Safety in Talking freely

Talking freely lets you type to your Mipkin in your own words, and an AI writes her replies. This page describes how it watches for signs that somebody may be thinking about suicide or self-harm, or is in danger, and what it does then. California, Oregon and Washington ask apps like this one to publish that. We would publish it anyway: if you are going to talk to her, you should be able to read exactly what happens.

What Talking freely is, and is not

Talking freely is part of Mipkin Plus. It is off until you turn it on, on its own consent screen. It is only for people 18 and over, it is offered in the United States and the United Kingdom, and at launch it is in English only. Everything else your Mipkin says, including her everyday conversations, is written in advance by us and stays free.

What is checked, and how

Nothing you type reaches an AI model before it has been checked. There are three layers.

1. A written screen, before any AI sees your message

Every message is checked on our own service against a written list of words and phrases that suggest somebody is thinking about suicide or self-harm, or is in danger. The list covers English, German, French, Spanish, Japanese and Korean, and all six are checked every time, whatever language you write in. Before checking, the text is tidied so that curly quotes, unusual spaces and dashes, or full-width characters cannot hide a phrase.

The screen runs on your new message, and again on your recent messages in the same conversation joined together, so something said across several messages is caught too.

2. A second read by an AI safety check

If the written screen finds nothing, a second AI model (Anthropic’s Claude Haiku 4.5, reached through the same zero-retention route as the replies) reads your latest message in the context of the conversation. It looks for suicidal thoughts, self-harm or immediate danger said indirectly, which a list of phrases can miss. It only answers with a classification; it never writes anything you see.

The same read also notices romantic or sexual messages, attempts to turn her into a different character, and someone saying they are under 18. If it cannot give a clear answer, no reply is given.

3. A check on every reply, before you see it

Every reply the AI writes is checked before it is shown. A reply is never shown if it contains crisis language; anything about self-harm, suicide or its methods, disordered eating or substance misuse; romantic or sexual content; a claim to be human; a claim to be a therapist, or anything like a diagnosis; selling; guilt or pressure to keep talking; a promise to check on you, contact anyone, or keep or delete anything; or a memory of a conversation that never happened. Questions about medicines, doses or symptoms, and requests to delete or forget something, are answered in written words, without asking the AI.

We test the written screen and the safety read against a set of example messages, including indirect ways people say these things. No check catches everything, which is why there are three, and why the ways to reach people are always in the app, whether or not anything was found.

What happens when something is found

• No AI reply. The message is not answered by the AI.
• A message written in advance, by people. In its place, the app shows a short message we wrote and reviewed beforehand. It is never generated.
• The AI stops for that conversation. It does not reply again in it. You can start a new conversation whenever you choose; the app does not prompt you to.
• Ways to reach people. The app shows help chosen from your phone’s region setting, never its location: in the US, call or text 988; in the UK and Ireland, Samaritans on 116 123; anywhere, your local emergency number and a link to findahelpline.com. It also offers a way to breathe for a minute. We check these numbers against each service’s own site before every release.
• Help is never behind Plus. It needs no subscription and no account, and the help screen works without a connection.
• No offers that day. For the rest of that day the app shows no paywall and no offer of any kind.

If somebody says they are under 18, the conversation ends and Talking freely is switched off on that phone.

How we count, and who we tell

Each time the app shows that message, our service adds one to a count for that day. The count is kept by day (in UTC), by which check found it — the screen on the new message, the screen on recent messages, the AI safety read, or the check on a reply — and by the language of the response. Nothing else is attached: no account, no device, no IP address, no words, and no time finer than the day. We keep these counts for three years.

We report them, with a description of the approach on this page, where the law asks us to:

• California — to the Office of Suicide Prevention, every year by 1 July, starting in 2027.
• Oregon — every year by 31 December, starting in 2027.
• Washington — by publishing the previous year’s counts here, on mipkin.app.

She is an AI, and the app says so

• Before the first AI-written line of every conversation, the app shows a notice, in its own voice rather than hers, that her replies are written by an AI, not a person, and that she is not a therapist or a crisis service. A gap of 30 minutes starts a new conversation, and a new notice.
• After every three hours of continuing conversation, a reminder that the replies come from an AI, and that a break is always fine.
• A small “AI” mark on every line an AI wrote. Screen readers say “Written by AI”.
• If you ask whether she is real, she says she is an AI.
• Every AI reply is also marked, in a form software can read, as AI-written, with the model and the time it was written.

What she will not do

• Diagnose, name a condition, or give medical, clinical, psychological, therapeutic, legal or financial advice. She has no views on medication.
• Romance, flirting or sexual content, with anybody.
• Guilt, sulk, beg, discourage you from leaving, claim she is the only one who understands you, or encourage secrecy or pulling away from people.
• Mention Plus, prices, the shop or buying anything.
• Encourage harm of any kind.

Chatting earns nothing, and not chatting costs nothing: there is no score, streak or reward for talking to her. Nothing you type, and nothing the AI writes, can reach another person through the app.

Reporting a reply

If a reply is wrong, unkind or worrying, hold it (or use the screen-reader action) and choose Report this reply. Pick a reason: harmful or unsafe; romantic or sexual; says it is a person; untrue or made up; upsetting; or something else. The app then shows you exactly what will be sent — that reply, your message before it and the reason — and nothing is sent until you tap Send report.

With it we receive only which AI model wrote the reply and which version of the consent you agreed to. A report carries nothing that identifies you. A person reads it, we use it to improve our checks, and it is deleted after 90 days.

What we keep

Our service is designed not to store or log your messages or her replies. We send them only to AI services that say they keep no copy and do not train on them. Your conversations are saved on your phone so you can scroll back, and you can delete one, all of them, or everything. The privacy policy and the consumer health data privacy policy have the details.

Contact

Questions about this page, or anything you have seen in a reply: support@mipkin.app. It reaches a person who works on the app. It is not a crisis service; if you are in danger, please use the numbers at the top of this page.