douce scroll
Legal · Regulation (EU) 2024/1689
AI Transparency Statement
Nearly everything you read and hear in douce scroll was written or spoken by a machine. We tell you so on every card, we mark it so software can detect it, and we give you a one-tap way to report anything wrong. Below: what we use, what we label, and what we are and are not under the EU AI Act.
1. Where AI is used
| Feature | What AI does | Labelled to you |
|---|---|---|
| Learning cards — vocabulary, expressions, culture, grammar | Writes the card, examples and explanations | Yes, on the card |
| Mini games and exercises | Generates items, distractors and answer keys | Yes |
| Spoken audio | Synthesises speech; no human speaker is recorded | Yes, on the player |
| Written answer grading | Assesses your answer and writes the feedback | Yes, on the feedback |
| Custom sets | Builds a set of cards from a topic you type | Yes |
| Word translation on tap | Translates the word in context | Yes |
| Placement test | Generates the questions; scoring itself is rule-based | Yes |
| Article and social cards | Summarises and glosses material from external sources | Yes, with a link to the original |
| Safety checks on what you post | Judges the text of a post, and any photo attached to it, before it is published. Classification only — it decides whether the post goes up, and writes nothing. | Yes, if it refuses |
| Feed ordering | Ranks and schedules cards. Rule-based and statistical, not generative. | Not applicable |
2. Our role under the AI Act
Under Regulation (EU) 2024/1689 (the AI Act), we are a provider of an AI system. We build douce scroll on general-purpose AI models supplied by third parties and place the resulting system on the market under our own name, which makes us the provider of that system under Art. 25(1)(a). We are also a deployer of the third-party models we call.
We are not a provider of a general-purpose AI model. We do not train, fine-tune or distribute models. The obligations in Chapter V of the AI Act fall on the model providers named in section 7, not on us.
3. Risk classification
We have assessed douce scroll against the AI Act's risk tiers. Our conclusion:
| Tier | Applies? | Reasoning |
|---|---|---|
| Prohibited practices (Art. 5) | No | We use no subliminal or manipulative techniques, no social scoring, no biometric categorisation, no emotion recognition, and no exploitation of age- or disability-related vulnerability. Streaks and XP are conventional, clearly visible progress feedback that you can ignore or switch off; they do not materially distort behaviour in a way that causes significant harm. |
| High-risk (Annex III, point 3 — education) | No | Annex III(3)(b)–(d) is confined to systems used in the context of or within educational and vocational training institutions. douce scroll is a consumer app for self-directed study; it is not such an institution and is not used by one. Annex III(3)(a) concerns determining access or admission to such institutions, which we do not do. Our placement estimate governs only which of our own cards you see. It grants no qualification and controls no access to education or employment. |
| Transparency obligations (Art. 50) | Yes | The system interacts directly with people and generates synthetic text and audio. Sections 4 to 6 below set out how we comply. |
| AI literacy (Art. 4) | Yes | Applicable since 2 February 2025. See section 10. |
We keep a dated internal record of this assessment and revisit it whenever we add a feature that changes how AI is used. If a future feature were to fall into the high-risk tier, we would meet the Chapter III obligations before shipping it, and by the applicable deadline of 2 December 2027 for Annex III systems.
4. Telling you it's AI
Art. 50(1) requires that people are told they are interacting with an AI system, at the latest at the point of first interaction. We do this:
- during onboarding, before you see your first card, in a screen you must acknowledge;
- on each generated card, as a persistent, visible label rather than something hidden in a menu;
- on the audio player, so it is clear the voice is synthetic;
- on AI-written feedback about your answers;
- in the Terms and here.
We disclose even where the AI involvement might be considered obvious. We would rather over-disclose than rely on the exception in Art. 50(1).
