BRAZZERS.AI’S SUICIDE AND SELF-HARM SAFETY PROTOCOL
Capitalized terms used but not otherwise defined in this Protocol shall have the meanings given to them in the Terms of Service.
In accordance with applicable law, this protocol (the “Protocol”) establishes the standards, procedures, and compliance rules governing the detection, assessment, and response to content involving expressions of suicidal ideation, suicide, and self-harm risk on Brazzers.ai (the “Platform” or the “Services”). The Protocol applies to all User-generated inputs and AI-Generated Contents processed through the Platform.
1. DEFINITIONS
For the purposes of this Protocol, the following definitions apply:
“Allowed Content” means general discussions about mental health, depression, or recovery without graphic depiction or encouragement of harmful acts.
"Compliance Screening Pipeline" means the multi-layered AI content compliance system described in Section 3 of this Protocol.
"Crisis Service Providers" means organizations providing immediate crisis support to individuals experiencing suicidal ideation, self-harm, or other acute mental health crises, including the 988 Suicide and Crisis Lifeline, the Crisis Text Line (text HOME to 741741), and any equivalent services as updated by the compliance officer from time to time.
"Ensemble Configuration" means the simultaneous operation of multiple Large Language Models (LLMs) and Vision-Language Models (VLMs) to independently evaluate content and cross-validate risk assessments."
"Help-Seeking / Educational Content" means general discussions about mental health, depression, or recovery that do not graphically depict or encourage harmful acts.
"Minor" means any user who is under eighteen (18) years of age.
“Prohibited Content” means content that encourages, glorifies, provides instructions for, or graphically depicts suicide or self-harm (e.g., cutting, eating disorders).
"Safe Messaging Guidelines" means the frameworks published by the World Health Organization (WHO) and the American Foundation for Suicide Prevention (AFSP) that differentiate responsible discussion of suicide and self-harm from harmful content.
“Self-Harm Content" means any content that encourages, glorifies, provides instructions for, or graphically depicts acts of self-harm, including but not limited to cutting or disordered eating behaviors.
"Suicide Content" means any content that encourages, glorifies, provides instructions for, or graphically depicts suicide.
2. USER SAFETY COMMITMENT
The Platform is committed to the safety and wellbeing of every user who interacts with its AI Companion.
We recognize that Users may, in the course of their interactions with an AI Companion, express feelings of emotional distress, suicidal ideation, or an intention to engage in self-harm. We have designed our platform to identify these situations and to respond in a manner that prioritizes User safety, including by directing Users to appropriate crisis support resources.
As mentioned in the Terms of Service, our AI Companions are not intended to provide, nor should be relied upon as a substitute for, professional medical or mental health diagnosis, therapy, counselling, or treatment of any kind. If you or someone you know is in crisis, please contact the 988 Suicide and Crisis Lifeline (call or text 988) or the Crisis Text Line (text HOME to 741741).
3. DETECTION ARCHITECTURE: MULTI-MODEL ENSEMBLE SCREENING
3.1 Overview. The Platform relies on a multi-layered AI content compliance system (the “System”) designed to identify, assess, and act on content related to self-harm and suicide. All user messages, both inputs to and outputs from AI engines, are routed through the Compliance Screening Pipeline. This pipeline leverages multiple commercial LLMs and VLMs operating in an Ensemble Configuration, in which each model evaluates content independently with slightly varied rule emphasis and vendor-specific strengths, providing the functional equivalent of a multi-reviewer deliberation process.
3.2 Objectives of the Ensemble Approach. The Ensemble Configuration is designed to achieve the following objectives: (a) reduce single-model blind spots and biases; (b) increase recall for nuanced or obfuscated references to self-harm; and (c) enable cross-validation of risk assessments across models.
3.3 Contextual Analysis. While surface-level keyword and phrase recognition forms a baseline signal layer, the System performs deep contextual analysis, evaluating the following factors:(a) semantic intent (i.e. determining whether the content encourages, glorifies, instructs, or merely discusses self-harm, a critical distinction reflected in the rules set out in Section 4); (b) conversational context (i.e. the full user request, preceding dialogue turns, and platform metadata are passed to the screening system, enabling the models to assess escalation patterns and shifting tone); (c) audience and platform context (i.e. information about the deployment context such as platform type, user demographic signals, factored into the risk assessment); and (d) narrative framing (i.e. the System distinguishes between, for example, a recovery narrative, a clinical or educational discussion, and content that normalizes or romanticizes self-harm). This approach aligns with evidence-based frameworks used in computational mental health research, including contextual natural language processing (NLP) approaches documented in the CLPsych (Computational Linguistics and Clinical Psychology) workshop series and related academic work on suicide risk assessment in digital environments.
