Legal
AI transparency notice
The German version is legally binding. This English version is a translation provided for information.
Binding version
The German version is authoritative; the English version is a translation provided for information. Version 2.0, effective from 18 September 2026.
What this is about
Artificial intelligence lends a hand in a few places in Tellersonne. We think you should know where that happens, what is decided there and who is ultimately responsible. This notice explains it in plain language and supplements our privacy policy (https://tellersonne.app/en/privacy).
Our role under the AI Act
The EU AI Act (Regulation (EU) 2024/1689) distinguishes between those who provide AI systems and those who use them. We do not develop AI models, we use models made by others - which makes us a deployer within the meaning of the regulation. We meet the transparency obligations of Article 50 with this notice and with the labels in the app. We do not operate a high-risk system under Annex III: we do not use AI to decide about access to education, work, credit or public services.
Where we use AI
- Recipe drafts: our editorial team has drafts written by a language model of our external AI provider - the idea, the ingredient list, the steps. A draft never goes live automatically. - Translation of editorial texts: we translate recipes, themes, help texts and app texts between German and English by machine and have them reviewed by people afterwards. - Checking photos: photos you post to the community or share with your household are checked for impermissible content by a vision model at Cloudflare (Workers AI) and, if that is unavailable or the result is unclear, by a model of our external AI provider. - Translating your texts: where needed we translate comments, captions, image descriptions, reports, complaints and support messages automatically so they can be read in both app languages. - Determining the language of a text: a simple method without AI decides first. If the language stays unclear and the text is long enough, we additionally ask a model of our external AI provider. - Comment pre-filter: a filter made of fixed rules (not an AI model) recognises typical advertising and insults and can only hold a contribution back provisionally.
Where we do not use AI
There is no AI chat and no AI assistant you talk to in Tellersonne, so Article 50 (1) of the AI Act does not apply. We do not use AI to assess you, to predict your behaviour or to set prices. The recommendations under "For you" come from your profile settings, the recipes you saved most recently, your cooking history and fixed rules - not from an AI model. Voice control in cooking mode uses the speech recognition of your operating system; neither audio recordings nor texts reach us. Nutritional values come from an official database, not from a model.
AI-generated images and videos
At present we do not use AI-generated photos or videos. Should we do so later, we will visibly label them as artificially created in the app and on the recipe pages - at the latest when they are first perceived and readable for screen readers too, as Article 50 (4) of the AI Act requires. We do not remove provenance data in the files (for example Content Credentials under C2PA).
What is labelled in the app
- Machine-translated contributions carry a "Translated" label and can be read in the original with "Show original". - If a photo or a comment is not visible because of an automatic check, we tell you so with the note "Under review" and send you a statement of reasons once a decision is made. - As soon as we use AI-generated images, they carry a visible "AI" marker at the image that is readable for screen readers as well. An individual recipe carries no AI marker today: that our editorial team works with AI drafts is disclosed here, and every recipe is reviewed by a person and test-cooked before publication. Article 50 (4) subparagraph 2 of the AI Act requires no labelling for such texts, because a person carries the editorial responsibility.
AI can get things wrong
AI models invent things, miscalculate and miss connections. That is why every published recipe is reviewed by a person and test-cooked before it goes live; nutritional values come from an official database and not from the model; allergens are checked separately. Machine translations can miss nuances; in case of doubt the original text applies, and moderation always judges the original too. The automatic photo check can be wrong - which is why it only publishes on a clear result and only rejects at high confidence. Please never rely on the app alone where allergies and intolerances are concerned, and read the packaging of the products you use.
People decide about people
No AI system takes a final decision about you here. The automatic check can release a photo that is clearly harmless or hold back a photo that is clearly not permitted; in every other case a person from our team decides. Warnings, comment bans and account locks are always issued by a person. For every restriction you receive a statement of reasons naming the ground, the rule that was broken and whether detection and decision were automated. You can lodge a complaint against every decision and ask for a review by a person; how to do that is described at https://tellersonne.app/en/dsa.
Which providers we use
- Cloudflare (Workers AI): checking photos and translations; runs inside the infrastructure of our host. - OpenAI Ireland Ltd, Dublin, Ireland, with OpenAI OpCo, LLC in the United States as a sub-processor: our external AI provider for recipe drafts, translations, language detection and the review of unclear photos. Both providers process the data on our behalf, and we have concluded data processing agreements with them. Your photos and your texts are not used there to train models. Which data goes where, and what we base the transfer on, is set out in the privacy policy: https://tellersonne.app/en/privacy. If we change provider, we update this notice with a new version.
What happens to your content
For the check we send the photo, and for the translation the text, to the provider concerned. Before that we replace e-mail addresses, links and token-like strings in texts with placeholders, and the instruction to the model explicitly says not to follow instructions contained in a text. All that comes back is categories and a confidence value for the check, and the translated text for the translation. Translation only happens when somebody reads your text in the other language, not when you write it. We keep translations in a cache for up to 90 days so the same text is not translated twice. Your original text is always kept and never overwritten.
When a provider is unavailable
If a model fails or our daily cost limit is reached, nothing is published silently: a photo stays invisible and goes into our team's queue, and a translation simply does not happen - you then see the original. With no provider at all, only a person reviews.
Reaching a human
Would you rather talk to a person than to a procedure? Write to support@tellersonne.app or use the app under Profile, Help, Feedback (https://tellersonne.app/en/help). We reply personally. The same applies if you believe an automatic check has treated you unfairly.
Changes to this notice
We update this notice as soon as we use AI in a new place or change providers. Every version carries a version number and a date, and earlier versions remain visible in the version history.