InnieHub AI Assistant Disclosure

Published by: T&T Consulting Business, LLC ("T&T"), which operates InnieHub
Lives at: /ai-disclosure
Version: 1.0
Last updated: 23 August 2026
Effective: 23 August 2026

Read this part, if you read nothing else

1. You are talking to software. Not a person. Not a therapist, counsellor, psychologist, psychiatrist, doctor or crisis worker. Nobody is reading your messages as you write them.

2. It can be wrong, and it will sound certain either way. AI models produce text that reads as plausible; they do not check whether it is true. Do not use it to decide anything about medication, treatment, a diagnosis, money, the law, or your safety.

3. It is a first step towards people, not a replacement for them. That is what we tell it to be, and it is what we mean it to be. If it is the only thing you talk to, something has gone wrong with how we built it.

4. The safety check never depends on any of this. It runs on every message, on every tier, whether or not you have paid, whether or not you have any allowance left, and whether or not you have accepted this document. Somebody typing that they want to die, on a screen they have not finished reading, gets help and not an error.

Everything below is the detail. You can come back to it whenever you like, this page stays reachable from the assistant.

Contents

1. What it is, technically
2. What we tell it to do, and why that is not a guarantee
3. What is sent to Anthropic
4. What is never sent
5. What it remembers
6. The crisis check
7. Why it never gives you a phone number
8. Your allowance, and what happens when it runs out
9. When the reply is not from the model at all
10. Your conversation, and deleting it
11. What we do not do with any of it
12. Languages
13. Changes to this document
14. Contact

1. What it is, technically

1.1 InnieCare is a conversational assistant built on a general-purpose AI model provided by Anthropic, PBC, a company in the United States. It is not a model we built, trained or fine-tuned, and it is not specific to mental health.

1.2 The model, as at the date of this version, is claude-haiku-4-5. We name it because you are entitled to know what you are talking to. We date it because models change more often than legal documents should: the model in use right now is shown in the assistant's information panel, and that is the authoritative answer if the two ever disagree.

1.3 Replies are capped short by design, a few sentences, not an essay. That is a deliberate limit, not a failure.

1.4 Nothing else in InnieHub uses AI. It does not write posts, moderate content, rank your feed, decide anything about your account, your tier or your role, or read anything you write anywhere else on the platform. The assistant is the only place, and the only thing it sees is what you send it.

2. What we tell it to do, and why that is not a guarantee

2.1 The model is given a written instruction before every conversation. Among other things, that instruction tells it:

• to listen without judgement and offer practical coping techniques, grounding, breathing, plain explanation of what anxiety and panic are;
• to remember that InnieHub is a community, and that connecting somebody to another person is worth more than any technique;
• never to diagnose a medical or psychiatric condition, or to suggest or rule one out;
• never to present itself as a licensed therapist or medical professional;
• never to give medical advice about medication;
• to raise a safety signal if it thinks somebody may be in danger, and not to wait for certainty.

2.2 And here is the honest part. Those are instructions to a model, not locks on a door. A model follows an instruction like that reliably, but reliably is not the same as provably, and there is no mechanism that makes it impossible for the assistant to say something it was told not to say.

2.3 Which is why the rule for you does not depend on the rule for it. Even if it never once broke an instruction, point 2 at the top of this page would still stand: do not rely on it for a decision that matters. We would rather tell you how the safeguard actually works than let its existence do reassuring work it cannot do.

3. What is sent to Anthropic

When you send a message, two things go to Anthropic so a reply can be generated:

3.1 The recent part of your conversation: the last 20 messages, yours and its, plus the one you just sent.

3.2 A catalogue of public things inside InnieHub, groups, events, courses and listed professionals, so that the assistant can point you at real things that exist here rather than inventing them. It is the same catalogue for everybody. It contains nothing about you.

