Three years is long enough for the novelty phase to pass and for patterns to emerge. In most schools, the question is no longer “Will staff use generative AI?” but “Which uses are now stable enough to standardise, and which still threaten reliability, safety or trust?” This matters because the wrong kind of certainty creates brittle policy. Overconfidence leads to lax assessment controls; over-caution leads to unmanageable bans and underground use.
If you want a practical way to take stock, the simplest route is an evidence pack: a small set of measures and artefacts you review termly, rather than reacting to headlines. If you are planning a structured review cycle, you may find the approach in [End-of-year AI audit evidence pack](/en-gb/blog/2025/05/29/end-of-year-ai-audit-evidence-pack-keep-stop-scale-summer-action-plan/ aligns well with what follows.
What we can now judge is the “shape” of impact: where AI reliably saves time, where it shifts the nature of pupil work, and where it increases risk. What we still cannot judge with high confidence is long-run attainment impact at scale, because implementation varies, comparison groups are messy, and assessment itself is changing. That uncertainty is not a reason to pause; it is a reason to measure the right things.
From late 2022 into 2023, access was the story. Staff and pupils experimented, often privately, and leaders scrambled for guidance. Early policy tended to swing between blanket bans and permissive “use it responsibly” statements that were hard to enforce.
By 2024, expectations shifted. Many teachers had tried AI for planning or rewriting resources, and pupils had discovered that “good enough” writing could be generated quickly. Schools began building routines: prompt guidance, citation expectations, and basic integrity statements. The more mature conversations moved from “Is it cheating?” to “What counts as learning evidence?”
In 2025, the story became controls and governance. Platforms added more admin features, and schools started evaluating tools rather than adopting them ad hoc. The best implementations looked less like a tech rollout and more like a curriculum-and-assessment adjustment, with clear boundaries and a review loop. If you are refreshing policy for the coming year, the [Annual AI acceptable use policy refresh checklist](/en-gb/blog/2025/08/18/annual-ai-acceptable-use-policy-refresh-checklist-2025-26/ is a useful companion.
The most visible change has been the move from bans to bounded use. Schools that tried bans often found they were unenforceable, inequitable, and pushed use into unobservable spaces. Bounded use is more workable: it defines permitted tasks (for example, planning, differentiation drafts, feedback preparation), prohibited tasks (for example, generating final assessed responses), and conditional tasks (for example, idea generation with process evidence).
The second shift is role-based controls. Policies now increasingly separate staff use, pupil use, and admin use, with different expectations for data handling and transparency. A teaching assistant using AI to simplify a text for accessibility needs a different rule set from a pupil using AI to draft an essay.
Governance becomes “evidenceable” when it produces artefacts: a tool register, risk assessments, staff training records, and a termly review note that says what changed and why. For schools navigating procurement and compliance pressures, the governance framing in [EU AI Act one year on: procurement and governance](/en-gb/blog/2025/09/17/eu-ai-act-one-year-on-uk-schools-procurement-governance-playbook/ can help you structure decisions even outside the EU context.
Settled practice has emerged where AI supports teacher work without replacing the learning task. The routines that have stuck are usually “prep-side”, not “performance-side”. A common example is lesson planning: a teacher asks for three alternative explanations of a concept, then chooses one and rewrites it in their own voice. Another is adaptation: generating a simpler version of a reading passage, plus vocabulary support, then checking it for accuracy and tone. A third is feedback preparation: drafting comment banks aligned to success criteria, then editing to match the pupil’s work.
These routines persist because they reduce blank-page time while keeping professional judgement central. They also create manageable safeguarding and integrity risk, provided staff avoid entering sensitive personal data and treat outputs as drafts.
Where routines remain volatile is when AI becomes the “doer” of the pupil task: full answers, full essays, or anything that removes productive struggle. If you are exploring voice and multimodal tools, the classroom lens in [Voice AI in schools: accessibility and safeguarding](/en-gb/blog/2025/10/13/voice-ai-in-schools-accessibility-fluency-formative-assessment-safeguarding-rubric/ is a helpful way to separate genuine inclusion benefits from new supervision needs.
Assessment integrity has matured from “catching cheating” to “designing out shortcuts”. Schools are increasingly drawing clear integrity boundaries: which tasks must be completed without AI, which can use AI with disclosure, and which are explicitly about using AI well.
The most robust approach is to treat process evidence as part of the assessment design. For example, a pupil might submit a planning page, a short in-class paragraph written without devices, and a reflection on how they revised their draft. The aim is not surveillance; it is to ensure the evidence matches the construct you are assessing.
Redesign patterns that reduce the value of “AI shortcuts” are now common. Teachers are using more in-class writing windows, more oral explanation, more personalised prompts tied to class texts or local contexts, and more tasks that require referencing lesson-specific material. Another pattern is “show your workings” for writing: outline decisions, source notes, and a brief commentary on changes. Used well, these approaches also improve learning, because they make thinking visible.
On workload, confidence is medium-high that AI can save time for some teacher tasks, especially drafting and reformatting materials. The savings are uneven: they depend on subject, experience, and the quality of the checking routine. A teacher who treats AI output as finished often pays later in corrections.
On attainment, confidence is low-to-medium. Small studies and local evaluations suggest potential benefits for practice and feedback, but results vary widely. The biggest variable is implementation discipline: when AI supports retrieval practice, modelling, and targeted feedback, it can help; when it replaces effort, it can hinder.
On equity, confidence is medium that unmanaged AI access can widen gaps. Pupils with more support at home can use AI more strategically, and those with weaker literacy may become over-reliant. However, confidence is also medium that well-scaffolded, supervised use can improve access, particularly for language support and additional needs, if schools explicitly teach how to use help without outsourcing the thinking.
