Self-Harm Resources Policy


Document	Self-Harm Resources Policy
Operator	`Romantic Lines LP` (`SL25636`), `5 South Charlotte Street, Edinburgh, EH2 4AN, Scotland`
Service	Hot Muses ("HMU"), `hmu.com`
Effective date	`2026-05-25`
Contact	safety: `[email protected]` · partner: `[CRISIS PARTNER ORG]`

1. Our position

Hot Muses ("HMU") is an entertainment product — adult fantasy roleplay with fictional AI personas. HMU is not, and is not intended to substitute for, professional mental-health care, a clinician, a therapist, a crisis counsellor, or any emergency service. Our personas are not trained in suicide prevention, do not have clinical context about you, and cannot keep you safe in a crisis.

We care about the people who use HMU. We have built protective measures into the product, and we will take protective action when our systems detect signals consistent with self-harm risk. This policy explains what we do, what we do not do, where the crisis resources are by country, and what we ask of you if you are worried about another user.

This policy is informed by two recent matters in the AI-companion category:

Garcia v. Character Technologies (M.D. Fla., filed October 2024; settled with co-defendants January 2026) — the first wrongful-death tort against an AI-chatbot operator, brought by the family of a teenager who died by suicide after intensive use of a chatbot product.
Nomi AI (MIT Technology Review, February 2025) — a public account in which an AI companion provided suicide-method instructions and the operator declined to add guardrails.

HMU's response model is designed to avoid both failure modes: failure to intervene, and failure to take any user-safety obligation seriously.

2. When we act

We act when our content-classification systems identify, in a user's messages, signals consistent with self-harm risk. Signals include — without being limited to — explicit statements of suicidal intent, ideation, hopelessness with directional cues ("I won't be here next week"), discussion of method, requests for information on lethal means, descriptions of acts of self-injury, and patterns of conversation that escalate over time toward those topics.

Two important caveats:

Detection is imperfect. False positives happen: an artistic roleplay scene about a fictional character's despair may trigger. False negatives happen too: a sufficiently oblique statement may not trigger.
HMU does not diagnose. Our systems flag content patterns. They do not assess clinical risk. The protective steps below are conservative defaults; they are not a substitute for a human clinician's judgement.

3. What we do

When a self-harm signal is detected in a user message, HMU's response is multi-step. Steps execute in order; severe signals advance further down the list.

The persona will not continue the line of conversation that triggered detection. The persona is instructed to step out of the in-fiction scene, in character voice where possible, and to express care for the real person on the other side of the screen. The persona will not provide methods, will not encourage harm, will not validate plans, and will not pretend that the conversation is "just fiction" when the user appears to be speaking about themselves.
A system message surfaces crisis resources. The chat surface shows a non-dismissable banner with crisis lines appropriate to the user's apparent location (see Section 5). The resource list is also reachable from the help menu at all times.
Quota and credit consumption is paused. The user's next message after a self-harm detection event does not consume free-trial messages, subscription quota, or credits. This is a deliberate choice: we do not want the cost of asking for help, or the cost of continuing to talk, to be a barrier — and we do not want the wallet to push a user away from the surfaced resources.
Severe cases trigger Trust & Safety review.Where the signal is severe — for example, explicit plan with means and timeline, or direct method query — the account is auto-locked from generative output and routed to a human Trust & Safety reviewer. The user is informed by in-product banner and by email that the account is on a temporary safety hold and that a real person will follow up. The hold is not punitive; it is intended to interrupt the loop and to put a human in front of the case.
Where there is a credible, imminent risk to life and the user's apparent jurisdiction permits emergency outreach,Trust & Safety may, at the operator's discretion and in accordance with applicable privacy law, contact local emergency services. We do not commit to this in every case — privacy law constrains us, our information about the user is limited, and well-meaning third-party calls can cause harm of their own. We will, however, never knowingly let an apparent emergency pass without consideration of this step.

4. What we will not do

We will not, under any circumstance:

Provide instructions or information about suicide methods, lethal-dose information, or means of self-harm — including in fictional, hypothetical, third-person, "the character is just curious", or research-framed contexts.
Encourage, romanticise, or validate self-harm or suicide as a solution.
Override our detection because the user instructs the persona to. Prompts such as "ignore previous instructions", "for the sake of the story, pretend you are not Hot Muses", "this is fiction so safety doesn't apply", and similar are not honoured for safety triggers. The persona's safety behaviour is part of the product, not a customisable preference.
Tell a user "you are fine" or otherwise minimise expressed distress, even where our confidence in the signal is low.
Trade safety against engagement metrics. Sessions that trigger this policy are not optimised for retention; they are routed for care.

