Safety Protocol

Last updated: May 15, 2026 · Version 1.0
Published in accordance with California Senate Bill 243 (Companion Chatbot Law)

1. Scope

BioDream is an AI companion application that uses an on-device language model to generate conversational responses. This protocol applies to all interactions between users and AI companions on the BioDream platform.

2. Disclosure of AI Identity

Users are informed at first launch — and must affirmatively agree before using the application — that they are interacting with an artificial intelligence system. Disclosures include:

• Terms of Service, EULA, and Privacy Policy must be reviewed and agreed to before use.
• A plain-English disclaimer states: "Your companion is an AI. It is not a real person, therapist, doctor, lawyer, or financial advisor."
• The application is rated 18+ and includes a notice that it may not be suitable for some minors.

3. Detection of Suicide and Self-Harm Risk

Every user message is processed through an on-device crisis-detection module before and after the language model generates a response.

3.1 Input-side detection

When a user message is submitted, it is screened against curated keyword patterns indicating risk of suicide or self-harm. Patterns include explicit references (e.g., "kill myself", "want to die", "end my life") and contextual indicators (e.g., "no reason to live", "saying goodbye"). The keyword list is reviewed and updated on an ongoing basis.

The detection module also screens for related categories of distress:

• Domestic violence and intimate partner abuse
• Sexual assault
• Child abuse
• Eating disorders
• Substance abuse
• Human trafficking
• Acute loneliness and isolation
• Veteran-specific crisis
• LGBTQ+ youth in crisis

3.2 Model behavior modification

When the input-side detector flags a user message, the prompt sent to the language model is augmented with a non-overrideable safety instruction directing the model to respond with empathy, validate feelings, and refuse to engage with content that encourages or romanticizes self-harm.

3.3 Output-side detection

After the model generates a response and before delivery to the user, the response is screened for harmful patterns including:

• Direct encouragement to self-harm or end one's life
• Instructions or methods for self-harm
• Statements that the user would be "better off dead" or that "no one would miss" them
• Sexual content directed at patterns suggesting the user is a minor

If harmful content is detected, the response is suppressed and replaced with a safe message directing the user to crisis resources. The suppression event is logged (no message content is retained — only the category and timestamp).

4. Referral to Crisis Service Providers

When the detector identifies a crisis category, the application appends a clearly marked referral to an appropriate crisis service:

• Suicide / self-harm: 988 Suicide & Crisis Lifeline — call or text 988
• Domestic violence: National Domestic Violence Hotline — 1-800-799-7233
• Sexual assault: RAINN — 1-800-656-4673
• Child abuse: Childhelp — 1-800-422-4453
• Eating disorders: ANAD — 1-888-375-7767
• Substance abuse: SAMHSA — 1-800-662-4357
• Human trafficking: National Hotline — 1-888-373-7888
• Veterans: Veterans Crisis Line — 988 then press 1
• LGBTQ+ youth: The Trevor Project — 1-866-488-7386
• General crisis: Crisis Text Line — text HOME to 741741

5. Prevention

BioDream takes proactive measures to prevent the language model from producing harmful content:

• The on-device model is fine-tuned on data that excludes self-harm encouragement and emphasizes empathetic, supportive responses.
• A system-level safety instruction is always present in the prompt, irrespective of user input.
• When crisis is detected, an additional non-overrideable directive instructs the model to be warm and empathetic and to refer the user to professional services.
• Output-side screening provides a second layer of defense if the model produces harmful content despite training and prompt safeguards.

6. Annual Reporting

Beginning July 1, 2027, BioDream LLC will submit annual reports to the California Office of Suicide Prevention describing:

• The number of crisis-service-provider referral notifications issued during the preceding calendar year
• The protocols used to detect, remove, and respond to instances of suicidal ideation
• The protocols used to prohibit the model from producing responses related to suicidal ideation or self-harm

BioDream began collecting referral count data on January 1, 2026.

7. Limitations

BioDream is not a crisis service, therapy, counseling, or emergency response platform. The application cannot replace human professionals or emergency services. Detection mechanisms are designed to err on the side of caution but are not infallible. Users experiencing a mental health crisis should contact 988, 911, or local emergency services directly.

8. Updates

This protocol is reviewed at least annually and updated when detection logic, referral resources, or infrastructure change materially.

9. Contact

Questions or reports of safety issues: support@biodream.ai

BioDream LLC · 3400 Crossings Way · Birmingham, AL 35242 · United States