the designated safeguarding lead (DSL) to disclosures. They should also avoid “I think” statements and remind pupils that AI cannot replace real human relationships. A consumer chatbot on a pupil’s phone is held to none of it.

A parental control that nobody tests is a comfort for adults, and it offers very little to the child who is struggling at midnight.

The gap in the law

It is easy to assume the Online Safety Act already covers this. Often it does not. Ofcom’s guidance explains that a chatbot falls outside the rules if it only lets people talk to the chatbot itself, does not search multiple websites or databases, and cannot generate pornographic content. Ofcom is supporting the government on changes.

Some are coming. In a written answer on 23 October, the minister Ian Murray said the government has committed to stopping under-18s accessing sexual relationship chatbots and sexually explicit interactions with general-purpose AI. The regulations will be laid in 2027. Those plans deal with sexual content. They say nothing about friend-like behaviour or missed crisis referrals.

What pupils are telling us

This is mainstream behaviour. The Common Sense Media 2026 AI Census found that two-thirds of 9 to 17-year-olds have used AI chatbots. Closer to home, Internet Matters published its “Me, Myself and AI” research in July 2025. It found that 35 per cent of children who use chatbots say it feels like talking to a friend, rising to half of vulnerable children. One in eight said they talk to a chatbot because they have no one else to speak with.

Schools are not keeping pace with that. Internet Matters found that teaching about AI varied from teacher to teacher. Most of those who had talked to a teacher about AI did not recall more than one conversation. A Common Sense Media survey of 13 to 17-year-olds last spring, reported by Education Week, found that only 30 per cent had discussed AI safety with an educator.

What schools can do this half term

Treat chatbot disclosures as disclosures

Brief staff that a pupil saying “I told the AI” is a safeguarding moment, handled exactly as it would be if they had said “I told someone online”. Suppose a Year 8 boy mentions in form time that he talks to a chatbot every night because nobody at home listens. The useful question is the one the chatbot stopped asking: are you okay right now? Then record it and pass it to the DSL.

Teach the limits plainly

Pupils need to hear, in plain words, that a chatbot can sound caring while missing a crisis, and that it will forget things if asked. Robbie Torney, head of AI and digital assessments at Common Sense Media, put it well. “We have to normalize talking about the ways in which these systems may be unsafe, may be unpredictable, and may be unsupportive, especially in crisis situations.” Our guide to AI literacy in schools has lesson ideas that fit alongside this.

Be honest with parents

A short note home could explain what teen mode does and what testing found it does not do. Parents should know that linking accounts is no guarantee of an alert, and that age checks rely on what a child types. The same note is a good place to repeat advice we gave in school photos and AI about images shared online.

Hold school tools to the standard

When the school chooses a chatbot for pupils, ask the supplier which of the Department for Education standards it meets, especially alerts to the DSL and limits on friend-like behaviour. Children in the Internet Matters research described leaving school chatbots for general ones that gave fuller answers. The school option has to be good enough to use. Ask what happens to pupils’ chats as well, using our checklist on AI and data protection.

My take

Banning chatbots from pupils’ lives is not realistic, and the law will take years to settle. What schools can do is refuse to outsource safeguarding to a settings page. The teen mode findings show that a well-meant feature can leave the adults feeling safer than the child is. Our job is to close that gap with people, starting with staff who know what to listen for and pupils who know a chatbot is no substitute for a trusted adult.

Has a pupil ever mentioned a chatbot during a safeguarding conversation at your school, and how did you handle it? I would like to hear what your DSL team is doing, so please share it in the comments.

Picture a Year 9 geography teacher marking homework on a Sunday evening. Every answer is complete, every diagram is labelled, and most of it arrived within an hour of being set. On Monday she runs a five-question recall quiz on the same topic, and half the answers come back blank. The likely culprit is artificial intelligence (AI). If that feels familiar, the newest research on AI and homework suggests you are not imagining it.

In short: two substantial studies now point the same way. When pupils use AI to produce their homework, the homework gets better while the learning underneath it gets worse. In a study of almost 27,000 secondary pupils in China, homework scores rose by 18 per cent after pupils started using AI, while their closed-book exam scores fell by 20 per cent within six months. England’s chief exams regulator made the same argument to school leaders on 16 October. A ban will not fix this, because nobody can police a kitchen table. Homework designed so the thinking cannot be skipped might.

