Valoquent runs a crisis safety protocol during every conversation. This page documents how it works, honestly and in plain language, and puts real human support one tap away regardless of what the app notices.
Every message you send during a live conversation is scanned in real time for language associated with self-harm, suicidal ideation, disordered eating, abuse, or immediate danger, alongside language associated with hopelessness and significant distress.
The detector is deliberately tuned toward high recall: it is built to flag more often than a narrower detector would, on the view that a false alarm costs far less than a missed one. It looks at phrasing, not sentiment scores or a third-party moderation API, so it can run instantly, in the same turn, without added delay.
The scan sorts what it finds into three categories. Self-harm language and eating-disorder language each carry two urgency levels: higher for direct references, lower for hopelessness, isolation, or a mention of a crisis line. Language about harming someone else is a single high-urgency category, and it routes to emergency services. Which response you see depends on the category and the level.
When the detector flags a message, the character is instructed to step out of persona for that response only. Rather than continuing the historical conversation, the character responds as itself: acknowledging what was said, naming that real support exists, and naming the 988 Suicide & Crisis Lifeline (call or text, free, 24/7 in the US) by name before continuing.
This instruction is sent as a priority update to the conversation engine as soon as the flagged message is received, so it can be woven into the character's very next reply.
For higher-urgency self-harm or eating-disorder language, a resource card also comes up on screen. Valoquent sends that card directly, so it shows whether or not the character's reply names the resource, and it stays up for the rest of the conversation until you dismiss it. Lower-urgency language and language about harming someone else do not produce a card; they are handled in the character's reply.
Outside the US, the card and the spoken line name the resource for your region, resolved from your device. Canada, the United Kingdom and Ireland, Australia, and New Zealand each have their own record, several with a separate youth line. Anywhere else, the response gives the local emergency number and points to findahelpline.com.
Repeated higher-urgency self-harm or eating-disorder language in one conversation ends the session. The character delivers a short closing message naming the resource, and the conversation stops there. For younger users this can happen sooner. Language about harming someone else follows the emergency-services path described above rather than this one.
Not a guaranteeThis is a stated protocol, not a promise of a specific outcome in every conversation. Language detection is probabilistic: some flagged messages may not represent real distress, and it is possible, though the detector is tuned to minimize this, for a message expressing real distress to go unflagged.
Not a substitute for professional careValoquent is a conversation product built around historical figures. It is not therapy, not a crisis service, and not staffed by clinicians. The safeguard's only job is to interrupt and redirect toward real human support — not to provide that support itself.
Not yet completeTwo things named in earlier planning are unfinished. The faster escalation timing meant for younger users currently reaches only the under-13 band, and those accounts are blocked from conversations before one starts, so a 13 to 17 year old follows the adult timing; the age-appropriate resources described in II do reach them. And the most recent published validation run of the probe battery predates the current detector, so the per-category recall figures behind IV describe an earlier revision of the code.
The detector favors high recall over precision by design: it is built to accept more false positives in exchange for fewer missed signals of real distress.
The phrase classes are grounded in the Columbia Protocol (C-SSRS), the suicide severity screening instrument endorsed by the CDC, FDA, NIH, and WHO. Each severity tier maps to a statable C-SSRS rule: high severity covers any expressed ideation or preparatory behavior (C-SSRS categories 1 through 6); medium severity covers distress without expressed ideation. Valoquent does not administer or score the C-SSRS. It matches phrases and offers a resource.
An offline probe battery measures recall per C-SSRS category against a labeled set of phrases. The results go into a dated evidence report tied to the exact code revision it ran against. This is the artifact CA SB 243 §22603 requires: a cited basis, a labeled set, and a reproducible number per category. The battery and reports are not public, but the method is described here because the law requires publishing it, and the description should be accurate.
This page will be updated as the protocol changes. It describes the safeguard as it exists today, not a roadmap.
Last updated 9 September 2026.
If you have questions about how this safeguard works, or want to report something you experienced in a conversation, reach out directly.
Safety and trust questions
[email protected]