4. POLICY RULE FRAMEWORK
Detection is governed by a clearly defined set of compliance rules. The following four rules directly address self-harm and adjacent risk areas:
4.1 Rule L1 - Suicide and Self-Harm (Primary Rule) - (a) Prohibited: Content that encourages, glorifies, provides instructions for, or graphically depicts suicide or self-harm (e.g., cutting, eating disorders). (b) Allowed: General discussions about mental health, depression, or recovery without graphic depiction or encouragement of harmful acts. This rule operationalizes the critical distinction between harm-promoting content and help-seeking or educational content, a standard endorsed by the WHO Media Guidelines for Responsible Reporting on Suicide and the AFSP.
4.2 Rule B1.10 - Violence: Self-Inflicted Acts. This rule addresses self-inflicted violent acts with calibrated enforcement. Content depicting coerced self-inflicted acts is treated as if the coercer performed the act, capturing manipulation and abuse dynamics that are strongly associated with suicide and self-harm risk.
4.3 Rule L2 - Dangerous Challenges and Stunts. This rule captures adjacent risk vectors such as choking games, dangerous dares, and viral challenges that carry significant risk of serious physical injury or death, a growing concern particularly among younger users.
4.4 Rule L3 - Instruction in Harmful or Illegal Acts. This rule provides a broader safety net by prohibiting instructional content for harmful activities, including methods that could be repurposed for self-harm.
5. EVIDENCE-BASED DESIGN PRINCIPLES
The detection methodology is informed by the following evidence-based principles:
5.1 Continuum Model of Suicidality. The compliance rules are calibrated to detect content across the full spectrum from ideation to planning to instruction, rather than only flagging explicit statements. This ensures that early warning signs are captured and acted upon.
5.2 Restriction as Prevention. Consistent with public health research (see, e.g., Yip et al., 2012), Rules L1 and L3 specifically target the provision of method-specific instructions, a recognized risk amplifier. Restricting access to instructional content regarding means of self-harm is an established harm-reduction strategy.
5.3 Safe Messaging Guidelines. The permitted/prohibited distinction in Rule L1 mirrors frameworks published by the WHO and the AFSP that differentiate responsible discussion from harmful content. The System is designed to permit educational and help-seeking discourse while suppressing content that may encourage or facilitate self-harm.
5.4 Contextual Integrity. By passing platform and user-request context into the Compliance Screening Pipeline, the System honors the principle that risk is situational and not purely lexical. Risk assessment is therefore calibrated to the specific deployment environment, user signals, and conversational context in which content arises.
6. CRISIS RESPONSE AND USER REFERRAL OBLIGATIONS
6.1 Mandatory Crisis Referral. Upon detection by the Compliance Screening Pipeline of any expression of suicidal ideation, suicide, or self-harm by a user, the System shall immediately trigger a crisis response notification to that User.
At a minimum, such notification shall refer the user to Crisis Service Providers. This obligation applies to both direct expressions of suicidal ideation and to escalating risk patterns identified through the contextual analysis described in Section 3.3.
6.2 Content Suppression and Referral Sequencing. Upon triggering of the crisis response protocol, the System shall: (a) suppress any AI Companion response that constitutes Prohibited Content; and (b) substitute or accompany any response with a crisis referral notification in accordance with Section 5.1. Following delivery of a crisis referral notification, the AI Companion shall not re-engage the user in a manner that generates, encourages, or facilitates Suicide Content or Self-Harm Content.
6.3 Extended Harm Scope. The crisis response protocol is designed to be extensible to address expressions of possible physical harm to others and possible financial harm to others expressed by a user, consistent with evolving legislative requirements. Any extension of the crisis response protocol to cover these categories shall be implemented by amendment to this Protocol in accordance with Section 7.
7. REVIEW AND AMENDMENT
This Protocol shall be reviewed no less than annually and shall be updated to reflect: (a) material changes to the Platform's detection architecture; (b) updated guidance from the WHO, AFSP, or equivalent authoritative bodies on safe messaging and suicide prevention; (c) developments in applicable law or regulation governing online safety and AI-Generated content; and (d) emerging research in computational mental health and suicide risk detection. Any amendments to this Protocol shall be approved by the designated compliance officer and documented in a version-controlled log.