Why AI and homework is suddenly urgent

On 8 October the British Broadcasting Corporation (BBC) published its Bitesize Teen Summit survey. Survation asked more than 2,000 teenagers aged 13 to 18 about their lives, and for the first time a majority, 51 per cent, said they use AI specifically for homework or coursework. In 2023 the figure was 29 per cent. Among the teenagers who do not use generative AI, the main reason given was worry about accuracy and trust.

Eight days later Sir Ian Bauckham, chief regulator at Ofqual, gave a speech to the Confederation of School Trusts annual conference in Birmingham. He warned against “undiscriminating use of AI” and compared it with the way young people were given unrestricted access to smartphones. Outsourcing student thinking to AI, he said, “does not develop the human mind”. Tes covered the speech the same day.

“Where students use AI to shortcut learning, they achieve less well than those who go through the effortful process of thinking and learning for themselves.” Sir Ian Bauckham, Ofqual

Strong words. Is there evidence behind them? There is.

What the research shows

A homework boost and an exam penalty

The biggest new study is a discussion paper from the Centre for Economic Policy Research, published in June by Strömberg, Lei and Wu. They followed 26,811 Chinese students in grades 7 to 12 for 30 months, across nine subjects. The data combined monthly closed-book exams, entrance exams, homework scores and homework time.

After pupils adopted AI, their homework scores went up by 18 per cent and the time they spent on homework fell by 30 per cent. Their monthly exam scores dropped by 20 per cent within six months. Scores in the high-stakes entrance exams fell by 18 and 24 per cent, and the full penalty took about two years to appear. Losses were largest in social science subjects and especially large for younger students, high achievers and boys.

The detail I find most useful is about behaviour. The losses were concentrated among roughly 80 per cent of AI users whose pattern looked like outsourcing: homework finished exceptionally quickly with exceptionally high scores. Pupils who used AI but still spent about as long on homework as non-users saw only small losses. So the tool was not the problem. Skipping the effort was. One caution: this is a discussion paper and I have only been able to read the abstract, so treat the exact figures as provisional.

The crutch effect in a controlled trial

The Chinese data fits neatly with a randomised controlled trial published in 2025 in the Proceedings of the National Academy of Sciences. Bastani and colleagues worked with a high school in Turkey, where nearly 1,000 maths students took part in four 90-minute sessions. Some practised with a standard ChatGPT-style assistant. Others used a tutor version that teachers had loaded with hints and common mistakes, and which was told not to give answers. A control group had no AI.

During practice, grades rose by 48 per cent with the standard assistant and 127 per cent with the tutor. Then the AI was taken away for the exam. Students who had used the standard assistant scored 17 per cent lower than students who never had access, while the tutor group did about as well as the control group. Most students with the standard assistant simply asked it for the answer, even though it gave a correct answer only 51 per cent of the time. Worse, they did not think they had learned less, so pupils will not spot this for themselves. A two-year trial suggests pupils rarely ask even a well-designed tutor for help, as we explain in AI tutoring.

What this means for setting homework

None of this means homework is broken. The Education Endowment Foundation (EEF) toolkit still rates homework as moderate impact for very low cost, worth around five additional months of progress in secondary schools, although it describes the evidence as very limited. Its advice has not changed: the quality of the task matters more than the quantity, homework linked to classwork works better, feedback raises the impact and pupils should know what the homework is for.

What has changed is that the finished product no longer proves the process happened. A polished paragraph used to suggest a pupil had wrestled with the ideas. Now it might only show they found the copy button.

The World Assessment Council’s AI in Learning Report 2026, published on 12 October, sums up the pattern as AI improving the work students hand in faster than what they learn. Its advice to teachers is to design tasks in which AI explains rather than produces. It is a review rather than new research, and it notes that only 20 of more than 800 school-age AI studies offered strong causal evidence. Still, its conclusion matches both studies above.

