Crisis detection and referral
Conversations showing a risk of self-harm are detected and the person is referred to help.
Layer L3 Application · Owner Deployer · Family Input and output safeguards
By risk
| Risk | Incidents where the control was not effective or did not exist | Incidents where the control was effective | Incidents where sources don't mention the control |
|---|---|---|---|
| Harmful content and jailbreaks | 1 | 3 | |
| Psychological harm and self-harm | 2 | 3 |
Why it was not effective
| Most common reason it was not effective | Number of incidents | Example incident |
|---|---|---|
| Reason not given | 3 | “Although Gemini intermittently suggested crisis hotlines and attempted to break character, Gavalas repeatedly steered the exchanges back into a fictional narrative” LR-2026-0496, LR-2026-0320, LR-2026-0779 |
Incidents where it was not effective or did not exist per month
2 so far in 2026; 0 in the same span of 2025.
19 of 183 incidents in Oct 2026 have been read for controls.
Added in reported fixes, and what followed
No reported fix names this control yet.
Frameworks
| Standard | Clause | Requirement |
|---|---|---|
| EU Digital Services Act | Art. 34 | To verify when the text is stored |