Run sheet, room setup, the twelve scenario cards with answer notes, the knowledge check key, and the two forms you must collect before anyone leaves.
Four hours of preparation. Skip the first item and the day still works; skip the second and it does not.
| Who | For which module | Why it matters |
|---|---|---|
| Designated safeguarding lead | M5, co-delivering | If they cannot attend, move the module. Do not run the safeguarding section without them. |
| Head of school | M1 open, M6 close | They name the accountable owner out loud at the end. The day lands differently when the commitment comes from the top. |
| Data protection lead | M4 | Often external. Brief them beforehand — most know GDPR well and the AI Act barely at all. |
| Every head of department | M3 | The struggle test is a subject-specialist judgement. Nobody else can make it for them. |
Timings assume a 09.00 start and a single 20-minute break. The right-hand column is what you do, not what you say.
Cabaret tables of six. Card sets face down. Projector tested with your screenshot open.
Slides 4–8. Three-runs demo, the confidence screenshot, the data columns, then the hallucination hunt.
Slides 9–12. Four rules, the red line, then red-amber-green with the card sets.
Say the M3 / M4 split twice before the break and once again as people leave.
Slides 14–16. Struggle test, assessment redesign, then rebuild one task in department groups.
Slides 17–18. Nine-question vendor screen run live against a real tool, then disclosure lines.
Slides 20–21. Rejoin as one room. DSL co-delivers the first section. Bias audit, then parent scripts.
Slides 23–25. Knowledge check marked as a room, commitment cards, then what happens next.
Cut the M5 parent scripts and shorten the M1 debrief. Never cut M6 — training that produces no artefact produces no evidence, and M6 is the entire evidentiary point of the day.
Print single-sided on card and cut. One set per table. Tables sort into banned, needs a decision, and fine. Do not tell them which three are contested — the friction is the activity.
A teacher uses AI to draft an email to a parent about a behaviour incident. She reads it, rewrites two sentences, and sends it under her own name.
The school is offered an exam proctoring tool that flags suspicious candidates by analysing facial expression and signs of stress during the paper.
A teaching assistant pastes a pupil's full SEN report into a free consumer chatbot on her personal account, to get a plain-English summary for a cover teacher.
A head of year generates all 180 of his tutor group's end-of-term report comments with AI from the grade data, reads each one, edits about a third, and signs them off.
During online lessons, a dashboard scores each student's attention and engagement from their webcam feed and shows the teacher a live ranking.
A science teacher generates twenty extra practice questions on moles for a revision sheet, checks the answers herself, and fixes three that were wrong.
Admissions uses an AI tool to rank applicants by predicted academic success, and the registrar interviews the top forty.
A wellbeing platform asks students to write a short check-in each Monday, and uses AI to analyse the written text for signs of low mood, flagging concerns to the pastoral team.
The marketing team publishes an AI-generated image of smiling students in the new prospectus. It is not labelled, and no real student is depicted.
A Year 12 student does not understand hypothesis testing. She asks a model to explain it four different ways, then writes her lab report entirely herself.
A teacher drafts a university reference for a student using AI, feeding in the student's grades and his own bullet-point notes. He edits it and signs it.
The MIS has a new feature that flags students at risk of underachieving, based on attendance and historical grade patterns. Tutors see the flags.
Nine of these sort themselves. Debrief only cards 4, 8 and 11, and write down where the room lands — those three become school policy.
