A Note on the Evidence Base

This case study is structured differently from our other case studies.  The other case studies rest on a mature body of peer-reviewed research with sample sizes in the hundreds of thousands. The evidence base on AI in education is younger, thinner, and in some areas still largely theoretical, a gap acknowledged in both the OECD AI and the Future of Learning report (2024) and the Education Endowment Foundation’s AI review (2024). Where evidence is strong, we say so. Where it is emerging, we say that too. Where we are drawing on practitioner experience and plausible inference rather than established research, it is clearly labelled as Meta Pedagogy commentary. A MAT CEO making strategic decisions deserves to know the difference.

The Strategic Challenge for MATs

Artificial intelligence is not coming to schools. It is already there.  Students are using ChatGPT, Copilot, and similar tools to draft essays, answer exam questions, and navigate coursework. Ofcom’s Online Nation report (2024) confirms that over 70% of UK 13-18 year olds have used generative AI tools. AI recommendation systems heavily influence what content young people see on every major platform they use. AI-generated imagery is circulating in school communities, including images designed to humiliate or harm specific students. AI companionship tools, largely unregulated and poorly understood, are being used by adolescents in ways that have no precedent in the research literature.

At the same time, AI tools offer genuine and documented benefits in the classroom: personalised feedback, differentiated learning support, reduced teacher workload, and improved accessibility for students with SEND.

MAT leaders are being asked to develop coherent policy responses to all of this simultaneously, without a settled evidence base, without clear statutory guidance, and frequently without the technical literacy to evaluate competing claims. The result, in most trusts, is either paralysis or reactive policy making driven by individual incidents rather than strategic thinking.

This case study sets out what the evidence supports, where the genuine uncertainties lie, and what a proportionate, forward-looking response looks like.

The Risks: What the Evidence Shows

Deepfakes and AI-generated content are a documented and growing safeguarding concern.

This is the area where the evidence is clearest and the risk most immediate. AI image generation tools have made it straightforward for anyone with a smartphone to create realistic fake intimate images of real people. The Internet Watch Foundation’s 2024 Annual Report documented a significant rise in AI-generated child sexual abuse imagery and deepfake abuse complaints. The Online Safety Act 2023 introduced new provisions around intimate image abuse, but enforcement against AI-generated content remains inconsistent.

For DSLs and school leaders, the safeguarding implication is direct: existing online safety policies that do not explicitly address AI-generated content are already out of date. KCSIE 2025 references AI and generative technology as areas requiring informed school response, but does not yet provide the operational detail most schools need.

Assessment integrity is under significant and immediate pressure.

The evidence here is consistent across multiple studies and jurisdictions. HEPI’s Student Academic Misconduct Survey: The Impact of Generative AI (2023) found that 53% of UK university students had used AI tools in their assessed work. Secondary school data is less systematically collected, but teacher-reported use in GCSE and A Level coursework is widespread. Turnitin and similar tools have moved rapidly to develop AI detection capability, but detection remains unreliable: false positive rates are high enough, as Turnitin’s own 2024 reliability disclosures and the Jisc briefing note confirm, to make AI detection evidence unsuitable as the sole basis for academic misconduct proceedings. Jisc (2024) has recommended integrative assessment design as the sustainable institutional response.

The deeper problem is structural. Assessment models built around take-home written tasks are increasingly unable to distinguish between student work and AI-generated work. This is a design problem that requires rethinking what and how schools assess.

AI companionship tools raise concerns that are plausible but not yet well evidenced.

Apps such as Replika and Character.AI allow users to form ongoing relationships with AI personas. Their use among adolescents is documented but not systematically studied. The theoretical concerns are coherent: displacement of peer relationships, reinforcement of social withdrawal, and the absence of the friction and reciprocity that characterise healthy human relationships. The Lancet Digital Health published an editorial in mid-2024 raising these concerns about relational AI and adolescents, but peer-reviewed longitudinal evidence is limited. This is an area where school leaders should be monitoring developments rather than responding to established findings.

The Opportunities: What the Evidence Shows

AI tools can meaningfully support differentiated learning and teacher workload reduction.

This is the most evidence-supported opportunity claim. A US-based pilot study of Khanmigo, Khan Academy’s AI tutoring tool, found measurable, if modest, improvements in maths outcomes for students using AI-assisted tutoring compared to control groups. The Education Endowment Foundation’s 2024 review of AI in education found promising early evidence for AI tools that provide immediate, specific feedback on student writing, particularly for lower-attaining students who receive less individualised teacher attention. Teacher workload studies consistently identify marking, planning, and administrative tasks as the highest-burden activities; AI tools are demonstrably useful for all three.

The evidence is not uniformly positive. The same EEF review noted significant variation in outcomes depending on implementation quality, teacher training, and whether AI tools were used to replace or supplement teacher interaction. AI tools used as substitutes for teacher feedback showed weaker outcomes than those used as supplements.

AI can improve accessibility and SEND support.