3.3 On a message that trips the safety check, the catalogue is not sent. Nothing competes with the safety reply.

4. What is never sent

Anthropic does not receive:

• your name, your username, your email address or any account identifier;
• your mood history, your exposure plans, your thought reframes, your social battery log, or any other wellbeing record;
• anything you have posted, commented, or written in a group, a forum or a message to another member;
• anything from InnieDate;
• your location, your device, or your IP address;
• and nothing at all from anyone else's conversation.

4.1 What Anthropic gets is a conversation with no name on it and a public catalogue.

5. What it remembers

5.1 It is given the last 20 messages. Not the conversation before that, and nothing from any other day beyond those 20.

5.2 You can scroll back further than it can remember, the screen holds the last 50. That gap is deliberate, and in that direction on purpose: you should be able to re-read more of your own conversation than the assistant is working from, never less.

5.3 It has no memory of you between conversations except those messages. It does not know your name, does not know you are the same person who wrote yesterday except insofar as yesterday is in the last 20, and builds no profile of you of any kind.

6. The crisis check

6.1 Every message you send is checked for signs of suicidal thinking or self-harm. Both in English and in Spanish.

6.2 The first check runs on our own servers, before anything leaves for Anthropic. It looks for defined phrases, it is a fixed list, not a judgement, and it does not use AI. That matters for a reason worth stating plainly: it works even when nothing else does. No allowance, no payment, no acknowledgement of this document, no working connection to Anthropic, the check still runs.

6.3 The model can also raise the signal itself, using a tool it has for exactly that, when a message worries it but matches no phrase on the list. Either signal is enough, and neither waits for the other.

6.4 What happens then. You get a reply written to be steady and direct, and a note pointing you at the Crisis Resources card. And a notification goes to the people on duty, administrators, and moderators who are active and have accepted the Moderator Terms, asking them to check in with you.

6.5 What that notification contains. That a safety signal fired during your InnieCare chat, your name and username, and a link to your profile. It does not contain what you wrote. Not the message, not the reply, not a summary. A moderator is told that you might need somebody, not what you said.

6.6 It is not sent more than once in 15 minutes. A long, hard conversation should not produce a stream of alerts about the same person.

6.7 A crisis message is answered by the model even when nothing else would be. If your allowance is spent, or the platform has hit its monthly ceiling, an ordinary message falls back to the simpler responder in section 9, but a message that trips the safety check goes to the model anyway. It is the one moment in this product where cost is not a consideration. It also means: past your allowance nothing is sent to Anthropic except a message the safety check has marked.

6.8 If you would rather not be checked, the answer is that you cannot use the assistant without it. The check is not a feature of the assistant; it is a condition of our being willing to offer one. Everything else in InnieHub works without it.

7. Why it never gives you a phone number

7.1 The model is instructed never to write a helpline number, a short code or an emergency number, for any country, ever. Instead it tells you to open the Crisis Resources card.

7.2 The reason is not caution, it is accuracy. The model does not know where you are. InnieHub does, and the card shows the line for your own country. A number produced from a model's memory is very likely the wrong country's, and a wrong number handed to somebody in crisis is worse than no number at all.

7.3 The Crisis Resources card is available on every tier, is never counted against any allowance, and is never behind payment. In the iOS application it works with no network connection; on the web the page has to have loaded. It covers five countries, see section 10 of the Health and Medical Disclaimer.

8. Your allowance, and what happens when it runs out

8.1 Model-backed replies are metered:

8.2 They are approximate on purpose. They are calculated from what one exchange costs us to run, so if that cost changes the numbers move with it. The current figures are shown in the app, and any material reduction is treated as a price change under clause 17.10 of the Terms.

8.3 Running out does not turn the assistant off. It falls back to a simpler responder that uses no AI model at all. In that mode nothing whatsoever is sent to Anthropic, not your message, not context, not a classification, with the single exception in 6.7.

8.4 There is also a ceiling on the whole platform, described in clause 13.8 of the Terms. If total AI spending across all members reaches it in a month, everybody falls back to the simpler responder until the month ends, including paid accounts that have used none of their own allowance. We tell you because it is the one circumstance where a paid tier stops delivering what it is sold for, for reasons that have nothing to do with you.

