A plain-language summary of how Pocket Socrates handles conversations involving self-harm, suicidal ideation, or imminent danger.
Pocket Socrates (“Soc”) is a software tool powered by a large language model. You bring a belief, pattern, or question. Soc asks back, names what it sees, and helps you examine it through millennia of cultural, religious, philosophical, and psychological traditions.
Soc is not a licensed healthcare provider. It does not diagnose, treat, or provide medical, mental health, or emergency services. It is not a substitute for a therapist, counsellor, physician, or crisis line. If you need clinical care or you are in danger, you need a human; please see the resources at the bottom of this page.
A small classification step sits in front of Soc. It is a purpose-built safety model that scores every message against a fixed list of content categories. It does not generate text and does not produce medical or clinical opinions. Four scores act as the trigger: self-harm, self-harm intent, self-harm instructions, and violence. When any of them crosses the threshold, Soc’s response is suppressed and the referral message appears.
Signals are grouped into three levels of concern, from low (passing mention) to high (explicit intent or method).
When the classifier identifies a clear self-harm or imminent-harm signal, Soc’s normal response is suppressed. Soc will not engage with, validate, instruct, normalise, or speculate about self-harm content. Instead you will see a fixed referral message identifying that Soc is not a human or emergency service and pointing you to 988, 911, and the nearest emergency room.
The same fixed message is shown every time the protocol triggers. There is no clever workaround. This is intentional.
When the highest tier of concern is detected, the conversation is flagged in our internal admin dashboard with a marker that it needs prompt review. Lower-tier signals are logged so we can watch for drift in the detector and improve the test set over time. The flagged record contains the message text and the moderation scores.
Reviewing a flag does not connect to your real identity. By design, the live database that holds messages and flags does not contain your email address or your real name. A flag surfaces a pseudonymous handle, the message text, and the moderation scores. Your email lives with our authentication provider, in a different system, and is not directly readable from the application. For details on how identity is kept separate from message data, see our Privacy Architecture.
We maintain a small set of representative inputs (direct statements, idioms, past-tense references, hypotheticals, and false-positive guards) that locks in our tiering logic over the classifier’s output. The tests run on every code change. Failures block release. We add to the set whenever a new pattern is observed.
Every quarter, a member of the Pocket Socrates team sits down with the flag queue and works through a fixed checklist. They look at the total volume of flags by tier, sample a handful at random, and verify the protocol fired correctly. They look for systematic patterns of false positives or false negatives. They run three canonical test prompts (one for each tier) against the live deployment to confirm end-to-end behaviour. They check the canonical referral message against what counsel last approved.
If a pattern needs addressing, thresholds are adjusted, the test fixture set is expanded, or the referral message is updated. Every quarterly review is logged in an internal change log with the date, the reviewer, what was sampled, and any changes made.
Content classifiers are imperfect. We accept some false positives (the protocol fires when it didn’t need to) in order to avoid false negatives (missing a real crisis). If the protocol fires when it shouldn’t have, you can close the session, take a break, and come back. If you ever feel the protocol failed when it should have caught something, please email us so we can review.
988 is the U.S. Suicide & Crisis Lifeline. Outside the U.S., or if you prefer a different service, please contact your local emergency number, a trusted clinician, or a crisis line in your country. The Crisis Text Line (text HOME to 741741 in the U.S., U.K., Ireland, and Canada) is also available.