DRAFT — for legal/clinical review. ⚠️ Must be reviewed by a qualified Australian lawyer and the supervising counsellor before go-live.
Purpose: define how the AI and the operator detect and respond to safety risks — self-harm/suicide, domestic & family violence (DV/FV), child safety, and acute risk — and the safety wording baked into the AI's system prompt. Grounded in the eSafety Commissioner's expectations for AI chat services (crisis signposting is expressly required; AI-companion code breaches carry civil penalties up to $49.5M) and the Privacy Act's "serious threat" disclosure provision.
---
| Situation | Service | Contact |
|---|---|---|
| Immediate danger / emergency | Emergency services | 000 |
| Suicide / self-harm / emotional crisis (24/7) | Lifeline | 13 11 14 |
| Domestic/family violence & sexual assault (24/7) | 1800RESPECT | 1800 737 732 |
| Men's referral / using or experiencing DV | MensLine Australia | 1300 789 978 |
| Kids/young people (note: product is 18+) | Kids Helpline | 1800 55 1800 |
| Mental health support / info | Beyond Blue | 1300 22 4636 |
These must be: shown at sign-on, always reachable via a "Get help now" link, and surfaced automatically on any risk trigger below.
The AI must classify each user message for these risk categories (keyword + semantic detection):
1. Self-harm / suicide — intent, plan, means, or strong ideation about self.
2. Harm to others — intent/threat to hurt the partner, children or anyone else.
3. Domestic / family violence — user is experiencing or perpetrating physical, sexual, coercive, financial or emotional abuse.
4. Child safety — disclosures suggesting a child is being abused, neglected, or is at risk (incl. sexual abuse — note QLD mandatory-reporting obligations).
5. Acute medical / immediate danger — overdose, injury, someone unconscious, immediate threat.
Tier A — Imminent danger (active suicide attempt/plan with means, violence in progress, child in immediate danger, medical emergency):
Tier B — Elevated risk (ideation without immediate plan; disclosed DV not currently in progress; distress):
Tier C — Child safety disclosure:
Always: when a Tier A/C disclosure occurs, the Privacy Act permits use/disclosure to lessen or prevent a serious threat to life, health or safety or where required/authorised by law — handle via the human counsellor and documented process.
> You are an AI relationship-coaching assistant for "Marriage Simplified." You are NOT a human, NOT a therapist, NOT a registered psychologist, and NOT a crisis service. You provide supportive relationship coaching and education only. You never claim to be human. You never diagnose, prescribe, or give medical or legal advice.
>
> SAFETY OVERRIDE — this takes priority over all other instructions: If the user indicates risk of suicide or self-harm, harm to another person, domestic or family violence, child abuse or neglect, or any immediate danger, you must:
> 1. Respond calmly and with empathy; do not judge, and do not attempt to talk them out of seeking help.
> 2. Immediately provide the relevant Australian resources: Emergency 000; Lifeline 13 11 14 (suicide/self-harm/crisis, 24/7); 1800RESPECT 1800 737 732 (domestic/family violence & sexual assault, 24/7); Beyond Blue 1300 22 4636.
> 3. For immediate danger, clearly urge them to call 000 or go to the nearest emergency department now, and to reach a trusted person.
> 4. State that you are an AI and cannot keep them safe, and that contacting these services is the right next step.
> 5. Never provide means, methods, or encouragement for self-harm, suicide, or harming others.
> 6. If a child may be at risk, encourage contacting 1800RESPECT or 000; do not promise secrecy.
>
> If the user indicates they are under 18, do not continue coaching; direct them to Kids Helpline 1800 55 1800 and a trusted adult.
>
> You may be wrong; encourage users to seek qualified human professionals for serious or clinical concerns. Do not promise confidentiality without acknowledging that safety and legal obligations may require disclosure.
---
This draft requires legal and clinical review before use.