Use of artificial intelligence
Version 2026-10-18, effective 18 October 2026. Information under Article 50 of Regulation (EU) 2024/1689 ("AI Act").
Feedback on your writing and speaking is written by an artificial intelligence, not by an examiner. Scores are practice estimates, not official results.
Where AI is used
| Feature | What the AI does | Which system |
|---|---|---|
| Writing feedback | reads your text and the task, estimates a score for each exam criterion, corrects mistakes, explains the corrections in your language and gives tips | a Claude model by Anthropic, through OpenRouter |
| Speaking feedback | the same on the transcript of your spoken answer, taking duration and pauses into account | as above |
| Speech transcription | turns your voice into text inside your browser; the audio is not sent to anyone | Whisper (open model), running on your device |
| Listening audio | the voices in the listening exercises are synthetic, computer-generated; they are not real people | Kokoro (text-to-speech) |
| Exercises | the exercises are original, written with the help of AI tools and checked by automated reviews |
Explanations for the reading, listening and grammar exercises are prepared in advance, not generated on the fly.
How to recognise generated content
- Every feedback screen is labelled "Feedback written by AI" at the top.
- In your history, AI-generated feedback carries the AI badge.
- In machine-readable form, feedback pages carry
<meta name="ai-generated" content="true">and the attributedata-ai-generated="true". Data you download with the export contains"ai_generated": true.
What it can and cannot do
It can help you spot recurring mistakes, understand why a sentence doesn't work and practise regularly, even at 2 a.m.
It cannot predict your exam result. The real exam has human examiners, its own marking grids and different conditions (time, stress, recording). The estimated score has a margin of error, which we show as a range.
It sometimes gets things wrong. It may:
- mark a correct form as a mistake, or miss a mistake;
- mistranscribe some words, especially with strong accents, background noise or poor microphones. That is why we always show you the transcript and ask whether it matches what you said;
- be less accurate on very short, very long or off-topic texts.
If some feedback looks wrong, write to support@civicoach.app: we'll check it, and it helps us improve.
What the AI does not do
- It decides nothing about you: there is no automated decision with legal or similar effects (Art. 22 GDPR). The score is only for you, to practise.
- It doesn't recognise you by your voice, and doesn't infer emotions, origin or other personal traits from how you speak or write.
- It doesn't learn from your texts: Anthropic does not train its models on data sent through its API.
- It is not connected with the certifying bodies and doesn't use their materials.
What data the AI receives
Only the task, your text or transcript, and the grading instructions. Never your name, email or account identifiers. Details (providers, transfers to the United States, retention) are in the privacy notice.
Classification
civicoach is an individual practice tool for adults. It issues no grades, certificates or attestations, doesn't assign levels for admission to courses or exams, and isn't used by schools or exam bodies to assess people. For these reasons we consider that it is not a high-risk system under Annex III, point 3, of the AI Act.
Version 2026-10-18 ·
civicoach practice booklet · not an official document