The EU AI Act's Transparency Rules Take Effect in August 2026 — What ESL Teachers Using AI Grading Need to Know

If you've used any AI grading or feedback tool in your classroom, you've probably never been asked to think about which EU regulation covers it. That changes this month. From 2 August 2026, the EU AI Act (Regulation (EU) 2024/1689) is fully applicable, including its transparency rules under Article 50 — and if you teach in the EU, or teach students who are, it's worth understanding what that actually requires from the tools sitting between you and your students' grades.

This isn't a call to panic about paperwork. Most of what the Act requires from a platform like an AI grading tool is already good practice: tell people when AI was involved, don't let it make unreviewed decisions about them, and be able to explain what it did. The point of this post is to translate the two provisions that matter most for classroom use — Article 50 and Article 6(4) — into plain language, and give you a short list of questions worth asking any vendor whose AI touches your students' work.

What actually changes on 2 August 2026

The EU AI Act phases in its obligations over several years, but 2 August 2026 is the date most of the rules relevant to a classroom platform take effect — including Article 50, the transparency requirement. In plain terms, Article 50 says that when an AI system is involved in producing something a person sees or is assessed on, that involvement has to be disclosed. Not buried in a terms-of-service document — visible, at the point where it matters.

For AI grading and feedback tools, that means a student reading AI-influenced feedback, or seeing an AI-assisted score, is entitled to know that AI was part of producing it. It's a low bar in principle, but it's not automatic — a tool has to be built to surface that disclosure, not just to comply with it on paper.

There's a second layer worth knowing about, even though it's less visible day to day: the Act distinguishes between a "provider" (the company that builds the AI system) and a "deployer" (whoever uses it — in a classroom context, that's you). Providers carry most of the technical obligations, but deployers aren't passive. As a deployer, you're expected to exercise the human oversight the system is designed to allow, and not to treat an AI-generated score or piece of feedback as final just because it's convenient to. In practice this usually just means: read what the AI drafted before you act on it, rather than rubber-stamping it because the class set is long and Friday afternoon is close.

What "not high-risk" actually means

The other provision worth understanding is Article 6, specifically Article 6(3) and 6(4), because it's the one vendors will cite when they tell you their tool is compliant. The Act sets out risk tiers, and education is one of the sectors called out in Annex III as a place where AI systems evaluating learning outcomes can be classified as high-risk — which comes with a much heavier set of obligations (conformity assessments, technical documentation, ongoing monitoring).

But Article 6(3) carves out an exception: a system in an Annex III category isn't high-risk if it's intended to improve the result of a previously completed human activity, or to perform a preparatory task ahead of an assessment — provided it doesn't materially influence the final decision without human oversight. In practice, this is the difference between "the AI decides the grade" and "the AI drafts a suggestion the teacher has to actively review and confirm." The first is high-risk. The second, done properly, isn't.

That word "properly" is doing a lot of work. Article 6(4) requires that this self-assessment be documented, not just asserted — and that's the part worth asking about, because it's easy for a vendor to say "a teacher reviews everything" in marketing copy and much harder to show that the product actually enforces it.

This distinction matters beyond a single classroom. If your school or language centre is procuring an AI tool rather than a single teacher picking one, Article 6(4) status is genuinely a procurement question — it determines whether the institution inherits high-risk obligations (conformity assessments, a registered technical file, ongoing monitoring duties) just by deploying the tool. A vendor who can't produce a documented classification is asking your institution to take that risk on faith.

Questions worth asking any AI grading vendor

Given those two provisions, here's what's actually worth asking before you trust a tool with your students' work — not as a compliance checklist, but because the answers tell you how the tool actually behaves:

• Can a score or piece of feedback reach a student without me taking an action first? If the answer is yes, ask what that action is and whether it can be skipped.
• Is there a documented risk classification, and can I see it? A vendor that has genuinely done the Article 6(4) assessment should be able to point you to it, not just describe it verbally.
• Where is my students' writing processed, and for how long is it kept? This is a GDPR question as much as an AI Act one, but the two overlap heavily in an EU classroom context.
• How is AI involvement disclosed to students, specifically? "It's in our privacy policy" isn't the same as a visible label on the feedback itself.
• What happens if I disagree with an AI-suggested score? You should be able to override it, not just flag it.

None of these are trick questions. A platform that's actually built around teacher oversight should be able to answer all five without hedging.

A worked example

Writing, no kidding went through this classification process, so it's a reasonably concrete example of what the answers can look like in practice. The platform completed a formal Article 6(4) risk classification and concluded it is not a high-risk system — full detail is on the AI Transparency page — on the basis that AI output is never binding: rubric scores and written evaluations are shown to the teacher as suggestions, and the interface itself enforces the review step rather than relying on the teacher remembering to do it. The "Save rubric" action is blocked until the teacher has clicked a level for every AI-scored criterion, and the AI-suggested grade is shown read-only — the teacher has to enter their own grade before a review can be submitted.

On the Article 50 side, students see a disclosure whenever AI corrections or AI rubric scoring were used on their work, and AI-suggested rubric levels are marked distinctly from teacher-confirmed ones, so the distinction is visible in the interface, not just documented somewhere. Every time a teacher overrides an AI-suggested rubric level, that override is recorded — the AI's original suggestion, the teacher's final decision, and when it happened — which is less about any single lesson and more about being able to show, months later, that human review actually happened on a given submission rather than being a formality that got skipped under deadline pressure.

On data handling, AI requests are processed in the EU/EEA rather than routed elsewhere, student names aren't included in what's sent to the AI model, and a GDPR Data Protection Impact Assessment covering the AI features has been completed as part of the same review — currently awaiting final data-protection-officer sign-off, alongside a Fundamental Rights Impact Assessment, both expected before the Act's application date. The Privacy Policy covers the data-flow specifics, and the broader trust posture — including how the same "teacher decides, AI assists" principle carries through to the academic integrity side of the platform — is summarised on the homepage. Worth naming plainly: this kind of review is ongoing work, not a box that gets ticked once. New AI features get run through the same risk-classification process before they ship, not after.

None of that is a guarantee against every possible future interpretation of the Act — this is a new regulation and guidance will keep evolving. But it's the kind of answer a vendor should be able to give you in specifics, not in reassurance.

The bottom line

You don't need to become a compliance officer to use AI grading tools responsibly this term. You need to know that a real classification exists, that a human — you — stays in the loop on every score a student sees, and that "AI-assisted" is something your students are actually told, not something buried in a document nobody reads. Ask the five questions above of whatever tool you're using, and you'll know within a few minutes whether you're looking at a compliance claim or a compliance design.

---

See the full risk classification behind Writing, no kidding. Read the AI Transparency page