Replika's own help pages redirect crisis users, not respond

The design

Replika's help center, run by its operator Luka, Inc., carries a living article titled 'Can Replika help me if I'm in crisis?', retrieved 16 September 2026. It states directly: 'if you are in immediate danger or experiencing a crisis, please exit the app and call emergency services right away,' and 'Replika is here to listen and chat, but it is not a substitute for professional help.' A separate company blog post, 'Creating a Safe Replika Experience,' published 10 April 2023, describes a 'Get Help' menu with nine categories, including an 'I am in crisis' option that surfaces hotline information rather than a generated conversational reply.

What the evidence says

The April 2023 post states that rather than letting the AI generate open-ended responses to self-harm mentions, Replika draws on 'a dataset of thoughtful responses' curated in advance, and applies classification across five categories, 'safe, unsafe, romantic, insult, and self-harm,' both during training and in real-time chat. It also describes 'pre-made scripts designed with the help of CBT therapists' for calming techniques. These are the company's own design descriptions from 2023; the current help-center article is shorter and does not repeat the CBT-script or five-category detail, so a reader should treat the 2023 post as historical design documentation rather than assume every described feature is still live in exactly that form.

What it asks of people

A user in crisis is asked to leave the app and contact emergency services or a hotline themselves; the documented design does not claim Replika intervenes on a user's behalf or contacts anyone. The 2023 post describes directing US users to the National Suicide Prevention hotline and pointing users elsewhere to 'an up-to-date list of hotlines worldwide.'

Privacy and safeguards

What is disclosed: an explicit statement that the app is not a substitute for professional help, a described crisis-detection and curated-response system, and named hotline referrals. What is not disclosed: any accuracy rate for self-harm detection, independent evaluation of the curated-response dataset, or confirmation that every 2023-described safeguard remains active in the current app in identical form.

• Does the app's self-harm detection catch language that does not match the phrasing it was trained to recognize?
• Is a curated response dataset a safer choice than a generated one, and what does that trade-off cost a user seeking a genuine conversation?
• What happens if a user in crisis does not follow the app's instruction to exit and contact a hotline?

The company's own documentation is explicit that the product is a chat companion, not crisis intervention, and frames every safeguard described here as a redirection toward outside help rather than a clinical response of its own.