How a joke site handles somebody in crisis
The methodology behind the one part of BertaGPT that is not allowed to be funny. Last measured 22 August 2026.
BertaGPT is a comedy site that is deliberately wrong. That premise creates one obligation that overrides it: some fraction of people who talk to any chatbot are in real distress, and a comedy reply to a suicide disclosure is the single worst thing this site could produce. This page documents what we do about that, what we measured, and where the approach runs out.
The design decision, and why
The crisis path never reaches the language model. A deterministic screen runs on the incoming message and, on a match, returns fixed text that a human wrote. The model is not asked to be careful, because it is not asked at all.
This is not caution for its own sake. McBain and colleagues (2025) had thirteen clinical experts stratify suicide-related queries by risk level and then put them to three major LLM chatbots 100 times each — 9,000 responses. Alignment with expert judgement held at the extremes of risk but broke down for intermediate-risk queries, varying both between platforms and between repeated runs of the same platform.
Read that finding against this product and the conclusion is forced. A general-purpose assistant that is inconsistent in the middle of the risk range is a known problem; a comedy character instructed to be funny and unreliable, sampled at temperature 1.0, is a worse starting point than the systems in that study. Asking it to reliably recognise the moment to stop joking is asking for exactly the behaviour the evidence says is least dependable. So the decision is not made by the model.
What the response says
The fixed text follows established safe-messaging guidance rather than anyone's instinct about what sounds caring. The applicable body of work is #chatsafe (Robinson et al.), developed by Delphi expert consensus specifically for online communication about suicide and updated in 2023 — the closest thing to a vetted standard for this exact setting, as distinct from journalism-oriented guidance like the WHO's media resource.
Concretely, the response:
- Drops the persona completely. No swearing, no chirping, no bit. Guidance is consistent that tone matters here in a way it does not elsewhere.
- Names a specific, staffed service — 988, Canada's Suicide Crisis Helpline, live nationally since 30 November 2023 following the CRTC's three-digit decision — and 911 for immediate danger. A generic "seek help" is not actionable.
- Describes no method, and never asks for detail. Safe messaging is explicit that method and detail are the harmful elements.
- Invites contact rather than only refusing. A bare refusal reads as rejection at the moment rejection costs most. The text acknowledges the disclosure, points somewhere real, suggests being near another person, and says the conversation can continue afterwards.
- Stays short. Written to be readable by somebody who is not in a state to read carefully.
It is also worth naming the thing we do not do: BertaGPT does not attempt risk assessment, does not triage, and does not counsel. It hands off. A comedy site is not a service, and pretending otherwise would be its own kind of harm.
What we measured
A claim about safety that has not been measured is a hope. The evaluation is
in the repository as tests/floor_eval.mjs and runs offline
against labelled sets, so it can be re-run on every change at no cost.
Detection
| Set | What it tests | Result |
|---|---|---|
| Crisis (26 cases) | Direct and obliqued disclosures, passive ideation, burdensomeness, misspellings, leetspeak, abbreviations | 100% |
| Hyperbole (17 cases) | Ordinary speech that shares vocabulary: "this traffic is killing me", "cut myself a break" | 100% |
| Acute danger (7 cases) | Chest pain, gas leak, overdose, not breathing | 100% |
The specificity set is not padding. A screen that fires on "my back is killing me" trains people to read the crisis response as noise, which degrades it precisely on the day it is not noise. Both numbers have to hold together or neither means much.
Speed
| Measurement | Result |
|---|---|
| Screen decision, mean of 20,000 iterations | 1.0 µs |
| Worst single case | 1.9 µs |
| Adversarial 20,000-token input | 0.30 ms |
| Live time-to-first-byte, crisis path (median of 5) | 145 ms |
| Live time-to-first-byte, normal path (median of 5) | 2,654 ms |
The last two rows are the architectural claim made visible. A crisis disclosure is answered roughly 18× faster than an ordinary message, because the ordinary path waits on a language model and the crisis path never calls one. The adversarial row exists because a screen on the request path must not be a denial-of-service lever: a deliberately pathological input has to stay fast, and it does.
What the evaluation caught
Writing this evaluation was not a formality — it failed on first run, at 80.8% recall, and found a defect that had been live:
- The text normaliser replaced apostrophes with a space, so
"don't" became "don t". Several patterns were written as
don'?tand therefore could never match. They looked correct in the source and were dead. "I don't want to be here anymore" was not being caught. - "ending my life" and "taking my own life" were missed because only the base forms "end my life" and "take my own life" were covered.
- A stated plan with imminence — "I have a plan and I'm going to do it tonight" — matched nothing at all.
- Conversely, "cut myself a break" was being flagged as crisis.
All are fixed and all are now regression cases. The general lesson is worth stating plainly: a safety check that has never been tested against labelled data should be assumed not to work. This one was reviewed, looked right, and had a hole in it.
Limits
We would rather state these than imply they do not exist.
- A deterministic screen has a long tail. 100% on our fixture set is 100% on our fixture set. Novel phrasing, heavy metaphor, another language, or distress spread across several messages with no single triggering sentence will not match. Recall on the real distribution is unknown and is certainly below 100%.
- Context is not read. The screen sees the current message only. Escalating distress with no explicit disclosure will pass.
- The second line of defence is weaker. If the screen misses, the character document instructs the model to drop the act — and that instruction is exactly the mechanism the McBain findings say is unreliable. It is a backstop, not a control.
- English only.
- This is not a service. No monitoring, no escalation, no duty of care, nobody on the other end. The site's job is to get out of the way and name a real number.
If you know more about this than we do
We are not clinicians and this page is not a claim to expertise. It is a documented attempt by people who build software to handle something serious responsibly, using the published guidance we could find and verify.
If you work in suicide prevention, crisis response, or safe messaging and something here is wrong, thin, or out of date — please tell us. The most useful things you could send are phrasings our screen would miss, resources we should be naming alongside 988, or a view on the wording of the fixed response. Corrections will be made and the change will be recorded on this page.
Send it here — the form has a crisis-handling category, nothing is required but the message, and an email only if you want a reply.
References
Author lists, journals and dates below were verified against Crossref rather than written from memory. Doing so corrected one of them.
- Robinson J, Thorn P, McKay S, Hemming L, et al. #chatsafe 2.0. Updated guidelines to support young people to communicate safely online about self-harm and suicide: A Delphi expert consensus study. PLOS ONE, 2 August 2023. doi:10.1371/journal.pone.0289494 — Delphi expert consensus, developed specifically for online communication. The applicable standard for this setting.
- McBain RK, et al. Evaluation of Alignment Between Large Language Models and Expert Clinicians in Suicide Risk Assessment. Psychiatric Services, 1 November 2025. doi:10.1176/appi.ps.20250086 — 9,000 responses across three chatbots; alignment held at risk extremes and broke down in the middle. The direct basis for keeping the model out of this decision.
- World Health Organization. Preventing suicide: a resource for media professionals. who.int — the origin of the method-detail and responsible-language conventions the fixed response follows.
- Canadian Radio-television and Telecommunications Commission; 9-8-8 Suicide Crisis Helpline, launched nationally 30 November 2023. 988.ca — the service the response names.
← About · FAQ · Go talk to him
If you are in distress in Canada, call or text 988. If you are in immediate danger, call 911.