Implementing content filtering for harmful output types. For example, detecting and blocking distressed responses, angry language, offensive content, biased statements, and deceptive information.
Content filtering rules, moderation API configuration, or classifier settings showing detection and blocking logic for harmful output types - may include filtering rules in code, third-party moderation tool configuration (e.g., OpenAI Moderation API, Perspective API), or custom classifier model settings with harm category definitions.
Implementing guardrails for advice generation. For example, restricting high-risk recommendations in sensitive domains, requiring disclaimers for guidance.
System prompts, guardrail rules, or domain restrictions showing safety controls on advice generation - may include defensive prompting, domain-specific output restrictions (e.g., medical/legal/financial advice blocklists), or conditional response templates that add warnings for sensitive topics.
Implementing bias detection and mitigation controls. For example, monitoring for discriminatory patterns, implementing fairness checks in outputs.
Documentation of bias eval results testing for stereotypical responses across demographic attributes, manual review logs documenting bias assessments, or output filtering rules blocking discriminatory patterns - may include automated fairness evaluation tools or bias monitoring dashboards if implemented.
Evaluating harm mitigation controls using performance metrics.
Test results, metrics dashboard, or evaluation report showing performance of harm controls - may include false positive/negative rates, coverage analysis of test scenarios, benchmark results against harm datasets (e.g., ToxiGen, RealToxicityPrompts), or confusion matrices showing filtering accuracy across harm categories.
Organizations can submit alternative evidence demonstrating how they meet the requirement.