EU AI Act compliance for EdTech: the audit trail for AI in education and assessment

If your product determines admissions, grades exams, places learners at a level, or monitors students during tests, Annex III, point 3 of the EU AI Act puts you in scope as high-risk. The full high-risk regime is in force. The penalty band for breaches of the operational obligations is €15 million or 3% of worldwide annual turnover.

EdTech sits in an uncomfortable position. The decisions your AI makes affect access to education, grades that go on official transcripts, and assessments that get challenged in writing by parents and lawyers. A wrong score is a regulatory artefact, not a backlog ticket.

What the Act expects from education AI

Annex III is specific. The high-risk category covers AI for:

• - "Determining access or admission or assigning natural persons to educational and vocational training institutions"
• - "Evaluating learning outcomes, including when those outcomes are used to steer the learning process"
• - "Assessing the appropriate level of education that an individual will receive or will be able to access"
• - "Monitoring and detecting prohibited behaviour of students during tests"

A reading-comprehension scorer, an essay grader, a level-placement quiz, a remote-proctoring computer-vision model — all in. A friendly chatbot that summarises a textbook is probably not; an AI that decides whether a learner advances to the next module is.

The audit trail you have to produce is built from the same articles every high-risk system has to satisfy:

• - Article 11 + Annex IV — technical documentation including intended purpose, training data, accuracy metrics, known limitations, and instructions for use. Kept for ten years.
• - Article 12 — automatic event logs over the lifetime of the system.
• - Article 14 — designed for effective human oversight, with the ability for a person to override or reverse outputs.
• - Article 19 / Article 26(6) — log retention of at least six months.
• - Article 72 — post-market monitoring system.
• - Article 73 — serious incident reporting within 15 days.

For schools and universities deploying your tool, Article 26 adds a clear obligation: assign human oversight to a competent person, monitor operation, and keep the logs.

Why "the model graded it" is not a defence

A parent emails. The essay was scored 47. The teacher says the work was a clear high pass. The parent asks to see the rationale. The product team checks the database. There is no rationale stored. There is no version history. There is no record of which prompt version graded the essay, what rubric was applied, or what the model's confidence was. The conversation that should have ended in a one-paragraph explanation ends in a complaint to the national supervisory authority.

The Act does not care that the model is good on average. It cares that you can document why this score was produced, and that you can show that the system performs equitably across protected characteristics.

The villain is the same villain as everywhere else: Ship and Pray. Three good test essays, a passing demo, a launch. The team finds out from parents which essay types break the model. In an Annex III category, that workflow is no longer legal.

The EdTech audit-trail checklist

For each AI feature touching admissions, assessment, placement, or proctoring, you should be able to produce:

1. - Annex IV technical documentation — including accuracy and bias metrics broken down by age, first language, special-educational-needs status, and any protected sub-group your evaluation set captures.
2. - Versioned prompts, rubrics, and model configurations — every change captured with date, author, rationale, and the evaluation that justified it.
3. - Evaluation reports — per-version, on a representative dataset, signed off by the curriculum or assessment expert who defined the criteria. Includes the score distribution across sub-groups.
4. - Runtime event logs — each scoring decision, the input, the rubric, the model version, the resulting score, the rationale text, and any human review.
5. - Human-override records under Article 14 — what teachers or assessors changed and why.
6. - Post-market monitoring data — periodic re-evaluation, drift detection, evidence that the model still performs as documented.
7. - Serious-incident register — any event that infringed fundamental rights or caused serious harm.

If a learner challenges a result, that record set is what you hand to the school's legal team. If a supervisory authority audits the system, that record set is what you hand them. There is no third option.

The deployer side: what schools will start asking for

European schools and universities procuring AI tools for high-risk uses are deployers under the Act. Article 26 requires them to assign human oversight, keep logs, and inform learners that AI is involved. Article 27 requires the Fundamental Rights Impact Assessment for public deployers.

In practical terms, the procurement questionnaire is changing. Schools will start asking:

• - Can you provide Annex IV technical documentation?
• - Can you export the audit trail for a given learner on demand?
• - What is the documented accuracy of your model on learners with English as a second language?
• - What is the human-override workflow, and how is it logged?
• - What is your post-market monitoring cadence?
• - How do you handle a serious incident — and what is the maximum time between event and report?

EdTech vendors that cannot answer those questions in writing will start losing tenders. EdTech vendors that can will win them.

Where domain experts have to lead

The Act calls for human oversight. In EdTech, the people qualified to provide it are teachers, curriculum leads, assessment specialists, and accessibility experts. They know what a rubric should produce. They know which essay-type the model will fumble. They know what bias looks like in their cohort.

Standard evaluation tooling shuts those people out. The eval set is a notebook only engineering can open. The rubric is a YAML file. The dashboard is built for ML engineers. The curriculum lead's only channel is to email a screenshot to engineering and wait two weeks for a response that may never come.

Lovelaice flips the role. The assessment expert writes the criteria for what acceptable looks like before the system runs against data — completeness against the rubric, factual correctness, consistency with the marking guide, absence of penalty for protected characteristics. They review outputs directly through a simple interface, rating and flagging issues. Their decisions are the audit trail.

Every evaluation is captured. Every prompt version is immutable. Every result is exportable. When the school's compliance officer asks how you know the model grades equitably for learners with dyslexia, you do not improvise. You export the run.

Retention and reporting windows to design for

• - Six months minimum log retention. Likely extended by national education-data laws and GDPR.
• - Ten years retention of technical documentation and the EU declaration of conformity.
• - 15 days to report a serious incident — 10 if a fundamental-rights infringement is widespread.

Education is where the consequences land directly on people who cannot choose another vendor. The audit trail is the work. Start producing it before the procurement questionnaire asks for it.

Sources

• - Regulation (EU) 2024/1689 — official text (EUR-Lex)
• - Annex III, point 3 — education and vocational training
• - Annex IV — technical documentation
• - Article 11 — technical documentation
• - Article 12 — record-keeping
• - Article 14 — human oversight
• - Article 19 — automatically generated logs
• - Article 26 — deployer obligations
• - Article 72 — post-market monitoring
• - Article 73 — serious-incident reporting
• - Article 99 — penalties

More in this series

The EU AI Act audit-trail series — one article per Annex III high-risk category:

• - Biometrics: remote identification, categorisation, and emotion recognition
• - Critical infrastructure: safety-component AI
• - Credit scoring and insurance: essential financial services
• - Employment and HR: recruitment, ranking, and workforce management
• - Law enforcement: risk assessment, polygraphs, and profiling
• - Migration, asylum, and border control
• - Justice and democratic processes