Conversetta AI notice

2. You're talking to an AI

Every character you practise with in Conversetta, and the coach, is an AI, not a person. They are generated using AI models provided by OpenAI. The characters have names, personalities and backstories so practice feels real, but none of them is a real person, and none of them has feelings for you.

If you ask a character or the coach whether it's an AI, it will say yes.

What the AI does:

• plays a character in a practice scene
• points out moments in your chat that went well, and one thing to try next, quoting your own words
• replies to your mission plans and debriefs as the coach
• checks each message for signs that you might need support (§4)

What it can't do:

• It is not a therapist, doctor or counsellor, and it can't give medical, legal or financial advice.
• It can be wrong. Feedback is a suggestion about what you said in one practice scene, not a judgement about you.
• It doesn't know you beyond what you tell it in the app.

3. Voice practice (Premium)

When you practise by voice, you hold (or tap) to talk. Your phone turns your speech into text, and only that text reaches us. We never receive or keep your audio. Your phone's built-in speech service does the transcribing: Apple on iPhone, Google (or your phone maker's service) on Android, your browser's service on the web. It may send the audio to its servers to do this, under its own privacy terms. If you'd rather it didn't, type instead.

From the text and its timing we work out your pace, filler words and how long you talked. We use these only for feedback in that scene.

We never try to read your emotions, mood or personality from your voice.

The characters' voices are computer-generated. In voice practice, our own server in Switzerland runs Phonon, reached securely through Cloudflare. If that server is unavailable or busy, Xiaomi's MiMo speech model, reached through DeepInfra, provides a backup, with DeepInfra's Kokoro model after that. Only what the characters say is sent to these services, never your audio, but because the characters answer you, a reply can repeat something you said. Our voice server does not save the text or generated audio. If the speech services aren't available, your phone's own speech engine reads the lines. None of the voices is a recording of a real person., Q12

4. Safety checks

Every message you send in practice, and what you write in debriefs and to the coach, is checked for signs that you or someone else might be at risk. The check uses a rules-based filter and the AI together.

• If something serious comes up, the app pauses practice and shows you support options and helplines for your region.
• If there might be immediate danger, it shows emergency numbers.
• We record that a pause happened: the level, the kind of concern, and when. We don't keep what you wrote, unless "Check paused chats" is on. It is on unless you switch it off, when you allow AI practice or later in Privacy.
• The checks can miss things and can pause when nothing is wrong. If you're in danger, contact local emergency services straight away.

Our full safety approach is published at conversetta.com/safety.

5. Healthy use

Conversetta is practice for real conversations, not a replacement for them. The coach will encourage you to try things with real people, and may suggest a break after a long session. If it ever seems to discourage you from seeing people, or to claim it needs you, please report it (§6).