Text-to-speech, speech-to-text, real-time captioning, and adaptive learning tools powered by AI are producing documented benefits for students with dyslexia, visual impairment, hearing impairment, and autism spectrum conditions. This is an area where the evidence base is more established, drawing partly on assistive technology research that predates the current generation of AI tools. The opportunity is real and the risk profile is relatively low.

AI literacy is becoming a foundational skill.

This is less an evidence claim and more a structural reality. The World Economic Forum’s Future of Jobs Report (2025) identifies AI literacy as among the most in-demand skills across all major employment sectors within five years. The UNESCO AI Competency Framework for Students (2023) provides a practical curriculum scaffold. Students leaving school without a working understanding of how AI systems function, what they can and cannot do, and how to use them critically will be at a significant disadvantage. Schools that treat AI purely as a threat to manage rather than a capability to develop are making a strategic error with long-term consequences for their students.

Where the Evidence Is Still Developing

Three areas warrant monitoring but do not yet support strong policy conclusions.

The cognitive effects of sustained AI tool use on student learning are not well understood. There is theoretical concern that over-reliance on AI for writing and problem-solving may reduce the cognitive effort that drives learning. The evidence base is not yet sufficient to quantify this risk or identify at what threshold it becomes significant.

The mental health implications of AI companionship tools for adolescents are largely unresearched. The mechanisms that make social media harmful to a subset of young people, compulsive engagement, identity validation through feedback, displacement of in-person relationships, are present in AI companionship tools in more concentrated form. This warrants serious attention as the evidence develops.

The equity implications of AI in education are beginning to emerge. Early evidence suggests AI tools may widen attainment gaps if access is unequal, but may narrow them if implemented well and targeted at lower-attaining students. The direction of effect depends almost entirely on implementation decisions that are within school control.

What a Balanced School Strategy Looks Like

The schools making the most coherent responses to AI are those that have resisted the pressure to treat it as either a universal solution or an existential threat. The evidence supports a strategy built around four principles.

Distinguish between AI as a safeguarding concern and AI as a learning tool. These require different governance structures, different leads, and different policy frameworks. Conflating them produces policies that are simultaneously too restrictive for learning and not restrictive enough for safeguarding.

Build assessment integrity around design, not detection. AI detection tools are not reliable enough to form the basis of misconduct proceedings. The sustainable response is to redesign assessment towards tasks that are difficult to complete with AI alone: oral components, process portfolios, in-class tasks, and assessments that require students to demonstrate and explain their reasoning rather than produce a written output.

Treat AI literacy as curriculum, not just policy. Students who understand how large language models work, what their limitations are, and how recommendation algorithms function are better equipped to use AI tools responsibly and to recognise when they are being manipulated by them. This is a curriculum design question as much as a safeguarding one.

Update safeguarding policy to explicitly address AI-generated content. Existing online safety policies in most schools were not written with AI image generation in mind. The gap between current policy language and current risk is significant and closing it is a straightforward, immediate action.

Recommended Actions

1. Audit current online safety and safeguarding policies for AI gaps Review existing policies against the specific risks identified in this case study: AI-generated imagery, deepfake abuse, and AI companionship tools. Update language to explicitly address these. Circulate to all staff with a briefing on implications. Owner: DSL/CEO | Term 2, 2025/26 | Staff time | Anchor: KCSIE 2025; Online Safety Act 2023
2. Redesign at least one high-stakes assessment per subject per year to be AI-resistant Work with heads of department to identify which assessment tasks are most vulnerable to AI completion and redesign them around oral, process, or in-class components. This does not require abandoning written assessment; it requires building in verification. Owner: Deputy Head Curriculum | Academic year 2026/27 | Staff time | Anchor: DfE Generative AI in Education Guidance (2024)
3. Introduce AI literacy as a named strand within the computing and PSHE curriculum Map existing curriculum coverage of how algorithms work, how AI systems are trained, and how to evaluate AI-generated content. Identify and fill gaps. This does not require a new subject; it requires deliberate sequencing of what most schools are already partially teaching. Owner: Head of Computing/PSHE Lead | From September 2026 | Staff time | Anchor: DfE AI in Education Strategy; UNESCO AI Competency Framework for Students (2023)
4. Establish a trust-wide AI working group with a structured review cycle The evidence base is moving faster than annual policy cycles. A working group that meets termly, monitors developments, and makes recommendations to the board ensures strategic decisions are informed rather than reactive. Include a governor, a DSL, a classroom teacher, and if possible a student representative. Owner: CEO | From Easter 2026 | Minimal cost | Anchor: DfE Generative AI in Education Guidance (2024)
5. Trial one AI tool for teacher workload reduction with structured evaluation Select one low-risk, high-impact use case: marking feedback, lesson planning, or SEND support. Pilot with a volunteer group of teachers, evaluate against workload and outcome measures, and use findings to inform trust-wide decisions. Avoid procurement of expensive platforms before evidence of benefit is established in your specific context. Owner: Deputy Head/Research Lead | Summer term 2026 | Variable depending on tool | Anchor: EEF AI in Education Review (2024)

Meta Pedagogy Commentary

The following reflects practitioner judgement based on 20-plus years working across schools and MATs, rather than the published evidence base above. It is clearly labelled as such.