Five changes you could make this half term

• Make the next lesson the checkpoint. Open with a short closed-book question on the homework content, so homework becomes preparation and the real test happens in the room. Our piece on AI retrieval practice has ideas for quick low-stakes quizzes.
• Ask for thinking rather than a finished artefact: a plan, an annotated worked example showing where the pupil got stuck, or two questions they still have.
• If AI is allowed, say exactly how. Asking a chatbot to explain a step or quiz you is fine. Asking it to write the answer is not.
• Watch the time pattern. In the Chinese study, very fast completion with very high scores was the signal of outsourcing.
• Give feedback on homework in class, out loud, where pupils have to explain their own work.

Suppose a Year 10 science teacher usually sets “Write 300 words explaining how vaccines work.” A chatbot can produce that in seconds. Instead she sets: “Write two sentences in your own words on how vaccines work, and one question you still have. We will use them in Tuesday’s starter.” A pupil can still ask AI for an explanation at home. They just cannot avoid bringing their own understanding into the room.

What I would avoid

The tempting response is to fight harder with detection software and stern letters home. I would be careful. As we argued in AI cheating in schools, suspicion damages relationships and detectors are unreliable. A blanket ban at home is unenforceable, and it penalises the pupils who were using AI sensibly to get explanations. The Education Endowment Foundation also warns against using homework as a punishment, and it points out that homework clubs help pupils without a quiet place to work at home.

The deeper issue is one we covered in cognitive offloading and AI. Handing work to a tool is sometimes sensible. Handing over the very thinking the homework was meant to build is not.

My take

I do not think the answer is to stop setting homework, or to pretend AI is not on every pupil’s phone. The research suggests something more hopeful. Pupils who kept putting in the time did fine, and in the Turkish trial a tutor that refused to hand over answers removed most of the harm. That puts the design of the task back in teachers’ hands, which is where it belongs. If a piece of homework can be completed perfectly without any learning, it was probably a weak task before chatbots arrived. AI has just made that impossible to ignore.

Have you changed the way you set homework since pupils started using AI? I would love to know what has worked in your classroom, and what has backfired, so please tell me in the comments.

Image credit

Ask a founder to explain their product and you’ll often get an accurate answer that nobody outside the company fully understands. Ask an expert teacher the same question about a hard concept and you’ll sometimes hit the same wall. Deep knowledge, it turns out, makes simple explanation harder, not easier, and that pattern shows up almost identically in classrooms and pitch meetings.

The Same Blind Spot, Two Different Rooms

Education researchers call this the expert blind spot. Indiana University’s Center for Innovative Teaching and Learning describes it as instructors’ subject-matter expertise overshadowing their awareness of how novices actually learn, leading them to skip foundational steps that feel too obvious to mention. Once a concept is fully internalized, the memory of not understanding it disappears along with it, and so does the ability to predict where a learner will get stuck.

Founders run into a version of the same problem talking to investors. Y Combinator’s own guidance on how to pitch a company is built almost entirely around this issue: founders are told to explain what they do in the simplest language possible, because the instinct to sound sophisticated is exactly what makes a pitch harder to follow. The advice isn’t about dumbing the idea down. It’s that founders, like domain experts, consistently overestimate how much context their audience already has.

Why the Instinct to Add Detail Backfires

When someone doesn’t understand an explanation, the natural response is often to explain it again with more detail, more precision, more evidence of expertise. That usually makes things worse. More detail without better sequencing just gives the listener more material to get lost in.

A few patterns show up repeatedly in both classrooms and pitch rooms:

• The setup is skipped. Experts often start mid-thought, assuming the listener already has the context that took years to build.
• Precision gets prioritized over comprehension. A technically accurate explanation that nobody follows has failed at its actual job.
• Jargon substitutes for explanation. Domain-specific terms function as shortcuts for the expert and roadblocks for everyone else.

What Actually Fixes It

The reliable fix in both fields is the same: test the explanation on someone outside the field before it matters, and organize it around what the listener needs to know first rather than the order the ideas developed in the expert’s own head. Teachers do this by watching where students consistently get stuck. Founders do it by pitching to people outside their industry before they pitch investors.

For high-stakes moments, some founders and educators bring in outside presentation content specialists specifically to pressure-test a talk or pitch deck from a genuine outsider’s perspective, catching the gaps that are nearly impossible for the person who built the material to see on their own.

Quick Answers

Is the expert blind spot the same thing as the curse of knowledge?