On unintended consequences, confidence is high that unclear expectations create friction: staff disagreement, inconsistent sanctions, and pupil confusion. That is why an evidence-pack approach is valuable: it makes the conversation concrete and reduces policy-by-anecdote.
Bias and hallucinations remain live risks, particularly when staff use AI to summarise complex topics or generate “facts” quickly. The settled mitigation is simple but non-negotiable: treat outputs as drafts, verify against trusted sources, and avoid using AI as the authority in the room.
Over-reliance is harder. Pupils can become dependent on instant phrasing help, which masks gaps in vocabulary and sentence control. The mitigation is routine: short device-free writing, explicit modelling, and teaching pupils how to use AI for planning and feedback rather than final answers.
Data protection drift is a quiet risk. It happens when staff gradually paste more sensitive information into tools because “it worked last time”. Your policy should name what must never be entered, and your training should include realistic scenarios. Admin controls and evaluation checklists can help here; see [Google Classroom/Workspace AI admin controls](/en-gb/blog/2025/10/01/google-classroom-workspace-ai-update-october-2025-uk-school-admin-controls-checklist/ for a model of the questions to ask, even if you use different platforms.
Finally, the new “agentic” risk surface is emerging: tools that take actions, not just generate text. If a system can send messages, modify files, or act across multiple steps, your supervision and logging expectations must change. For leaders tracking the next wave, [GPT-5 readiness pack for schools](/en-gb/blog/2025/10/30/gpt-5-watch-week-1-readiness-pack-for-schools/ is a useful way to think about capability shifts without panic-buying tools.
Ready to Revolutionise Your Teaching Experience?
Discover the power of Automated Education by joining out community of educators who are reclaiming their time whilst enriching their classrooms. With our intuitive platform, you can automate administrative tasks, personalise student learning, and engage with your class like never before.
Don’t let administrative tasks overshadow your passion for teaching. Sign up today and transform your educational environment with Automated Education.
A workable impact assessment is small enough to run termly. Start by separating measures (numbers you track) from artefacts (evidence you can show).
Measures worth keeping include staff time saved on specific tasks (captured as short pulse checks), the proportion of assessed tasks with a clear AI condition statement, and the number of integrity incidents where AI use was suspected and investigated. Also track training coverage and tool usage at a high level, not per-teacher surveillance.
Just as important is what to stop measuring. Stop trying to “detect AI writing” as a primary metric; it is unreliable and encourages false certainty. Stop counting prompts or tool logins as a proxy for impact; high usage can mean confusion, not effectiveness. Stop aiming for a single whole-school “AI impact score”. You need a small dashboard by domain: policy compliance, classroom routine quality, and assessment integrity.
For artefacts, keep a one-page tool register with purpose and risk notes, two or three examples of redesigned assessments with process evidence built in, and a sample of staff-created materials showing the “AI draft → teacher edit” workflow. Add one pupil-facing guidance sheet that explains permitted use in plain language, and a short record of termly decisions: what you kept, stopped, or scaled, and why.
Prepare for consolidation. Schools will tire of tool sprawl, and the winning pattern will be fewer platforms with stronger controls, clearer pricing, and better reporting. Multimodal use will continue to grow, particularly voice and image input, which will help accessibility but increase safeguarding and privacy considerations.
Agent-like features will be the biggest leadership challenge. Even if you do not adopt full “agents”, more tools will offer automated workflows. Your 2026 readiness is less about predicting the next model and more about setting procurement rules: no adoption without an evaluation sprint, clear data boundaries, and an integrity impact note.
A lightweight way to do this is to run a one-week evaluation cycle with a small staff group and a defined rubric; the structure in [One-week evaluation sprint](/en-gb/blog/2025/09/05/openaidevday-2025-to-monday-uk-schools-one-week-evaluation-sprint/ adapts well to any vendor.
Use the template below as a single-page governor/SLT briefing. Keep it to one sheet, termly, and insist on evidence attached.
1) What is now settled (standardise):
Name 3–5 permitted staff uses (for example, planning drafts, adaptation drafts, feedback comment banks) and the checking routine expected.
2) What remains risky (tighten boundaries):
List 3–5 volatile areas (for example, take-home assessed writing, unmanaged pupil accounts, sensitive data entry, agentic actions).
3) Assessment integrity decisions (this term):
State which assessment types are AI-free, which are AI-conditional (with disclosure), and which explicitly assess AI-supported process.
4) Evidence pack (review termly):
Measures: time saved pulse check; assessment statements coverage; integrity incidents; training coverage.
Artefacts: tool register; redesigned task examples; pupil guidance sheet; termly decision log.
5) Controls and compliance:
Confirm data boundaries, account provisioning approach, and who approves new tools.
6) Next 30 days (actions):
Decide: policy refresh date; one assessment redesign pilot; one staff training focus; one evaluation sprint; one pupil communication.
May your next integrity review be calmer, clearer, and backed by evidence.
The Automated Education Team
Black Friday can feel like a rare chance to “save” on AI subscriptions, but …
December in schools brings a familiar spike: cover changes, heightened …
Microsoft Ignite can feel like a firehose of AI updates, but schools need a …
Report writing in 2025 is less about “which chatbot is best” and more about …
Three years into LGR22, many schools report real gains in clarity and …
Anti-Bullying Week works best when it moves beyond awareness and into …
Remembrance teaching asks for careful language, accurate sources, and …
Mock season often fails for predictable reasons: revision plans are …
Kolm aastat pärast ChatGPT avalikustamist on koolidel piisavalt kogemust, et hinnata, mis on …
Tre år efter ChatGPT:s lansering har skolor tillräckligt med erfarenhet för att bedöma vad som …
Kolme vuotta ChatGPT:n julkaisun jälkeen kouluilla on riittävästi kokemusta arvioida, mikä on …