EU AI Act Article 50: The Transparency Rules That Now Apply to AI-Moderated Course Evaluations

Bottom line up front: On 2 August 2026, Article 50 of the EU AI Act (Regulation (EU) 2024/1689) began imposing direct transparency obligations on the providers and deployers of certain AI systems — including the two kinds most relevant to modern course evaluation: AI systems that interact directly with people (a conversational AI interviewer) and AI systems that generate content (an AI-written summary of open-text feedback). These are not the high-risk obligations of Annex III; they are lighter, horizontal transparency duties that apply regardless of risk classification. But they are enforceable, they carry fines of up to €15 million or 3% of worldwide annual turnover (Article 99), and if your evaluation platform uses generative AI you are now in scope. This article sets out precisely what Article 50 requires, where universities as deployers pick up duties, and how to comply without abandoning the richer methods that conversational AI makes possible.

Why course evaluation is suddenly in scope

For years, the AI conversation in higher education quality assurance focused on whether student-feedback analysis might be "high-risk" under Annex III. That framing missed a nearer obligation. Article 50 is a transparency article that applies to broad categories of AI system whatever their risk tier. Two of its paragraphs land directly on the evaluation workflow:

• Article 50(1) — AI systems intended to interact directly with natural persons. Providers must design the system so that the people interacting with it are informed they are interacting with an AI, unless that is obvious to a reasonably well-informed and observant person. A conversational AI that interviews students about a module is the textbook case.
• Article 50(2) — generative AI producing synthetic content. Providers of systems that generate synthetic audio, image, video or text must ensure outputs are marked in a machine-readable format and detectable as artificially generated. There is an explicit carve-out where the AI performs an assistive function that does not substantially alter the input, but an AI that writes a thematic summary of hundreds of free-text comments is generating content, not just fixing grammar.

The Commission confirmed the timeline: the transparency obligations apply from 2 August 2026, with the European Commission having adopted guidelines on Article 50 and the AI Office publishing a voluntary Code of Practice on marking AI-generated content. Providers of generative systems already on the market were given until 2 December 2026 to meet the machine-readable marking requirement, but the disclosure duty for conversational systems is the immediate one.

What each obligation actually requires

1. Tell students they are talking to an AI (Article 50(1))

The disclosure must be given at the latest at the time of the first interaction, in a clear and distinguishable manner, and it must meet accessibility requirements. In practice: before a student begins an AI-moderated evaluation interview, the interface must state plainly that the interviewer is an AI system. The "obvious" exception is narrow — you should not rely on students inferring it from tone. A one-line, unmissable disclosure at the start of the session is the compliant default.

This obligation overlaps with, but is stricter than, the ethical case we made in should you tell students an AI is running their course evaluation. Article 50 converts a good-practice recommendation into a legal one.

2. Mark AI-generated feedback summaries as artificial (Article 50(2))

If your platform produces an AI-written synthesis — "the three themes across this cohort were assessment workload, lab access, and pacing" — that output is synthetic text and must be detectable as AI-generated. For a public-facing report there is an additional Article 50(4) layer where text is published to inform the public on matters of public interest. For internal QA reporting, the core duty is machine-readable marking plus clear labelling so that a committee reading the summary knows it was machine-produced and can trace it to the source comments.

This is where the provenance question we raised in AI summaries of course feedback and hallucination risk becomes a compliance question, not just a quality one: a summary that cannot be traced back to the underlying student comments fails both the reliability test and the spirit of the transparency regime.

Where universities pick up duties as deployers

Article 50 assigns most content-marking duties to providers (the vendor), but deployers — the university running the tool — carry their own obligations, and they cannot be contracted away entirely. The deployer is responsible for ensuring the disclosure actually reaches students in the live deployment, in the institution's languages, and in an accessible form. This dovetails with the wider deployer obligations under Article 26 and with the AI literacy duty under Article 4: staff who run evaluation cycles need to understand what the tool discloses and why. Practically, a quality office should confirm three things with any vendor before an evaluation cycle: (1) the first-interaction AI disclosure is present and accessible; (2) generated summaries are marked as AI-produced and traceable; and (3) the vendor can evidence its own provider-side compliance.