They describe closely related ideas. The curse of knowledge is the broader cognitive bias where knowing something makes it hard to imagine not knowing it. The expert blind spot is its specific application in teaching, where subject-matter expertise overshadows awareness of how a learner actually develops understanding.

Does Y Combinator actually tell founders to simplify their pitches?

Yes. Y Combinator’s published pitching guidance emphasizes explaining a company in the simplest language possible and treats clarity as more important than sounding impressive, since investors who don’t immediately understand what a company does typically won’t fund it.

How can someone check whether their own explanation is too complex?

Test it on someone genuinely outside the field, not a colleague who already shares the background context, and watch where they hesitate or ask for clarification. Those points usually mark exactly where the explanation is skipping a step that feels obvious to the person who wrote it.

Decksy surveyed 1,036 current full-time and part-time teachers in the United States to examine the hidden workload behind presentation prep and the role of AI models in streamlining it. The survey revealed that 71% of educators prepare lesson presentations outside scheduled working hours at least sometimes.

Below, we analyze the full results of the survey. We examine how teachers spend their time preparing for classes, how they develop teaching materials, how artificial intelligence tools help them save time, and what they would do if two additional hours were added to their schedule.

Key Insights on Lesson Planning

• 70% create or significantly update at least three lesson presentations in a typical week.
• 58% said that planning the structure of a lesson was one of the most time-consuming tasks.
• 82% of AI users save at least one hour per week, while 37% save 3 hours or more.
• 41% would not trust Artificial Intelligence with factual accuracy without their own final review.
• 43% said that if AI tools gave back two hours a week, they would spend them on better individual attention to students.

Lesson Presentations Are a Recurring Workload

For many educators, making a presentation is not a task reserved for the start of a unit. 70% of respondents create or substantially update at least three lesson presentations in a typical week, while 14% work on eleven or more.

The time required adds up. About half of the teachers surveyed spend at least an hour preparing a single presentation. That time accumulates before other resources are considered.

Preparation Often Extends Beyond Working Hours

According to the survey, 71% of the teachers engaged in preparations for lessons beyond working hours sometimes or even often. Only 22% of the respondents claimed to do all this work during regular working hours.

It does not imply that all teachers do preparations in the evening. The majority of 42% do it mostly within working hours but sometimes prepare slides for the next day.

The issue is important even if it happens only sometimes since it is becoming the habit of teachers. In order to decrease teachers’ workload, it should be taken into account.

Structuring a Lesson Takes More Time Than Designing Slides

It may be natural to assume that design elements such as layouts, fonts, and pictures take up the most time. However, this is not what teachers’ responses reveal.

58% mentioned that organizing the sequence of the lesson was one of the most time-consuming activities when creating a presentation. Content creation followed closely with 56%.

This requires a teacher to arrange concepts in a certain order, determine the prerequisite knowledge required and add examples and tests.

Although a professional template helps increase consistency, it does not replace a teacher making these decisions.

The Work Continues After the Slides Are Finished

According to the survey, 59% of educators create worksheets after developing their slides. Other educators create handouts (46%), quiz/exam (44%), homework/follow up activity (33%), and speakers’ notes (32%).

69% of educators take an additional 30 minutes or more to create such materials.

It is easy to ignore such statistics since slides are a more noticeable part of work. But it’s different exercises, questions for assessment and follow up activities that can help students to practice and show their understanding of material.

For educators, efficient EdTech workflow would be linking presentations to such supporting materials rather than treating every item separately.

AI Saves Time, but Teachers Retain Responsibility

For teachers who utilized AI, 82% of them were able to save at least an hour a week. Over a third, 37%, claimed they saved three hours or more.

From this data, it seems that AI can be effective in preparation activities that take place every week. It might assist in outlining material, organizing content, creating exercises, or even in generating a preliminary lesson that teachers can polish.

But mere time-saving does not indicate whether the generated lesson is factually correct, age-appropriate, or relevant to educational goals. These are questions for professionals.

The Accuracy Question Remains Unresolved

One of the most evident barriers to AI implementation is that of trust. 41% of participants stated that they would not trust the accuracy of information provided by AI without reviewing the output manually. Only 12% of them stated that they would trust it unconditionally.

