AI transparency

1. Where we use AI

• The materials: exercise texts, instructions, explanations, images and audio recordings (synthetic voices) are prepared in advance with the help of AI. They are produced editorially, with no user data at all, and we review them before publication — they may nonetheless contain errors.
• AI marking: your written work, your picture descriptions and your short recordings are assessed automatically by an AI system. The marking is a learning aid: it can be wrong, it is not an official exam result, and we disclose it to nobody but you.
• What the AI does not do: it does not decide what you are shown, it builds no profile of you, and it does not predict how you will do in the exam. Automatically marked exercises (reading, grammar, listening) are assessed by comparison with a prepared answer key, with no AI involved.

2. How we mark AI-produced content (Art. 50 AI Act)

• A notice before every interaction with the AI: before you even send your answer, the exercise tells you that an AI system will assess it (Art. 50(1) AI Act). The notice sits next to the control itself, not buried in this document.
• A note beside every marking: "Automated AI marking — this is not an official exam result."
• A synthetic-voice note on the listening exercises.
• A machine-readable mark: images produced with AI carry the standardised term "trainedAlgorithmicMedia" in their metadata (IPTC/XMP), and audio files carry the same term in their ID3 metadata. The AI marking output is additionally marked in the page itself by an attribute reserved for AI-produced content. Pages carrying such content declare the term "ai-generated-content" in the document head with the same standardised value, and media files as well as the exercises’ data files are served with the response header "X-AI-Generated", whose value is that same standardised identifier.
• What is not marked, and why: the Site’s wordmark is an image, chosen by the publisher from AI-generated candidates, and published as the service’s mark of origin rather than as content produced for the learner — so it is not marked, following the same convention our other sites use for their own wordmarks. The favicon, the home-screen icon and the social preview image are still drawn directly by the page code from a piece of text and a simple colour gradient: no model output enters them. Nor are this page, the Terms of Service and the Privacy Policy marked: they are not content within the meaning of Art. 50, they are written by us, and the marking duty does not apply to them.
• The case of text: there is as yet no single technical standard for marking text. The attribute in the page and the term in the document head are our documented best effort in the light of the state of the art, not a claim to a verifiable watermark.

3. Our role and our classification under the AI Act (Annex III, Art. 6)

Our role: we offer the AI marking system under our own name, so we are its provider within the meaning of Art. 3(3) AI Act — the fact that it is free of charge changes nothing, the definition says so expressly. The third-party AI models with which the materials are prepared in advance and your answers assessed we merely use; their provider is the processor named in the Privacy Policy (OpenAI Ireland Ltd.), and speech synthesis is carried out by a further provider that receives nothing but our own script texts. We name no particular model: it changes, and naming one would publish a fact that goes stale.

The AI Act places certain AI systems used in education among the high-risk systems (Annex III, point 3(b)): systems intended to evaluate learning outcomes. We checked whether our marking function falls within it.

The result of our documented assessment: the function is not a high-risk system within the meaning of Art. 6(2) AI Act, because the exception in Art. 6(3) applies. It is a voluntary, free aid to self-study; the marking is purely formative, it produces no legal or comparable effect, it leads to no certificate, it conditions access neither to education nor to an exam, and it is disclosed to no exam organiser and to no third party. A single, voluntarily submitted answer is assessed at a time; no profiling takes place. The exam that counts is taken before independent human examiners at the official bodies.

We recorded that assessment in writing in accordance with Art. 6(4) AI Act and will produce it to the competent authorities on request. It is our own documented assessment, not a lawyer’s or an authority’s confirmation. Art. 6(4) AI Act attaches to such a documented assessment the registration duty in Art. 49(2), in the EU database: we will discharge it within the statutory deadlines. If our assessment were ever to change, the function would fall under the registration duty in Art. 49(1).

Since when we mark: the transparency obligations of Art. 50 AI Act (Chapter IV) have applied since 2 August 2026. The amending Regulation (EU) 2026/1744 defers only Sections 1 to 3 of Chapter III (high-risk systems); Art. 50 is not covered by that deferral. We therefore mark from the moment the service is publicly reachable, making it available to the public being the placing on the market. We rely on no later date.

4. What we ask of you — and what we do not

Please do not include, in your answers or in your recordings, data falling within the special categories of Art. 9 GDPR (for example health data, origin, religious beliefs or political opinions), nor data about other people: we do not need it for the marking. This is a request, not a filter: the system neither searches for nor separates out such data, your answer is transmitted in full and assessed as a whole. The request sent contains neither your name, nor your email address, nor any other user identifier.

Whether and how you use our materials is entirely your decision. AI marking replaces neither a teacher nor an official exam result: only sitting the exam at an official body produces a certificate. We are an independent private project with no connection to any examination centre or public authority.

A question, or an error to report? Write to us at contact@francaisa2.fr

5. Further information

• How our materials are produced and what we answer for: Terms of Service
• How data is processed during AI marking, and the controller’s details: Privacy Policy