The single most consistent pattern I observe when working with school leaders on AI is the gap between what governors think is happening and what is actually happening in classrooms. Students are using AI tools extensively, often creatively and sensibly, and often in ways their teachers are not fully aware of. Teachers are using AI tools for planning, differentiation, and feedback, often without formal permission or support. The policy vacuum is not preventing AI use; it is preventing schools from shaping it.

The leaders who are getting this right are not those who have produced the most comprehensive AI policy document. They are those who have created the conditions for an honest conversation about what AI use is actually occurring, what is working, what is causing problems, and what the school’s values require. That conversation cannot happen if the implicit message from leadership is that AI use is presumptively suspicious.

Assessment integrity is the area where I see the most disproportionate anxiety. The fear of AI-generated coursework is real and legitimate, but the response in many schools, blanket prohibition combined with unreliable detection tools, is producing a situation where students who use AI skillfully are not caught and students who use it clumsily are, which is neither fair nor educationally coherent. The schools doing this well have accepted that the assessment landscape has changed permanently and are redesigning accordingly, rather than trying to hold a line that cannot hold.

The deepfake risk is, by contrast, systematically underestimated. I have spoken with DSLs in multiple schools who were unaware that the tools to create realistic fake intimate images of students are freely available and widely known among secondary age pupils. This is not a future risk. It is a present one, and it is arriving in schools that have not yet updated their policies to address it.

The opportunity in all of this is genuine. Schools that develop genuine AI literacy in their students, that model thoughtful and critical engagement with AI tools, and that build the institutional capacity to adapt as the technology develops, will be significantly better positioned than those that do not. The evidence base will mature. The technology will develop. The schools that have built the strategic infrastructure to respond will be the ones that serve their students well over the next decade.

Evidence Validation Appendix

For governance and audit purposes. Each key claim mapped to its supporting source and confidence rating.

Claim	Source	Confidence
Over 70% of UK 13-18 year olds have used generative AI tools	Ofcom Online Nation (2024)	High
AI recommendation systems influence content young people see	Established algorithmic design across TikTok, YouTube, Instagram	High
Rise in AI-generated child sexual abuse imagery	IWF Annual Report (2024)	High
53% of UK university students used AI in assessed work	HEPI Student Academic Misconduct Survey (2023)	High
AI detection tools have high false positive rates	Turnitin reliability disclosures (2024); Jisc briefing	High
Integrative assessment design recommended as response	Jisc (2024)	High
Lancet editorial on relational AI and adolescent mental health	Lancet Digital Health (mid-2024)	High, limited longitudinal follow-up
Modest maths gains from AI tutoring	Khanmigo US pilot, Stanford GSE (2023)	Moderate, US-based, small scale
AI feedback tools benefit lower-attaining students	EEF AI in Education Review (2024)	Moderate, implementation-dependent
AI literacy among top in-demand skills by 2030	WEF Future of Jobs Report (2025)	High
UNESCO curriculum scaffold for AI literacy	UNESCO AI Competency Framework for Students (2023)	High
Cognitive effects of AI tool use under-researched	OECD AI and the Future of Learning (2024)	Emerging
AI companionship mental health risk plausible but unproven	Lancet Digital Health (2024); theoretical literature	Emerging
Equity implications of AI implementation context-dependent	Early evidence, multiple sources	Emerging
AI policy lag in schools documented	NAHT Headteacher Survey (2024); ASCL briefings	Moderate

References

Department for Education (2024). Generative AI in Education: Guidance for Schools and Colleges. https://www.gov.uk/government/publications/generative-artificial-intelligence-in-education

Education Endowment Foundation (2024). Artificial Intelligence in Education: Evidence Review. https://educationendowmentfoundation.org.uk

HEPI (2023). Student Academic Misconduct Survey: The Impact of Generative AI. https://www.hepi.ac.uk

Internet Watch Foundation (2024). Annual Report 2024. https://www.iwf.org.uk/about-us/who-we-are/annual-report

Jisc (2024). Assessment and Feedback in the Age of Artificial Intelligence. https://www.jisc.ac.uk

Keeping Children Safe in Education (2025). Statutory Guidance for Schools and Colleges. https://www.gov.uk/government/publications/keeping-children-safe-in-education–2

OECD (2024). AI and the Future of Learning. https://www.oecd.org/en/topics/sub-issues/ai-in-education.html

Ofcom (2024). Online Nation Report. https://www.ofcom.org.uk/research-and-data/internet-and-on-demand-research/online-nation

Online Safety Act 2023. https://www.legislation.gov.uk/ukpga/2023/50

UNESCO (2023). AI Competency Framework for Students. https://www.unesco.org/en/digital-education/ai-future-learning

World Economic Forum (2025). Future of Jobs Report 2025. https://www.weforum.org/publications/the-future-of-jobs-report-2025