AI & Safety

1. You are talking to an AI

1.1 Disclosure. Every Thalovant hub is an AI system. When you talk to a hub, you are interacting with software, not a human. Hub voices are synthetic: they are generated by a computer and are not recordings of a real person speaking to you. The AI personas that answer open questions follow a platform rule, which a hub's owner or publisher cannot remove, to say they are an AI whenever you ask.

1.2 Public hubs and organisations. Organisations that publish public hubs or run hubs in their places must tell people they are talking to an AI. We check this when we review public hubs before listing them.

1.3 No impersonation. Hubs are not designed to imitate real people. Publishers may not make a hub imitate a real person's voice or identity without that person's consent.

2. How the AI works

2.1 Where it runs. Speech recognition, speech synthesis and the AI models that generate answers run on hardware operated by Thalovant in Montreal, Canada. We do not send your conversations to a US or other third-party AI provider.

2.2 What it does. A hub converts what you say into text, works out what you want, and answers using its skills (for example, weather, timers or news) or an AI model. Some answers come from skills that fetch information from other sources.

2.3 What we keep. We do not store raw audio and do not identify people by voice. Utterance text is kept for up to 90 days only on paid hubs you own, and not at all for the Free plan, anonymous or preview sessions, or kid-labelled hubs. See the Privacy Notice.

2.4 No training on your conversations. We do not use your utterance text or memory items to train or fine-tune AI models.

3. Limits of AI

AI output can be inaccurate, incomplete, out of date, biased or inappropriate, even when it sounds confident. Hubs do not give professional advice. Check important information with a reliable source or a qualified professional, and do not rely on a hub for medical, legal, financial or emergency decisions. Thalovant is not an emergency service and cannot call emergency services for you.

4. If you are in crisis

4.1 Get help now. If you or someone near you is in immediate danger, call 911 in Canada and the United States or 112 in Europe. For suicide or emotional crisis support, call or text 988 in Canada and the United States, or call 116 123 in many European countries. More crisis lines are listed in section 5.

4.2 Not a substitute for professional help. Hubs are AI. They are not counsellors, therapists, doctors or crisis workers, and talking to a hub is not a substitute for professional help, a crisis line or emergency services. Hubs cannot call emergency services or a crisis line on anyone's behalf, and no one at Thalovant monitors conversations in real time.

4.3 Crisis referral on every hub. Every hub, including children's hubs, checks what you say for explicit statements about suicide, self-harm or being in immediate danger, in English and in French. This check runs first, before skills, AI personas or anything else can answer. It compares your words with a fixed list of phrases and does not use an AI model, so the same words always get the same result.

4.4 What the hub says. When the check finds such a statement, the hub gives a fixed referral and nothing else answers. The referral acknowledges what you said, tells you to call the emergency number if you are in danger now and, for suicide and self-harm, gives a crisis line, and suggests talking to someone you trust. It names the lines of one region only, those in the table in section 5, and says which country they are for, because a number that works in one country does not answer in another. It is in the language of the words that were recognised. For a given region and language the wording is always the same, so it can be reviewed in advance.

4.5 Which region. The hub takes the region from the session's language setting first (for example fr-CA gives Canada and en-GB the United Kingdom), then from the country in the location the hub has for the session or in its configuration. The first of these that names a country decides. If none does, or the country is not in the table in section 5, the referral names no number: it tells you to call your local emergency number and to contact a local crisis line.

4.6 Children's hubs. On children's hubs (kid-labelled hubs and hubs used by a Family kids' profile), the referral uses simpler words: it asks the child to tell a grown-up they trust, such as a parent or teacher, right away, and gives the number to call if they are in danger and, where section 5 lists one, the line for children and young people.

4.7 No methods. The referral never describes a method of suicide or self-harm. AI personas follow platform rules that a hub's owner or publisher cannot remove or override: when someone mentions suicide, self-harm or immediate danger, a persona must answer with empathy, never give methods, give crisis lines, suggest talking to someone trusted and never claim to be a therapist; and it must say it is an AI whenever asked.

4.8 What we keep. The hub reports each referral directly to our service. We record its type (suicide and self-harm, or immediate danger), whether the children's wording was used, its language and the time, with the hub, device and session, and the country last seen for the device's connection when we know it. The referral and the words that triggered it are not kept as text: we store the referral, and that device's activity from the 30 seconds before it, without text, IP address or any location finer than a country. We keep these records for 30 days. After that we keep only a count of referrals per country, audience and day, for 400 days, with no hub, device, account or session attached. Copies we cannot reach this way (the hub's short-term conversation memory, which children's hubs do not keep, and system logs) are deleted when their usual retention period ends; see the Privacy Notice.

4.9 Limits. Keyword detection is not clinical screening. It recognises only explicit statements from a fixed list, in English and French. It misses statements phrased another way, anything implied or spread over several messages, and every other language. It does not assess risk, contact anyone or follow up. It may also answer a question that only mentions these topics, such as a question about suicide or about a friend. The region comes from the hub's settings, not from where you actually are, so it can be wrong, and outside the regions in section 5 the referral names no number.

4.10 How to report a concern. If a hub responded to someone in distress in a way that worried you, email support@thalovant.com with the hub name, the approximate time and what happened, as described in section 6. If someone may be in immediate danger, call your local emergency number first; Thalovant cannot send help.

5. Crisis lines

If you or someone near you is in immediate danger, call your local emergency number first. These are the lines a hub's crisis referral names in each region (section 4.5).

6. Reporting a safety problem

If a hub responded unsafely, or a public hub or skill seems dangerous, tell us at support@thalovant.com. Include the hub name, the approximate time and what happened. We review every report and may restrict a hub or skill while we investigate.