AI Transparency Notice

Last updated: 8 October 2026

Sprichst contains an AI tutor. This notice tells you what it is, how it works, where your messages go, and what to watch out for. It supports the transparency duties of the EU AI Act (Article 50) and the GDPR.

You are talking to an AI

The coach, its corrections, explanations, follow-up questions and spoken replies are generated by artificial intelligence, not written or spoken by a person. The voice you hear is a synthetic voice. The app labels AI content as AI-generated wherever it appears.

What it is used for

Correcting your German, explaining grammar and vocabulary, practising conversation, transcribing your speech, and reading replies aloud. It does not make decisions about your rights, grades, work or access to services. The lesson plan and "readiness" figures are calculated by ordinary code in the app from your own answers, and you can ignore them.

The models

On your phone (the default, once downloaded): Qwen2.5 and Gemma 3 language models, Whisper speech recognition, and Piper German voices. They run locally. Nothing leaves your device.

With your own AI key (optional): your device sends the request straight to the provider you chose (Groq, Google Gemini, OpenAI, Anthropic Claude, xAI Grok, Mistral, DeepSeek or OpenRouter, or another service that works like OpenAI), which runs its own model for the answer and, where it offers one, speech recognition; providers without it leave speech recognition to your phone. Sprichst runs no AI server and does not see the request. No provider has a German voice the app uses, so spoken replies always come from the voices on your phone. In a web browser the models cannot run, so the web version needs your AI key.

What the model sees

Your message or recording, the last few messages of the current conversation (so the tutor does not ask the same thing twice; they are forgotten when you leave the coach screen or reset progress), plus a short learning summary: your CEFR level, current unit and lesson, a few weak skills, some known words and recent mistakes. It does not see your name, email or account. Use "This phone only" in Account → AI & voice if you want nothing to leave your phone. When your provider's allowance runs out, your key is rejected or you are offline, the app tells you and asks before it answers on your phone instead; it never switches without asking.

Limits and risks

• It may produce wrong, biased, or inappropriate output. Answers can look convincing and still be incorrect.
• Small on-device models are weaker than the larger models cloud providers run.
• It may misunderstand an accent, background noise, or a dialect.
• Do not enter secrets, health data, or other people's private information. You do not need to give any for learning.
• We filter and size-limit what is sent, but we cannot guarantee that the model's output is safe in every case.

Training

The translator works the same way: the text you type or paste is translated on your phone, or sent straight to the provider you chose with your own key. It is not kept by the app, and translations can contain mistakes, so check anything important.

We do not train or fine-tune any model on your messages, recordings or learning data. With your own key, how the provider treats what you send is governed by your agreement with it; check its data and training settings in your account there.

Human oversight and reporting problems

Content is not reviewed by a person unless you report it. If an answer is harmful, wrong in a way that matters, or you think the AI misbehaved, tell us at sripathiyadav66@gmail.com (include the date and what you saw). Account → Contact developer prepares a report you can copy and send.

Your rights

You have the rights listed in the Privacy Policy and Your Rights, including to object to processing, to ask for human review of anything that affects you significantly (we make no such decisions), and to complain to a supervisory authority.

Licences of the models

Qwen2.5 0.5B and 1.5B: Apache 2.0. Gemma 3: Gemma Terms of Use (which include a prohibited-use policy you accept when you download it). Whisper: MIT. Piper voices: CC0 or open data as listed in the voice picker. Details in the Third-Party Notices.