8.5 You can send at most 20 messages a minute. That is a limit on the machinery, not on you.

9. When the reply is not from the model at all

9.1 There is a simpler responder behind the assistant. It uses no AI: it recognises patterns in what you wrote and answers from a written set of replies. It is more limited, and it is still crisis-aware, the check in section 6 runs the same way.

9.2 It answers in three situations:

a) your allowance is spent;
b) the platform ceiling in 8.4 has been reached;
c) the call to Anthropic failed.

9.3 The third one is the one worth telling you about. When Anthropic is unreachable or the request errors, we do not show you an error, you get a reply from the simpler responder instead, so the conversation does not break in the middle of something difficult. That is the right behaviour. But it means there are moments when you believe you are talking to the model and you are not.

9.4 So we will mark it. A reply that did not come from the model carries a small, plain note saying so. Not an apology and not an alarm, just the fact, so that "this answer felt oddly generic" has an explanation available instead of being something you wonder about.

9.5 We already record which is which on every reply we store, so this is a matter of showing you something that exists rather than measuring something new.

10. Your conversation, and deleting it

10.1 Your conversations are stored against your account. They are never shown to another member, are included in the data export you can generate yourself at any time, and are deleted when your account is deleted.

10.2 You can erase the whole conversation whenever you want, from the assistant. It is a real deletion, not a hide. This is mental-health content, and being able to close a subject properly matters more than any use we might have for the rows.

10.3 Three consequences, none of them accidental:

a) The assistant forgets too. Its memory is that same conversation, so your next message starts a genuinely new one.

b) Your allowance does not come back. What you have used this period is recorded before the messages go. We are saying this out loud because otherwise it looks like a bug: deleting and finding the counter unchanged is confusing, and the alternative, deletion restoring your allowance, would make the allowance meaningless.

c) Crisis notifications already sent to moderators stay. They carry no message content (see 6.5), and a safety record that a person can erase is not a safety record.

10.4 Withdrawing your consent to health-data processing stops the assistant working, and is described in section 7 of the Privacy Policy. It stops the processing; it does not by itself delete what you already wrote. Deleting is 10.2, and it is yours to do whenever you like.

11. What we do not do with any of it

11.1 We do not train anything on your conversations, not our own models, and Anthropic does not train theirs on them under our contract with them.

11.2 We do not read them as a matter of course. No member of staff, moderator or administrator sees your InnieCare conversation. There is no review queue for it and no sampling.

11.3 We do not send them to our analytics provider, ever, in any form.

11.4 We do not use them to profile you, to target anything at you, to decide anything about your account, to change what you see anywhere else in InnieHub, or to inform any decision about your membership, your tier or your role.

11.5 When something goes wrong technically, we record the failure with our error-monitoring provider. What is recorded carries no message, no reply, no name and no account identifier, only that a failure happened, on which feature, on which tier, and whether it was a crisis turn. That last flag is there because a crisis conversation degrading is a different severity of problem from an ordinary one, and on its own it identifies nobody.

12. Languages

12.1 The assistant works in English and Spanish. It replies in your language, not the language it was instructed in.

12.2 Those two, and not others, for a safety reason. Crisis detection and the helpline directory exist in English and Spanish. Offering the assistant in a language where we cannot recognise somebody saying they want to die would be offering the part that comforts without the part that catches.

13. Changes to this document

13.1 If we change what the assistant does in a way this document describes, we update this document and ask you to acknowledge it again. A version you never saw is not one you agreed to.

13.2 Changing the underlying model is not by itself a change to this document, provided what is sent, what is kept and what is never sent all stay as described here. The current model is always in the assistant's information panel, per 1.2.

13.3 Previous versions are at the legal centre.

14. Contact

If you are in immediate danger, do not write to us. Contact your local emergency services, or open the Crisis Resources card, it is on every tier, and in the iOS application it works with no connection.