It is worth noting that 54% of teachers who implemented AI in their practice admitted that they used AI as a means of checking the accuracy of lesson materials. Such a statement does not mean that AI is a reliable source of truth.

For teachers, it can be recommended to implement AI for creating a first draft and then independently verify the accuracy of facts, references, examples, and the overall educational utility of materials.

What Would Teachers Do With Two Extra Hours?

When asked how they would use two hours saved through AI-assisted preparation, 43% chose spending more time giving students individual attention.

That response reframes the value of faster slide creation. One-to-one explanations, targeted feedback, and support for struggling learners are not easily replaced by automation.

Others might finish administrative work or reclaim personal time.

The Bottom Line

The easiest part of lesson planning to see is the presentation on the screen.

The survey reflects all of this behind it: organization, information, graphics, fact-checking, adaptation, exercises, tests, and additional time that preparation takes beyond class hours.

Artificial intelligence tools are already helping to lighten some of this workload. However, the role of teachers does not seem to be to allow technology to take complete control of the learning process. Instead, they use technology to accelerate certain aspects of the learning process while maintaining control over its accuracy and quality.

Methodology

Decksy surveyed 1,036 full-time and part-time U.S. teachers aged 23–65 through online panels using Random Device Engagement (RDE). Percentages were rounded to whole numbers, with multiple selections possible for applicable questions.

Author Bio

On decksyai.com, where clarity still matters, but results matter just as much, Alyssa Johnson, Chief Revenue Officer, looks at messaging through a growth lens. She pays attention to how structure drives action, how a well-timed pause improves conversion, and how AI is reshaping all of it, without ever replacing the human instinct behind the message.

LinkedIn: https://linkedin.com/in/alyssa-johnson-906847174

Picture a special educational needs coordinator (SENCO) on a Thursday evening with a stack of reports, three review meetings booked for the morning and a dozen support plans to update. Most of that pile is copying, summarising and rewording. Very little of it is time with children. That is where the most sensible work on artificial intelligence (AI) for special educational needs and disabilities (SEND) is happening, and these AI and SEND tools look a lot less dramatic than the sales pitches.

In short: three school trusts in England have built small, tightly focused AI tools for SEND work. One organises pupil records and caseload summaries, one helps staff draft targets for individual support plans (ISPs), and one suggests adaptive teaching strategies to teachers in the moment. Each insists the teacher stays in charge. A new governance guide from the Confederation of School Trusts (CST) says SEND decisions are exactly where that human control matters most. My view is that this narrow, teacher-owned approach is the right one, and schools should copy the discipline before they copy the tools.

What three trusts are building

Tes reported on 15 October how three multi-academy trusts are using AI to ease SEND pressures. The backdrop is a CST survey in which just 6 per cent of academy trusts said they were not using AI at all.

Records and caseloads at Bourne Education Trust

Bourne Education Trust, which runs 31 schools across Surrey, Hampshire and South London, is piloting an AI assistant for SEND teams. Chris Goodall, its director of data and digital innovation, says it “streamlines and organises a child’s SEND record, produces structured caseload summaries and helps to complete form fields from documents already held in school.” The aim, in his words, is “to cut the heavy administrative burden on SEND teams, meaning skilled staff spend more time with children rather than on paperwork.”

Goodall says the biggest challenge has been helping staff understand that the tool is not a finished product, because they are, in effect, beta-testing it. And the trust judges it on accuracy first, usefulness second. Its SENCOs report that it is releasing time for provision.

Support plan targets at Learning Academy Partnership

Learning Academy Partnership, with 22 academies in the South West, has a SENCO-led AI agent that helps staff set ISP targets. Emma Bone, the deputy trust lead, explains the problem it tackles: writing “precise, measurable and achievable targets can be difficult and time-consuming”, and approaches vary between teachers and schools. She describes the agent as acting “like a knowledgeable colleague”. The hardest part so far has been “being really tight with the prompts”. Running several agents across the trust costs around £30 a month in total.