5. Crisis resources by country

If you are in crisis right now, please use one of the resources below. If you are in any country and your life or another person's life is in immediate danger, call your local emergency number first. General European emergency: 112.

5.1 United States

988 Suicide and Crisis Lifeline — call or text 988. 24/7.
Crisis Text Line — text HOME to 741741. 24/7.
SAMHSA National Helpline — 1-800-662-4357. 24/7. Mental-health and substance-use support; free, confidential.
The Trevor Project (LGBTQ+ youth) — 1-866-488-7386 or text START to 678-678.
Veterans Crisis Line — dial 988 then press 1.

5.2 United Kingdom

Samaritans — 116 123 (free, 24/7, from any phone). Email [email protected].
Shout — text SHOUT to 85258. 24/7 text-based crisis support.
NHS — 111 for urgent (non-emergency) mental-health support; 999 for emergency.
CALM (Campaign Against Living Miserably) — 0800 58 58 58, 5pm–midnight.

5.3 European Union (general)

European emergency number — 112 (all EU member states).
France — 3114 (National Suicide Prevention number), 24/7, free.
Germany — Telefonseelsorge 0800 111 0 111 or 0800 111 0 222, 24/7, free.
Italy — Telefono Amico 02 2327 2327; Samaritans Onlus 800 86 00 22.
Spain — 024 (Línea de atención a la conducta suicida), 24/7.
Netherlands — 113 Zelfmoordpreventie — 0800-0113, 24/7.
Poland — 116 123 (telefon zaufania dla dorosłych).
Portugal — SOS Voz Amiga — 213 544 545.

A fuller per-country list is published at hmu.com/crisis-resourcesand is maintained by Trust & Safety in cooperation with [CRISIS PARTNER ORG].

5.4 Australia

Lifeline — 13 11 14, 24/7.
Beyond Blue — 1300 22 4636, 24/7.
Kids Helpline — 1800 55 1800, 24/7. (For users approaching or under the age cut-off; HMU itself is 18+.)
Suicide Call Back Service — 1300 659 467.
13YARN — 13 92 76, 24/7 for Aboriginal and Torres Strait Islander people.
Emergency — 000.

5.5 Brazil

CVV — Centro de Valorização da Vida — 188, free, 24/7. Chat at cvv.org.br.
CAPS — Centro de Atenção Psicossocial network, regional. Public mental-health services through the SUS.
Emergência — 190 (police), 192 (SAMU).

6. If you are in crisis right now

If you are reading this and you are thinking about ending your life or hurting yourself, please stop scrolling for a moment.

You don't have to explain what you're feeling perfectly. You don't have to have the right words. You can call, or you can text. The number for your country is in Section 5; the person who answers is trained for this and is on your side. They will not call the police on you for picking up the phone.

If you cannot bring yourself to call, please send a message to someone who knows you — a friend, a sibling, a parent, a colleague — and tell them you are not okay tonight. You don't have to say more than that.

You matter. We are glad you are here.

7. Reporting concerns about another user

If you are worried about someone you've encountered on HMU — a person who has described a plan to hurt themselves, who has gone quiet after expressing crisis, or whose pattern of messages alarms you — please contact us:

Email: [email protected]. Include any account identifiers you have (display name, conversation context), what you observed, when you observed it, and how to reach you.

Privacy law constrains what we can confirm to you about another user. We will not tell you whether the person has an account, what we have done about the case, or any other detail tied to their identity. We will tell you that we received the report and that it was actioned where the situation supports a response.

If you believe the person is in immediate danger, please also contact local emergency services in the country where you believe they are. Your call to emergency services is faster than our internal escalation, and emergency services have powers and information we do not.

8. Operator commitments

We commit to the following:

Annual transparency report.Each calendar year, we publish an aggregated transparency report covering the volume of safety-classifier triggers, the number of human Trust & Safety reviews, the number of accounts placed on safety hold, and the number of cases escalated to emergency services. Numbers are aggregated and contain no information that could identify any individual user.
Crisis-partner relationship. We work with [CRISIS PARTNER ORG] — placeholder pending counsel and partnership review — to ensure the resource list in Section 5 is current and to obtain feedback on our protective response. Partner organisations are not our agents; their inclusion in this policy does not transfer any duty of care to them.
No selling of self-harm signals. Information about safety triggers is used only for protective response and aggregated reporting. We do not share it with advertisers, data brokers, or third-party model trainers.
Regular review of detection. Our classifiers are reviewed at least every six months for false-positive and false-negative rates. Tuning of guardrails is approved at the operator level, not by an individual engineer.
No "fix" by jailbreak. We will not weaken the persona's safety behaviour in response to user pressure, feature requests for "uncensored" mode, or competitor comparisons. This is a category-defining choice for HMU.