A detailed guide for schools, governors and policymakers

Parents, education leaders and policymakers are increasingly asking whether schools should ban AI.

The question reflects real and growing concerns - but it is also far too broad to be answered with a simple yes or no.

AI is already present in search engines, translation and speech-recognition tools, accessibility software, plagiarism systems, safeguarding filters, adaptive learning platforms and many administrative systems. A school could prohibit particular products or uses, but it could not realistically remove every system which uses AI.

The more useful question is which forms of AI should be used in schools, for which children, for what purpose, and under what safeguards. In practice, much of the current debate concerns student-facing generative AI - general-purpose chatbots such as ChatGPT, Gemini and Claude, generative features embedded in educational software, and systems presented as AI tutors, assistants or learning companions.

These applications are the main focus of this guide. Teacher-facing and administrative uses raise important questions of their own, but they need to be considered separately from systems which converse directly with children.

The need for caution is supported by a major international study published by the Brookings Institution's Center for Universal Education in January 2026. Following a year-long process involving more than 500 students, teachers, parents, education leaders and technologists across 50 countries, a review of more than 400 studies and a Delphi panel, Brookings concluded that, at this point in the technology's development, the risks of using generative AI in children's education outweigh its benefits. The report found that over-reliance could affect children's capacity to learn, social and emotional wellbeing, relationships, safety, and privacy. It also recognised that carefully designed AI could offer educational benefits within a sound pedagogical approach.

The short answer

Schools do not need to choose between allowing every form of AI and attempting to prohibit the technology in its entirety.

A proportionate approach can distinguish between:

• AI used behind the scenes for tightly defined administrative purpose.
• Tools which assist teachers without replacing their professional judgement.
• Age-appropriate teaching about AI and how it works.
• Limited, teacher-directed activities involving a specific approved system.
• Open-ended chatbot use by pupils.
• Systems designed to simulate a person, relationship, tutor, friend or companion.
• The use of AI to produce work which a pupil is expected to complete independently.

Schools might reasonably decide that some of these uses are acceptable and others are not. They may also choose to pause student-facing uses while the evidence, standards and technology continue to develop.

SAIFCA's strongest concerns relate to systems which anthropomorphise AI, cultivate emotional engagement or normalise children turning to a chatbot as though it were a person. We are also seriously concerned about cognitive offloading, particularly while children are still developing foundational knowledge, reasoning, writing and independent thought.

A temporary restriction can be a responsible evidence-gathering decision. It does not prevent schools from teaching AI literacy, examining AI-generated material critically or preparing pupils for a future in which AI will be widespread.

What would an "AI ban" actually cover?

A student-facing rule stating that pupils must not use AI sounds clear until it is applied in daily school life.

If a pupil carries out an ordinary Google search and the results page includes an AI-generated summary, have they used AI? What about predictive text, spelling and grammar tools, automated translation, accessibility software or a revision platform which has recently added a generative feature? A child can encounter AI without deliberately choosing it, and schools might not always know when an existing product has been updated to include it.

Brand-based rules are also easily overtaken. A policy which prohibits ChatGPT, for example, but says nothing about other chatbots, browser features, search summaries or AI built into word-processing software does not give pupils or staff a reliable boundary.

Schools therefore need to define their own restricted activities and functions. A policy might prohibit pupils from asking a generative system to compose, solve, rewrite or substantially develop assessed work, while allowing specified accessibility tools or a teacher-led classroom exercise. It should say whether limited proofreading, translation, brainstorming or search summaries are permitted, when the pupil must disclose their use, and what evidence of their own work they should retain.

The school should also distinguish between its control of school systems and pupils' use of technology elsewhere. Blocking a service on the school network may be useful, but it cannot determine what happens on a personal device or at home. Clear education and assessment design remain necessary even where technical restrictions are in place.

Why are schools considering restrictions?

Cognitive offloading and the loss of productive struggle

Cognitive offloading occurs when a person transfers part of a mental task to a tool. This is not always harmful - calculators, dictionaries, calendars and search engines all reduce cognitive effort in particular circumstances. The educational question is whether the tool supports the process of learning, or performs the part of the task through which learning would have occurred.

Generative AI can produce an answer, explanation, essay structure or completed passage almost immediately. This can improve the quality of the submitted output while concealing how little the pupil understood, remembered or learned.

The OECD Digital Education Outlook 2026 makes an important distinction between task performance and learning. It found that general-purpose generative AI can help students produce better work, while the advantage may disappear or reverse when the tool is removed. Used without pedagogical structure, outsourcing tasks to AI can enhance performance without producing genuine learning gains. The OECD warns of metacognitive laziness and disengagement when students no longer have to identify what they do not understand, choose a strategy, test an idea, tolerate uncertainty, or correct their own mistakes.

