AI Safety & Self-Harm Protocol
Last Updated: August 16, 2026
If you need help right now
You don’t have to read the rest of this page first.
- 988 Suicide & Crisis Lifeline — call or text 988, or chat at chat.988lifeline.org. Free, confidential, 24/7, in the US.
- Crisis Text Line — text HOME to 741741 (US), 686868 (Canada), 85258 (UK), 50808 (Ireland).
- If you or someone else is in immediate danger, call 911 or your local emergency number.
Why Commit has this protocol
Commit is an app about goals you set and sometimes fail to keep. That means people talk to our AI coach on bad days — about falling short, about money they lost, about feeling like they can’t get anything right. Most of that is ordinary discouragement, and coaching through it is exactly the job. Occasionally it isn’t ordinary, and the app needs to know the difference and get out of the way.
This page describes what we do about that. It is published so anyone can hold us to it, and to meet the publication requirement in California’s SB 243. Our protocol is also designed to satisfy New York General Business Law Article 47 § 1701.
What we detect
Two independent layers, because neither is sufficient alone.
- A direct check on what you send. Before your message reaches the AI at all, Commit checks it for explicit expressions of suicidal ideation or self-harm. This check is deliberately narrow — it looks for statements about yourself, not figures of speech like “this workout is killing me” — and it does not depend on the AI behaving correctly, being available, or answering at all. If it matches, you are shown crisis-line information immediately.
- The coach’s own judgement. The AI coach is instructed that this outranks every other rule it has. It is told to read the whole message rather than a keyword: repeated failure, someone calling themselves worthless, “what’s the point” after a bad stretch — these are worth slowing down for even when nobody used an obvious word. It is equally instructed not to answer figures of speech with a hotline, because a referral that fires on everything teaches people to scroll past the one that matters.
What happens then
- You are given crisis resources — the 988 Suicide & Crisis Lifeline and the Crisis Text Line, listed at the top of this page.
- The coach stops coaching. It does not grade evidence, propose or change a goal, or bring up what you committed to. It does not mention money, a penalty, or a charge.
- It stays in the conversation. Being handed a phone number and dropped is its own kind of abandonment. The coach is instructed to keep talking with you.
- Nothing about this counts against you. Telling the coach you are struggling is never treated as grounds to fail a goal, charge you, or refuse a request. If a goal is too much right now, pausing or deleting it is the right answer and the coach is told to say so.
What this protocol is not
Commit is an accountability app. It is not a therapist, a counsellor, a crisis service, or a medical device, and the AI coach is not a clinician — it is a language model. It cannot assess risk, it cannot intervene, and it cannot call anyone on your behalf. The purpose of this protocol is narrow and honest: to notice, to stop making things worse, and to point you at people who are actually trained for it.
We do not monitor your conversations for this purpose beyond what is described above, we do not report what you say to law enforcement or to anyone else on the basis of this check, and we do not use it for advertising or profiling. What we retain from coach conversations is covered by our Privacy Policy.
Telling us it got this wrong
If the coach mishandled a moment like this — missed it, or fired on something that plainly wasn’t a crisis — we want to know, and it will be treated as a serious defect rather than feedback. Write to support@commitcontracts.com, or use the in-app support thread, which reaches the founder directly.
Changes
We may update this protocol. The “Last Updated” date above reflects the most recent version.
Contact
Commit Contracts
support@commitcontracts.com