"But doesn't a disclosure just make students distrust the process — or stop responding?"

This is the strongest objection, and it is worth taking seriously. There is a legitimate worry that announcing "you are talking to an AI" depresses response rates or candour. Two responses.

First, the evidence on candour is not one-directional. Students often disclose more to a non-judgemental automated interviewer than to a human, a dynamic we examine in will students be honest with an AI interviewer. Transparency and candour are not necessarily in tension.

Second, and decisively: the choice is not yours to make. Article 50 is not an optional UX preference; it is a legal obligation with a supervisory-authority enforcement mechanism and turnover-based fines. The design question is therefore not whether to disclose but how to disclose well — framing the AI as a means of consistent, unbiased questioning rather than surveillance, and pairing the disclosure with a clear statement of purpose and data handling. A well-framed disclosure can increase trust by signalling that the institution takes both honesty and data protection seriously. That said, transparency is a floor, not a ceiling: Article 50 governs disclosure, while GDPR still governs lawful basis, and the emotion-recognition prohibition still constrains sentiment inference. Compliance is a stack, not a single checkbox.

What good compliance looks like — and how Koji is built for it

Koji for Education was designed for exactly this regulatory shape. Its conversational AI interviewer discloses at first interaction that it is an AI, in the student's language and in an accessible format — meeting the Article 50(1) duty by default rather than by bolt-on. Its automatic thematic analysis produces summaries that are labelled as AI-generated and, critically, traceable back to the underlying quotes, supporting both the Article 50(2) marking obligation and the provenance standard a QA committee needs. Because moderation is standardised and bias-aware, and because Koji handles data on a GDPR/AVG-appropriate basis, the transparency layer sits on top of a defensible lawful-basis and data-handling foundation rather than papering over gaps in it. The same AI interview engine underpins general user research on the main Koji platform; the education product wires it into the compliance obligations specific to student feedback.

To be precise about the claim: Article 50 compliance does not make an evaluation system trustworthy on its own — it makes it honest about what it is. The trust comes from what you do with the disclosure: consistent questioning, traceable summaries, and human oversight of any decision that affects staff.

The takeaway

Article 50 quietly changed the baseline. As of August 2026, any course-evaluation tool that talks to students or writes their feedback back to you must say so — clearly, accessibly, and at first contact. Treat this as an opportunity rather than a burden: a transparent, conversational, traceable evaluation process is both the compliant path and the better methodology.

Frequently asked questions

Does the EU AI Act apply to course evaluation software?

Yes, where the software uses AI. Article 50 transparency obligations apply from 2 August 2026 to AI systems that interact directly with people (such as a conversational AI interviewer) and to generative AI that produces content (such as an AI-written feedback summary), regardless of risk tier. Separate high-risk rules may also apply depending on use.

What does Article 50 require for an AI course-evaluation interviewer?

Under Article 50(1), students must be informed that they are interacting with an AI at the latest at the first interaction, in a clear, distinguishable, and accessible way — unless it is obvious. A plain disclosure at the start of the evaluation session is the compliant default.

Do AI-generated feedback summaries need to be labelled?

Under Article 50(2), synthetic content generated by AI — including text summaries — must be marked in a machine-readable format and be detectable as artificially generated. There is a carve-out for purely assistive edits that do not substantially alter input, but a thematic synthesis of student comments is generative output and should be labelled and traceable.

What are the penalties for non-compliance with Article 50?

Breaches of the transparency obligations can attract fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher, under Article 99 of the AI Act, enforced by national supervisory authorities.

Is the university or the vendor responsible?

Both. Providers (vendors) carry most content-marking design duties, but deployers (universities) must ensure the disclosure actually reaches students in the live deployment, in the right languages and accessible form, and should verify the vendor's provider-side compliance. This aligns with deployer obligations elsewhere in the Act.

Will an AI disclosure reduce response rates or honesty?

Not necessarily. Evidence suggests students can be more candid with a non-judgemental automated interviewer, and a well-framed disclosure can build trust. Regardless, disclosure is a legal requirement, so the practical question is how to disclose well — clearly explaining purpose and data handling — rather than whether to disclose.