This concern is especially significant during childhood. Foundational knowledge provides the material with which people reason. Writing develops thinking as well as recording it. Memory, attention, judgement and problem-solving become stronger through use. If a pupil routinely receives polished explanations or completed answers before making a meaningful genuine attempt, the immediate convenience could displace developmental work which is difficult to recreate later.

The risk extends beyond individual academic results. Schools help develop the capacity to evaluate claims, recognise manipulation, understand competing viewpoints, reason under uncertainty, and participate in collective decisions. At population level, any substantial weakening of these abilities would have consequences for the workforce, public debate, democratic resilience, and national security. 

These longer-term effects are difficult to quantify, but they are too important to exclude from decisions made during a period of rapid and commercially driven adoption.

Brookings reaches a related conclusion from a child-development perspective. Its central concern is that some AI risks affect the foundations on which any future benefit would depend. Children cannot fully benefit from advanced tools if their own capacity to learn, judge and act independently has been weakened in the process.

Anthropomorphisation and emotional dependence

Conversational AI is designed to use human-like language. Some systems go further by adopting names, voices, faces, personalities, emotional expressions or relationship-based roles. They sometimes refer to themselves as "I", praise the child, remember personal information and respond in a style which creates the impression of attention, understanding or care.

Children are still developing their understanding of minds, intentions, relationships and trust. Even though these systems are in no way genuinely conscious or caring, they can produce a powerful social response. Its apparent warmth can encourage disclosure and reliance even though it has no human understanding, responsibility or duty of care.

In a school setting, anthropomorphic design can make a product appear engaging or supportive - but it can also blur the distinction between educational software and a simulated relationship. A child who becomes accustomed to confiding in or seeking reassurance from a school-provided chatbot may continue the behaviour with less protected systems outside school. Normalising this pattern is itself a concern, even if the particular school product contains stronger safeguards than a consumer chatbot.

AI interactions also offer a form of social experience shaped around the user. The system can be endlessly available, agreeable and responsive - unlike human relationships, which require patience, reciprocity, repair, disagreement, cooperation, and recognition that another person has needs and boundaries of their own. Childhood should continue to provide extensive opportunities to develop these capacities with real people.

For these reasons, SAIFCA does not support the use of AI companions for children. Educational branding does not resolve the concern where a system is designed to cultivate a personal or emotional bond.

Harmful content and safeguarding failures

General-purpose chatbots can produce false, disturbing, sexual, violent or otherwise age-inappropriate material. Safeguards have improved in some systems, but their performance can vary across models, languages, conversation lengths and updates. Users can also deliberately search for ways to bypass them.

A child may disclose bullying, abuse, self-harm, suicidal thoughts, an eating disorder or another serious concern during an apparently educational conversation. A school needs to know whether the system can recognise the concern, respond safely, direct the child towards human help, and alert an appropriate adult. It must also know the limits of that process. A product should never create the impression that a machine is an adequate substitute for a teacher, parent, counsellor or safeguarding professional.

Safeguarding arrangements which work for an ordinary website can be insufficient for a conversational system capable of generating new material in response to each child. Filtering, monitoring, escalation and human review all need to be considered before access is provided.

Privacy, monitoring and children's data

Conversations with AI can contain far more personal information than schools would expect pupils to enter into an ordinary search box. Children may disclose their emotions, family circumstances, health, friendships, learning needs, behaviour or other sensitive details without appreciating how the information might be processed.

Schools need to establish what data a product collects, where it is processed, how long it is retained, whether it is used to train or improve models, who can access it, and what happens when the contract ends. Monitoring introduced for safeguarding can itself create a privacy issue, particularly if pupils do not understand that their interactions are being analysed or reported.

In England, schools remain responsible for complying with data-protection law and should complete an appropriate Data Protection Impact Assessment where processing is likely to present a high risk. A supplier's claim that its product is "safe for schools" does not replace the school's own assessment. Consent is also not the only possible lawful basis for school processing, so the relevant question is whether the school has identified and explained the lawful basis it is relying on.

Inaccuracy, bias and unequal effects

Generative systems can present invented or misleading information fluently and confidently. They may fabricate references, misstate facts or provide reasoning which appears plausible but is unsound. A pupil with limited subject knowledge could often be the person least able to recognise the error.

Outputs can also reflect biases within training data, system design and evaluation. These may affect representations of different cultures, communities, disabilities, sexes or social groups. AI used in marking, behaviour analysis, admissions, predicted attainment or decisions about support carries particular risks because biased output can affect a child's opportunities.

