The European Union’s AI Act is beginning to turn broad principles about artificial intelligence into practical obligations. For teachers and school managers, one of the most important ideas is transparency.

At first sight, this sounds simple. If a school uses AI, people should be told. But transparency can mean several different things. It might mean telling a student that they are interacting with a chatbot. It might mean giving a school enough information to understand a supplier’s system. Or it might mean explaining why an AI-supported decision affected a learner.

These are not the same thing. The real question is not simply whether AI is transparent. It is who needs what information, when do they need it, and what can they do with it?

What the EU AI Act says

The AI Act does not require every AI system to publish its source code or training data. Instead, it creates different obligations depending on what the system does and how serious its effects may be.

From August 2026, providers of AI systems that interact directly with people must generally make clear that the user is dealing with AI, unless this is already obvious. Systems that generate images, audio, video or text must also make their outputs identifiable as artificially generated where this is technically possible.

For schools, the more important provisions concern high-risk uses. The Act identifies systems used for admission, access or allocation; the evaluation of learning outcomes; assessment of a learner’s educational level; and the detection of prohibited behaviour during tests. These systems are considered high risk because they can influence a person’s educational and professional future.

This does not mean that every classroom chatbot, writing assistant or learning platform is automatically high risk. The classification depends on the system’s purpose and actual role in the educational process.

The timetable is also important. The full set of high-risk obligations for educational systems has been postponed until December 2027. Schools should therefore avoid both extremes: assuming that nothing needs to be done yet, or assuming that every educational AI tool is already subject to the complete high-risk regime.

When the full requirements apply, providers will have to give institutions information about the system’s intended purpose, performance, limitations, risks, data requirements, logging and human oversight. Schools will need to understand how the system is meant to be used and where it may fail.

The Act also requires AI literacy. For education, this should mean more than learning how to write effective prompts. Teachers and managers need to recognise unreliable outputs, automation bias, privacy risks and the limits of a system’s approved use. The Act also prohibits most uses of emotion-recognition AI in education institutions. Tools claiming to measure attention, engagement or stress should therefore be treated with particular caution.

There is a further safeguard for people affected by certain high-risk decisions. In defined circumstances, they may have a right to a clear and meaningful explanation of the AI system’s role in the decision and its main elements. This is important, but it is not a general right to an explanation of every automated recommendation, chatbot answer or teacher judgement.

The General Data Protection Regulation provides an additional layer of protection where personal data are involved. In some cases of solely automated decisions with significant effects, people must have an opportunity for human intervention and to challenge the decision. A member of staff who simply accepts an AI recommendation without examining it is unlikely to provide meaningful human oversight.

What will implementation mean in practice?

The immediate implication is that schools should stop asking only, “Is this AI tool approved?” They should also ask what the tool does, what evidence supports it and what happens when it is wrong.

Before introducing an AI system that affects assessment, admissions, placement, discipline or learner support, a school should be able to identify its educational purpose, the learners affected, the data it uses, its known limitations and the person responsible for the final decision. The school should also know how a teacher or learner can challenge an output.

Contracts with suppliers will matter. A school should expect information about system updates, performance, data retention, security, subcontractors, incident reporting and access to relevant records. It should not promise teachers or learners transparency that it cannot obtain from the supplier.

The principle of human oversight also needs to be taken seriously. A teacher cannot meaningfully oversee a system if they have no time to examine its recommendation, no access to the relevant evidence and no authority to reject it. “Human in the loop” should not become a phrase that hides automatic decision-making.

For students and families, transparency should be clear and timely. A general statement buried in a privacy policy is not enough if AI has helped determine a grade, placement or disciplinary outcome. People need to know what role AI played and how they can seek correction or review.

The danger of transparency theatre

Transparency can easily become another form of paperwork. A school may publish an AI policy while staff do not know which tools are actually being used. A supplier may provide a detailed model card that describes general performance but says little about how the system works with local learners. An impact assessment may be completed after a procurement decision has already been made.

This is transparency theatre: information is made visible without giving anyone the power to understand, question or change the decision.

The solution is not necessarily to publish everything. Publishing source code or training data could create privacy, security and commercial risks. A better approach is layered transparency. The public needs basic information about the purpose of a system, who is responsible and how complaints can be made. Teachers and managers need practical information about performance, limitations, data and oversight. Regulators and independent auditors may need deeper technical access. Learners need a clear explanation of how the system affected them.

Transparency only becomes meaningful when it is connected to action. Can a teacher override the system? Can a learner correct inaccurate information? Can a manager suspend a tool when repeated errors appear? Can an auditor reconstruct what happened in a disputed case?

A wider political question

This matters because the debate about AI is increasingly shaped by the executives who develop it. Some company leaders warn that advanced AI could threaten civilisation. Others present the same technology as an almost limitless solution to economic and social problems.

Brian Merchant describes this combination as a form of “doom marketing”. His argument is not that every warning is dishonest. Some researchers and executives clearly believe that advanced AI could create catastrophic risks. His concern is that presenting AI as all-powerful, either for good or for harm, can increase the authority of the companies selling it.

The appropriate response is neither to dismiss serious risks nor to allow private forecasts to become public policy without evidence. Companies that make dramatic claims should be expected to provide testing methods, evidence, incident records and independent access to their safety assessments.

Merchant’s reporting on proposed copyright-transparency legislation offers a useful contrast. Instead of asking whether the public trusts an AI company, it asks what the company can document about the material used to train its systems and whether people whose work is affected can obtain relevant information.

This may be the more practical direction for education. Schools should not have to choose between believing a supplier’s promises and rejecting every new technology. They need evidence on which to make proportionate decisions.

Looking ahead

The next stage of AI regulation should move beyond the simple principle that systems must be “transparent”. It should require a chain of responsibility.

Providers should document what their systems are designed to do, how they have been tested and where they are known to fail. Schools should document why a system is being used, who is affected and what alternatives were considered. Teachers should have the training, time and authority to challenge outputs. Learners should receive understandable information and a practical route to appeal. Regulators and independent experts should have access to the evidence needed to investigate serious problems.

This will not make AI automatically accurate or fair. Transparency cannot prove that a system is educationally effective. It cannot remove bias simply by describing it. But it can make unsupported claims easier to challenge and harmful decisions easier to investigate.

For teachers and school managers, the most useful question is therefore not, “Is this system transparent?” It is: “What information will we receive, what will our learners receive, and what power will that information give us?”

That is a more modest ambition than opening the entire black box. It is also much more likely to produce responsible use of AI in education.

References

[1] Regulation (EU ) 2024/1689, Artificial Intelligence Act, consolidated text

[2] Regulation (EU ) 2026/1744 amending the EU Artificial Intelligence Act

[3] Regulation (EU ) 2016/679, General Data Protection Regulation

[4] Brian Merchant, Afraid of AI? The Startups Selling It Want You to Be

[5] Brian Merchant, AI Disagreements

[6] Brian Merchant, There Has to Be a Way

This article is an analytical discussion of policy and practice, not legal advice. Schools should consult the current official legislation and national guidance when making decisions about particular AI systems.