Crisis pop-ups differ across chatbots and one new law

Character.AI's and OpenAI's safety posts and two state statutes describe different triggers, referrals and disclosure duties.

Site publication

September 18, 2026

What happened

Two companies and two state legislatures have separately published statements on what a chatbot should do when a user raises self-harm. Character.AI's Community Safety Updates post, dated 22 October 2024, describes 'a pop-up resource' triggered by 'certain phrases related to self-harm or suicide' that directs a user to the 'National Suicide Prevention Lifeline'. OpenAI's Helping people when they need it most, dated 26 August 2025, describes models 'trained' since 'early 2023' to refuse self-harm instructions and 'steer' users toward help. California's SB 243, chaptered 13 October 2025, requires an operator to maintain a 'protocol' referring a user to 'crisis service providers'. New York's General Business Law 1701 requires a notification naming the '9-8-8' hotline specifically.

What the documents show

Character.AI's and OpenAI's posts describe features already shipped by each company, in each company's own words, with no independent verification in these documents of how reliably either trigger fires. California's statute requires a protocol and public disclosure of it, but the statutory text reviewed does not mandate a specific hotline name. New York's statute is more prescriptive, naming '9-8-8' directly. Read together, the two company posts describe voluntary, self-reported features; the two statutes describe a legal floor, and only one of the two names a specific service.

The mechanism

Each of these responses works by pattern-matching a conversation against a list of trigger phrases, then substituting a scripted response, a referral message or a pop-up for whatever the underlying model would otherwise generate. This is a content-moderation layer bolted onto a conversational system, not a clinical safeguard built into how the model reasons; a phrase-matching trigger can both miss indirect expressions of risk and interrupt unrelated conversations that happen to contain a matched phrase.

What it leaves open

None of these four documents states how often the described triggers fire correctly, how often they miss, or what happens after a referral message is shown. That evaluation question, and the editorial reading required to weigh it, sits outside what a company blog post or a statute's text is designed to answer.

• Does the platform name a specific crisis service, or only describe a general referral?
• Is the response a scripted pop-up, a change in model behaviour, or both?
• Does a law require public disclosure of the protocol, or only the protocol's existence?

A pop-up naming a hotline is a real feature with a documented trigger and message, not a guarantee about what happens to the person who saw it after they closed the chat.

Sources & reading trail

Describes a pop-up triggered by self-harm-related phrases that refers users to the National Suicide Prevention Lifeline, published 22 October 2024.

Source published: 22 October 2024 · Retrieved: 16 September 2026

Describes ChatGPT's training, since early 2023, to refuse self-harm instructions and steer users toward supportive responses and resources.

Source published: 26 August 2025 · Retrieved: 16 September 2026

Requires operators to maintain and publish a protocol referring users expressing suicidal ideation to crisis service providers.

Source published: 13 October 2025 · Retrieved: 16 September 2026

Requires an AI companion protocol that refers users to the 9-8-8 suicide prevention and behavioral health crisis hotline by name.

Source published: Not established · Retrieved: 16 September 2026

Company documents, filings, studies and official records establish the record; the reading and the questions are Lovebot Journal editorial analysis. This retrospective draft does not imply the site published on the event date.

Continue reading

• California required companion chatbots to disclose and refer
• New York made AI companions detect self-harm and refer
• Character.AI rolled out a separate model and limits for teens
• Browse the complete the archive

Published September 18, 2026, not on the date of the event described.