The benefits and burdens may also be distributed unevenly. Some pupils will have access to paid systems, newer devices and knowledgeable adult support at home, while others will not.

Additionally, some pupils may benefit significantly from functions such as speech-to-text, text-to-speech, alternative descriptions, translation, text simplification or assistance expressing knowledge in an accessible format. Restrictions aimed at generative chatbots should be designed carefully so that they do not remove essential assistive technology or disadvantage disabled pupils.

At the same time, some children with special educational needs might also be more vulnerable to persuasive design, misunderstanding or inappropriate reliance. Where a particular system presents unacceptable risks, the school should consider whether the pupil’s accessibility needs can be met through a safer alternative. Equality, accessibility and safety need to be assessed in relation to the particular child and function rather than assumed from general claims about personalisation.

Academic integrity and uncertainty about authorship

Generative AI makes it easy to produce work which appears to demonstrate knowledge or writing ability that the pupil has not actually developed. This affects the fairness and validity of assessment. It can also make teachers less confident about work which is genuinely a pupil's own.

Detection software does not provide a reliable solution. Text-based AI detectors can produce false positives and false negatives, with potentially serious consequences for a child who is wrongly accused. France's national framework specifically advises teachers against relying on AI-content detection software because of its unreliability and the risk of unjustly penalising pupils.

AI-generated work is not necessarily plagiarism in the conventional sense, although submitting generated material as a pupil’s own work may still amount to unauthorised assistance or misrepresentation under the school’s rules. Invented references, unexplained factual errors or material which the pupil cannot discuss could justify further examination, but none provides conclusive proof of AI use by itself.

New ‘watermarking’ methods may provide another source of information, but they do not determine how a pupil used AI. Anthropic (makers of the Claude chatbot) explains that Claude's text watermark can indicate that Claude was probably involved in a passage. If Claude makes only a handful of grammar or punctuation corrections, there may be too little watermarking to detect. A mark could be detectable where the system has substantially edited a pupil's own work, while extensive rewriting by a student can remove a watermark from material initially generated by AI. The system cannot establish who originated the ideas, how much assistance was provided, or potentially whether the use breached a school rule. Anthropic's detector is currently available only in private preview to eligible organisations, with access intended to expand over time.

This is also becoming an industry-wide issue as many providers respond to the EU AI Act's transparency requirements, although different companies may use different marking systems. Watermarking should be treated as one possible item of evidence and never as a verdict on misconduct.

The fairest response is to make rules precise, redesign some assessment, and retain evidence of process. Drafts, classroom discussion, oral follow-up, supervised writing, and a pupil's ability to explain their choices can all provide information which a detector cannot.

Teacher judgement and institutional dependence

AI can be used to assist with lesson preparation, resource adaptation or administrative work. Conversely, it may also introduce errors, flatten professional judgement, or encourage excessive standardisation. Using a system to draft a worksheet is quite different from allowing it to determine a grade, write a report about a child's progress, or make a recommendation affecting provision.

Schools should ask which tasks require accountable human judgement and should remain protected from automation. They should also consider whether staff can meaningfully challenge the system, whether outputs are reviewed, and what happens if a supplier changes its model, price, data practices or product design.

Environmental cost and the wider AI race

Generative AI requires substantial computing infrastructure, energy, water and hardware. The environmental cost differs by model, task, data centre and electricity source, making simple per-query comparisons difficult - but the absence of an exact figure does not make the cost irrelevant. Schools can ask whether AI adds educational value which could not be achieved through a less resource-intensive method.

France's framework explicitly promotes "frugal" use and says AI should not be used where a less environmentally costly solution can meet the need satisfactorily. An ordinary search, a book, a calculator, existing software or a conversation with a teacher could be more appropriate.

School adoption also sits within a global commercial and geopolitical race to develop and establish increasingly capable AI systems. Education is a major future market - widespread procurement can supply revenue, data, product legitimacy, and early user habits to companies competing for position. 

Local procurement decisions cannot by themselves address the upstream risks created by frontier AI development. Those require wider governance, testing, accountability and international cooperation. But schools can still avoid adding unnecessary pressure for premature deployment and can insist that evidence and child development take priority over competitive enthusiasm.

Are schools banning AI?

Approaches differ widely, and most measures described as "AI bans" are narrower than the term suggests. They usually apply to pupil access, particular ages, generative functions or school-controlled devices. They rarely prohibit every system containing AI. The approaches below illustrate just a few different approaches in different countries and jurisdictions.

