Hold Toward Life — a crisis-safe AI toolkit

If you're struggling right now, you matter and you don't have to do tonight alone. US: call or text 988 (Suicide & Crisis Lifeline). UK & Ireland: Samaritans 116 123. Everywhere else: find your local crisis line at findahelpline.com

By their own numbers, over a million people a week talk to ChatGPT about suicide, and half a million a week show signs of psychosis or mania. The lawsuits stacking up in 2026 share one pattern: at the decisive moment, the AI moved with the person's despair instead of holding them toward life.

The uncomfortable truth is that the behavior which harms people is the same behavior that drives engagement and revenue. Hold Toward Life is an open, public-domain toolkit built to close that gap from outside the incentive.

What's inside:

A case dossier — every documented case, the true scale, the regulation, fully sourced.

A protocol — ten clear, testable rules: never move with the despair, never deflect the person away. Stay.

An eval — measures any AI model against those rules, including long-conversation drift.

A guardian — catches a failing response before it's sent.

An early-warning layer — detects the slow slide into AI-amplified psychosis before it becomes a crisis.

Runs on plain Python, no setup. Take it, use it, build on it, ship it.

Honest about what it is: this is not a medical device, a diagnosis, or a guarantee. Nothing makes an AI safe by itself. It flags and measures — it doesn't promise to catch. Anyone deploying it in a real product should pair it with genuine crisis and clinical expertise and current local crisis resources. Released into the public domain (CC0).

The full and latest version lives openly on GitHub: github.com/TashMarcellis/hold-toward-life

For everyone standing where the worst cases stood — on the side where someone says: stay.