| Card | Sort | Why |
|---|---|---|
| 01 | Green | She read it, edited it, and owns it. This is drafting assistance, and the human decision is intact. |
| 02 | Red | Inferring emotional or stress states from expression in an education setting. Article 5(1)(f). Not a procurement decision — a prohibition. |
| 03 | Red | Special-category personal data into an unapproved consumer tool. Banned by data protection law and school policy rather than by Article 5 — a useful distinction to draw out. |
| 04 | Contested | Run this one. He read and signed every comment, so the human decision survives. But 180 comments reviewed in one sitting is review in name only, and report comments feed a reported judgement. The room usually splits on volume, not principle. Ask: at what number does reviewing stop being reviewing? |
| 05 | Red | Attention and engagement inferred from webcam. The clearest breach on the deck. Nothing to weigh. |
| 06 | Green | She verified and corrected. This is the model use case. |
| 07 | Amber | Admission and assignment is an Annex III high-risk use. Obligations deferred to December 2027, so not banned — but it needs a written decision, human review of every rejection, and a place in the register today. |
| 08 | Contested | Run this one. The Article 5 prohibition is tied to inference from biometric data. Text is not biometric, so this probably falls outside it — but it is emotion inference on children in a school, at scale, with pastoral consequences. Legal and right are not the same question. Ask the room which one they are answering. |
| 09 | Amber | Article 50 transparency became enforceable on 2 August 2026. Add a label. Cheap to fix, embarrassing to be caught on. |
| 10 | Green | She used it to understand, then did the work. If your policy makes this cheating, your policy is wrong. |
| 11 | Contested | Run this one. No prohibition applies. But a reference is a personal testimony whose whole value is that a named human vouched for the student. Rooms divide sharply and often along age lines. Ask: would you tell the student? Would you tell the university? If not, why not? |
| 12 | Amber | Arguably assessing the appropriate level of education. Needs a decision, a named owner, and a rule that a flag never triggers an action on its own. |
The instinct is to give the room an answer. Resist it. Whatever the school decides on cards 4, 8 and 11 becomes policy, and policy the staff argued their way into is complied with very differently from policy they were handed. Minute the outcome and add it to the policy annex the same week.
Not a test. Where the room disagrees is where your policy is unclear — note those and put them on the review agenda.
2 February 2025. Not August 2026 — that is when national market surveillance authorities began supervising and enforcing it.
Watch for: staff assuming the obligation is new. It is not; only the checking is.Enforcement of the AI literacy duty by national market surveillance authorities, and the Article 50 transparency obligations becoming enforceable.
Watch for: confusion with the high-risk deferral to December 2027. Different articles.Using AI to infer the emotions of a person in an education institution. Article 5(1)(f). The exception is narrow — medical or safety reasons only, read restrictively.
Accept any answer that gets "emotions" and "banned". Precision on the article number is not the point.Attention or focus scoring from webcam · engagement analytics using facial expression · sentiment analysis of student faces or voices · proctoring that flags stress from expression · a wellbeing dashboard inferring mood from biometrics.
If someone names a tool your school actually runs, stop and write it down. That is the most valuable thirty seconds of the day.You are. Anything issued under your name is your responsibility, whatever produced the first draft. AI may support marking; a human makes and owns the decision.
That they are interacting with an AI system. Where someone could reasonably assume a human wrote or answered something, say otherwise.
Admission and assignment · evaluating learning outcomes, including steering the learning process · assessing the appropriate level of education · monitoring behaviour during tests.
Add that obligations are deferred to December 2027, and that the register you build now is the same register you will need then.Talk to the student. Ask about the work — how they got there, why they made a choice, what they would change. A detection score is not proof, and detectors are wrong most often for EAL writers and unusual voices.
This is the question that reliably splits a staffroom. Whatever the room concludes, write it down — it is your academic integrity procedure and it needs to exist before you need it.Any pupil personal data. And never safeguarding records, medical information, SEN reports or staff HR matters into any tool without written approval.
The accountable owner, within one working day. Reportable: incorrect output acted on, personal data exposed, a suspected prohibited practice, a disputed integrity decision, a safeguarding concern. Reporting in good faith carries no penalty.
The room should be able to say the name without looking it up. If they cannot, that is your finding.Print, collect, file. These two forms are evidence item E-05, and they are what turns a good day into a defensible one.
School ____________________________ Date ____________________ Facilitator ____________________________
| Name | Role | M1 | M2 | M3 | M4 | M5 | M6 | Card in |
|---|---|---|---|---|---|---|---|---|
Record M3 and M4 separately — the differentiated split is how you evidence that training was proportionate to role and technical knowledge, which is what Article 4 actually asks for.
File the attendance record, the commitment cards, the hallucination-hunt sheets, the photographed card sorts and the redesigned tasks. Update the register with everything named on the day. Minute the decisions the room reached on cards 4, 8 and 11 and add them to the policy annex. Diarise the annual review.