5. Machine-readable marking
Art. 50(2) requires providers to mark synthetic output in a machine-readable format so it can be detected as artificially generated. What we do:
| Output | How it is marked |
|---|---|
| Synthetic audio | Carries the SynthID watermark applied by Google DeepMind at generation. It is inaudible, survives MP3 compression, and is detectable by Google's verification tooling. |
| Generated text, in our API and shared pages | Every generated item carries provenance metadata recorded at generation — that it is AI-generated, the model family, and the timestamp. This is exposed as structured metadata on public card pages and in API responses, so a crawler or downstream tool can detect it without parsing prose. |
| Generated text, in the app interface | The same provenance record drives the visible label described in section 4. |
There is no established, interoperable watermarking technique for short passages of text that survives copy-and-paste — this is a known limit of the state of the art, and Art. 50(2) requires marking that is effective and interoperable as far as technically feasible, taking account of what is available. Our approach marks text wherever we control the surface it is served on. If you copy a card into a document, the marking does not travel with it. We track the EU Code of Practice on Transparency of AI-generated Content, finalised in June 2026, and will adopt a stronger technique for text as soon as a workable standard exists.
6. News and matters of public interest
Some cards are built from public news sources so you can read current events at your level. Where AI-generated or AI-summarised text concerns a matter of public interest, Art. 50(4) requires disclosure, since we do not put that text through human editorial review before publication.
Such cards are therefore labelled as AI-summarised and always carry a link to the original article. Read the original for the facts; read our version to practise the language. We do not generate synthetic images of real people or events, and we do not produce deepfakes of any kind.
7. Which models we use
| Provider | Used for |
|---|---|
| Anthropic (Claude) | Card and game generation, grading written answers, safety checks |
| Google (Gemini) | Card generation, word translation, safety checks |
| Google Cloud Text-to-Speech (Chirp 3: HD) | Spoken audio |
We route between these for cost and quality reasons and the mix changes. This page is updated when a provider is added or dropped.
8. Human oversight and correction
- Generated content passes an automated verification step — checking level, correctness signals and safety — before it can reach the feed. Items that fail are rejected and logged.
- Prompts you submit to the custom-set feature pass a safety filter that refuses illegal requests and requests for medical, legal or financial advice.
- Every card has a flag control. Flagged cards are reviewed by a human, and removed or corrected. This is the fastest way to get something fixed.
- Rejection reasons are monitored in aggregate so recurring failure modes get fixed at the prompt level rather than card by card.
9. Known limitations
We would rather state these plainly than let you discover them:
- AI can produce fluent, confident, wrong output. A grammar explanation may be subtly incorrect and read perfectly well.
- Regional variation is a persistent weakness. A phrase that is ordinary in one country can be odd or rude in another that shares the language.
- Register is hard. Content may be more formal or more casual than a native speaker would choose.
- Cultural content can flatten nuance or repeat a stereotype present in training data.
- Level estimates are approximate. Treat A1–C2 as a guide, not a measurement.
- Synthetic speech may mispronounce names, loanwords and dialect.
None of this is a reason to avoid the app — it is a reason to keep your own judgement switched on, and to flag what looks wrong.
10. AI literacy
Art. 4 AI Act requires providers and deployers to ensure a sufficient level of AI literacy among staff and others operating AI systems on their behalf. douce scroll is run by one person, who maintains that competence and keeps a dated record of it. This public statement, the in-app labelling and the limitations above are also how we discharge that duty towards you — an informed user is the point.
11. Your data and AI
Your written answers, custom-set prompts, tapped words, text queued for speech and your side of a role-play are sent to the providers above for processing. So is what you post — douce posts, the photos on them, and questions and answers you send to native speakers — for the safety check before it appears. Your email address, password, comments, voice recordings and progress history are not. Under those providers' business terms, data sent through their APIs is not used to train their models. The full picture is in the Privacy Policy, section 3.
12. Contact and review
Questions about how we use AI, or a card you think is wrong: ai@douce.live. For authorities, including market surveillance authorities under the AI Act: legal@douce.live.
We review this statement at least every six months, and whenever we change a model provider or add a feature that uses AI in a new way.