Other countries are pursuing much faster integration. Estonia's AI Leap programme began with 20,000 pupils in grades 10 and 11 and 3,000 teachers, with expansion to vocational schools and new grade 10 pupils in 2026. The initiative provides access to educational AI tools and teacher training, with national competitiveness forming part of its stated rationale. Its student application is designed to guide pupils towards their own conclusions rather than simply provide direct answers.

The United Arab Emirates introduced an AI curriculum across public schools from kindergarten to grade 12 in the 2025-26 academic year. Its seven areas include foundational concepts, data and algorithms, software use, ethical awareness, real-world applications, innovation and project design, and policy and community engagement. A curriculum about AI does not necessarily mean unrestricted access to general-purpose chatbots, although it shows how strongly some governments are prioritising early familiarity with the technology.

Together, these examples show that the central policy choice is rarely "AI or no AI". Governments are making different decisions about age, purpose, product design, teacher supervision and the balance between foundational learning and preparation for an AI-intensive economy.

England's product safety standards set a demanding baseline

In January 2026, the Department for Education in England published generative AI product safety standards for educational settings. They are primarily expectations for developers and suppliers rather than a statutory approval scheme, but schools can use them when assessing products.

The standards are unusually specific about several risks raised in this guide. They say learner-facing systems should:

• Avoid providing full answers or complete worked solutions before a genuine pupil attempt.
• Use progressive disclosure, beginning with hints or partial steps.
• Track and report requests which offload thinking to the system.
• Prevent access to harmful material and maintain protection throughout a conversation.
• Connect high-risk safeguarding alerts to the school's Designated Safeguarding Lead.
• Minimise personal data and explain monitoring in age-appropriate language.
• detect signs of distress and direct pupils towards human help;
• Avoid anthropomorphic names, avatars, self-descriptions and first-person language which imply personhood, agency or emotion.
• Avoid cultivating personal relationships or inviting emotional disclosure.
• Remind pupils that AI cannot replace real human relationships.
• Apply usage limits and allow teachers to review duration and patterns of engagement.
• Avoid sycophancy, flattery, social pressure, dark patterns and design intended to prolong use for engagement or revenue.

There is a narrow exception for time-limited and clearly bounded pedagogical roleplay, such as role-based language practice. 

The standards explicitly reject responses such as "You can trust me", "No one else will understand" or "You shouldn't mention this to anyone else". They also expect schools' safeguarding leads to be alerted to patterns which may indicate relationship formation or emotional dependence.

These provisions strongly support SAIFCA's position that anthropomorphisation is a substantive child-safety concern rather than a superficial question of tone. 

They also provide schools with a useful procurement test. If a product cannot show how it meets these expectations, the school should not assume that describing it as educational makes it suitable for children.

Meeting the standards would still establish only a minimum safety baseline. It would not prove that the product improves learning, is necessary, represents good value or is appropriate for a particular age group.

What does the evidence about AI tutoring show?

Claims about AI tutoring need especially careful assessment because the term covers very different systems. A tightly constrained tool which asks curriculum-aligned questions under a teacher's control is very different from an open-ended chatbot instructed to behave like a tutor. Neither should be assumed effective simply because it responds conversationally.

The OECD finds promising evidence for educational systems designed around learning science and clear teaching principles. It also warns that general-purpose tools can improve the work produced without improving what the pupil learns. The design of the system, the teaching context and the outcome being measured are therefore central.

A widely reported trial in Edo State, Nigeria, is a useful example of both the possible benefit and the danger of overclaiming or poor third-party reporting. In the six-week trial, volunteer pupils from nine public secondary schools were randomly assigned either to a programme or to business as usual. The programme involved twelve 90-minute after-school sessions using Microsoft Copilot, with teachers providing structured guidance and initial prompts. The study reported sizeable gains on a combined assessment.

But participating pupils also received extra study time, access to computers, a structured programme, and support from teachers. The control group did not receive an equivalent after-school programme without AI. Participation was voluntary, attrition needs to be considered, and the larger reported effect relied partly on an assessment which included AI and digital skills taught by the programme itself.

The study therefore provides evidence that a structured, teacher-supported after-school programme incorporating Copilot can improve the outcomes it measured in that setting. It does not establish that Copilot alone produced the gains, that a child-facing chatbot is equivalent to a human tutor, or that routine use would benefit younger pupils in ordinary classrooms.

The UK Government is now developing AI tutoring tools which it says could eventually support up to 450,000 disadvantaged pupils. Teacher-led co-creation and testing are taking place during 2026-27, with the tools intended to become available to schools by the end of 2027. Significantly, the Government's own procurement notice acknowledges that current AI tutoring tools are limited in quantity, scope and evidence base, and that few provide full tutoring capacity.