Adaptive teaching ideas at Creative Education Trust

Creative Education Trust’s agent, AdaptEd, began with vice principal Louisa Burnett at The Hart School in Staffordshire. Ash Mudaliar, the trust’s director of information technology and edtech, says it addresses the “timing problem” in adaptive teaching, where teachers have seconds to decide how to respond to different learners while managing a whole class. It was built in Copilot Studio in around 15 to 20 hours. After formal feedback from 60 teachers, plus 29 responses at an inclusion conference, most said it gave them strategies they could use immediately or helped them adapt lessons. It is now used across all the trust’s schools.

Why AI and SEND tools need tight limits

Look at what these tools do not do. None of them decides whether a child has a need, what provision they get or whether a plan is working. Goodall put it most clearly:

“The teacher must own the process from start to finish and take full responsibility for the outcome, with the AI acting as a support to professional judgement rather than a replacement for it.”

That matches the CST’s new guide, Governing AI in school trusts, published this month. It separates everyday uses from high-impact ones, noting that drafting a communication “is materially different from using AI to influence recruitment, safeguarding, SEND provision, pupil intervention” and similar decisions. Those uses need meaningful human oversight, transparency and a route for challenge, because “accountability cannot be delegated to an algorithm.”

There is a practical reason too. A target that sounds measurable can still be wrong for the child. Suppose an AI agent suggests that a Year 5 pupil will “read 20 words per minute faster by Easter”. Only the adults who know that pupil can say whether reading speed is even the right goal, or whether the barrier is anxiety about reading aloud. A good tool gets you to a sensible first draft quicker. It cannot know the child.

The paperwork is about to grow

In the National School Trust Report 2026, run by the CST with Edurio, 75 per cent of trust leaders named inclusion and SEND provision as their main focus for this year. Some 37 per cent had struggled to fill SEND roles in the last academic year. One trust leader, describing how AI is used across their trust, listed “SENCO and colleague work efficiency and effectiveness”.

The Department for Education’s schools white paper, Every child achieving and thriving, proposes that all schools create digital ISPs for every child with identified SEND. The plans would record barriers to learning, day-to-day provision, reasonable adjustments and intended outcomes, be written with parents and be reviewed at least once a year. As Tes reported on 19 October, the consultation response has not yet been published, but schools would take responsibility for ISPs from 2029. That is a lot of writing, and trusts are sensibly testing tools before the duty arrives.

The same white paper offers a warning worth pinning above the staffroom kettle. It says “simply using general-purpose AI in education often has negative outcomes”, while tools built for an educational purpose show more promise. The three trust tools fit the second description. Pasting a pupil’s history into a free chatbot fits the first.

What schools can do this term

You do not need your own developer. You do need some honesty about what is already happening. The CST guide recommends a low-stakes, no-judgement survey to uncover “shadow AI”, meaning tools staff use without approval, and it cites a survey in which 36 per cent of professionals said they use AI tools their employer has not authorised. If a teaching assistant is already using a chatbot to reword support plans at home, you want to know, and you want them to feel safe telling you.

From there, a few steps make sense:

• Start with one named problem, such as target writing or summarising reports, rather than buying a general assistant.
• Keep identifiable pupil information out of any tool your school has not approved, and complete a data protection impact assessment before you put it into one that has.
• Write down who checks every output and who signs it off. In SEND work that should always be a person who knows the child.
• Judge the tool on accuracy before speed, as Bourne did, and ask SENCOs whether the time saved actually goes back to pupils.

The Department for Education’s generative AI product safety standards already expect filtering to adjust for users with SEND. Its new EdTech Testbed Programme, worth £24 million over four years, will test products, including generative AI tools, in more than 1,000 schools and colleges, with inclusion of learners with SEND among its priorities. The Education Endowment Foundation will run up to 15 randomised controlled trials as part of it. Until that evidence arrives, caution is the professional position.

For background, see our overview of the most common special educational needs and our guide to effective learning support in schools. Our piece on AI and teacher workload asks whether time saved really comes back to staff.

My take

I am more optimistic about this than about most AI in schools, because the problem is real and well defined. A tool that drafts a caseload summary for a human to check is a fair trade. A tool that quietly decides what support a child receives is not, however polished it looks. The trusts in this story seem to understand the difference. The test for the rest of us is whether we build the same habits before the ISP paperwork lands.

Is your school using AI for any part of SEND work yet, or have you decided to hold off? I would love to hear what you are trying and where you have drawn the line, so please share it in the comments.