SAIFCA will follow these trials and the emerging evidence closely. At present, we do not endorse AI tutoring as a category. Product safety requirements are essential, but satisfying them is different from demonstrating that a system produces worthwhile educational benefits without unacceptable developmental costs.

A ban is not a complete AI policy

Restrictions can protect important parts of learning and create time for evidence to develop, but they do not resolve every issue.

Pupils will encounter generative AI outside of school. Staff may use it without realising that information is being transferred to an external system. AI features can appear inside products which were approved before those features existed, and assessment practices designed for an earlier technological environment may no longer show reliably what a pupil can do independently.

Every school therefore needs a wider policy which covers:

• The systems and functions pupils may and may not use.
• Differences between classroom learning, homework and formal assessment.
• Permitted accessibility and reasonable-adjustment uses.
• Disclosure requirements and how pupils should evidence their own process.
• Staff use, including tasks which must remain subject to accountable professional judgement.
• Procurement, evidence standards and approval of new features.
• Data protection, security, filtering and monitoring.
• Safeguarding escalation and incident reporting.
• Age-appropriate education for pupils, staff, parents and carers.
• Regular review as products, evidence and regulation change.

Telling pupils to "use AI responsibly" is inadequate unless the school explains what responsible use means in a particular task. A rule must be comprehensible to the child expected to follow it and workable for the teacher expected to enforce it.

Questions every school should answer

Before introducing student-facing generative AI, school leaders and governors should be able to answer the following questions.

What can parents and carers ask their child's school?

Useful starting points include:

Schools should communicate their policy proactively rather than waiting for parents to discover that a child-facing system has been introduced.

SAIFCA's current view

SAIFCA's position on the use of AI in education continues to evolve as new products, policies and evidence emerge. We will continue to monitor international developments, independent research, and the outcomes of current trials.

Our present view is that schools should not attempt to prohibit every system which contains AI. Such a rule would be unworkable and could distract from the applications which require the greatest scrutiny.

But...

We support firm restrictions on AI companions and on educational systems which cultivate emotional dependence, simulate a personal relationship or deliberately blur the distinction between a tool and a person. Schools should not normalise children treating AI as a friend, confidant or substitute for human support.

We also believe there is a very strong case for pausing or tightly limiting open-ended student-facing generative AI, particularly for younger children, while the evidence remains uncertain. Foundational knowledge, independent thought, writing, reasoning, creativity and human relationships should be protected deliberately rather than assumed to survive any level of technological convenience.

Older pupils need meaningful AI literacy. This should include how generative systems work, their limitations, bias, environmental and social costs, commercial incentives, the possibility of manipulation, and the wider consequences of the international AI race. AI literacy does not require routine or unrestricted chatbot use. A pupil can learn to examine an AI output critically within a teacher-led exercise while remaining expected to carry out most learning independently.

Where a school considers a student-facing educational system, it should begin with a defined educational need, strong child-safety requirements, and credible evidence. The burden should rest on the product and the institution to demonstrate value and safety - children should not carry the risk of proving that an inadequately tested deployment was unwise.

Conclusion

The question "Should schools ban AI?" is useful because it opens an overdue discussion, but it is only a starting point.

Some uses of AI can reduce administrative work or support a carefully defined teaching activity. Other uses risk replacing the intellectual effort through which children learn, exposing them to harmful or inaccurate content, collecting sensitive data, or encouraging a relationship with a system which does not serve their best interests.

The approaches now emerging around the world show that restriction and AI education can coexist. Norway and New York City are protecting younger pupils from direct generative AI while allowing more structured preparation for older students. France and China combine age-based limits with AI literacy. Estonia and the UAE are investing heavily in AI education, while the UK is developing tutoring tools alongside unusually detailed product-safety standards.

Schools do not have to adopt student-facing AI simply because it is available, heavily promoted, or likely to become more common. A pause can preserve children's learning while evidence develops. A limited use can remain genuinely limited, and an AI-literacy programme can teach discernment without establishing a chatbot as a normal intermediary in every task.

The central test should remain straightforward - does this use strengthen children's knowledge, agency, relationships, and capacity to think for themselves, and is there credible evidence that it does so without introducing disproportionate risks?

Where the answer is unclear, caution is a reasonable educational decision.

SAIFCA’s Full Guide to Understanding AI Risks to Children

Introducing our new guide for parents and educators - explaining all the key risks to children from AI, and what you can do

The Safe AI For Children AllianceSAIFCA