HIP-220: Bias Detection & Mitigation. Status Draft. Hanzo's own standard — read this before implementing against it.
This HIP establishes the framework for detecting, measuring, and mitigating bias in Hanzo AI systems. It defines bias categories, evaluation methodologies, fairness metrics, and remediation processes to ensure AI systems treat all users equitably.
Algorithmic bias: Systematic and repeatable errors in a computer system that create unfair outcomes for certain groups.
Fairness: The principle that AI systems should treat individuals and groups equitably, without discrimination based on protected characteristics.
| Category | Characteristics | |----------|-----------------| | Demographic | Race, ethnicity, nationality, religion | | Personal | Age, gender, sexual orientation, disability | | Socioeconomic | Income, education, occupation | | Geographic | Region, urban/rural, language |
| Type | Source | Example | |------|--------|---------| | Historical | Past discrimination in data | Hiring data reflecting past discrimination | | Representation | Underrepresentation in data | Fewer examples of certain groups | | Measurement | How data is collected | Sensors less accurate for certain skin tones |
| Type | Source | Example | |------|--------|---------| | Aggregation | One-size-fits-all models | Medical model trained primarily on one population | | Learning | Algorithm amplification | Feedback loops reinforcing initial bias | | Evaluation | Biased benchmarks | Test sets not representative |
| Type | Source | Example | |------|--------|---------| | Deployment | Context mismatch | Model used in unintended population | | Interaction | User behavior patterns | Different usage across groups | | Temporal | Changing contexts | Society changes, model doesn't |
| Method | Application | Tools | |--------|-------------|-------| | Demographic parity | Equal prediction rates | Statistical analysis | | Equalized odds | Equal TPR/FPR across groups | Fairlearn | | Calibration | Consistent probability meaning | Reliability diagrams | | Individual fairness | Similar treatment for similar individuals | Distance metrics |
| Method | Application | Approach | |--------|-------------|----------| | Output auditing | Review generated content | Human evaluation | | Prompt testing | Test with demographic markers | Structured prompts | | User studies | Perception of fairness | Surveys, interviews |
| Benchmark | Bias Type | Metrics | |-----------|-----------|---------| | WinoBias | Gender | Accuracy parity | | StereoSet | Stereotype | Language model score | | CrowS-Pairs | Multiple | Preference score | | BBQ | Social bias | Accuracy in ambiguous contexts |
| Dataset | Coverage | Size | |---------|----------|------| | Hanzo-Bias-1K | Multi-category bias probes | 1,000+ | | Regional-Fairness | Geographic/cultural bias | 500+ | | Intersectional-Set | Intersecting identities | 300+ |
1. Define evaluation scope (groups, use cases)
↓
2. Select appropriate metrics
↓
3. Run quantitative evaluation
↓
4. Conduct qualitative review
↓
5. Document findings
↓
6. Determine if thresholds met| Activity | Frequency | Scope | |----------|-----------|-------| | Automated metrics | Weekly | Key fairness metrics | | Sample audits | Monthly | Human review sample | | Full audit | Quarterly | Comprehensive evaluation |
Definition: Prediction rates equal across groups
Formula:
P(Ŷ=1|A=a) = P(Ŷ=1|A=b) for all groups a, bThreshold: <10% difference across groups
Definition: Equal true positive and false positive rates across groups
Formulas:
P(Ŷ=1|Y=1,A=a) = P(Ŷ=1|Y=1,A=b) (Equal TPR)
P(Ŷ=1|Y=0,A=a) = P(Ŷ=1|Y=0,A=b) (Equal FPR)Threshold: <15% difference in TPR, <10% difference in FPR
Definition: Equal precision across groups
Formula:
P(Y=1|Ŷ=1,A=a) = P(Y=1|Ŷ=1,A=b)Threshold: <10% difference across groups
Definition: Similar individuals receive similar treatment
Formula:
|f(x) - f(x')| ≤ d(x, x') for similar x, x'Definition: Tendency to associate groups with stereotypes
Measurement: Compare likelihood of stereotypical vs. anti-stereotypical completions
Target: Score ≈ 50% (no preference)
| Metric | Definition | Target | |--------|------------|--------| | Mention parity | Equal mention rates | Within 20% | | Sentiment parity | Equal sentiment | No significant difference | | Association | Co-occurrence patterns | No stereotypical clustering |
Analyze bias at intersection of multiple characteristics:
| Intersection | Analysis | |--------------|----------| | Gender × Race | Check for compounded bias | | Age × Disability | Assess unique patterns | | Region × Language | Cultural intersections |
| Strategy | Application | |----------|-------------| | Data balancing | Ensure representative training data | | Data augmentation | Add underrepresented examples | | Data filtering | Remove biased content | | Source diversification | Include diverse data sources |
| Strategy | Application | |----------|-------------| | Debiasing techniques | RLHF, constitutional AI | | Regularization | Fairness constraints in loss | | Adversarial training | Train against bias detectors | | Multi-task learning | Include fairness objectives |
| Strategy | Application | |----------|-------------| | Output filtering | Flag biased outputs | | Calibration | Adjust outputs for fairness | | Prompt engineering | Include fairness instructions | | Human review | Manual review of sensitive outputs |
| Strategy | Application | |----------|-------------| | Use case restrictions | Limit high-risk applications | | User warnings | Inform users of limitations | | Feedback loops | Collect bias reports | | A/B testing | Test mitigation effectiveness |
1. Complete bias assessment questionnaire
↓
2. Run standard benchmark suite
↓
3. Conduct demographic analysis
↓
4. Document known limitations
↓
5. Implement mitigations
↓
6. Verify mitigation effectiveness
↓
7. Obtain fairness sign-off| Activity | Frequency | Owner | |----------|-----------|-------| | Metric monitoring | Weekly | MRM Team | | Benchmark re-run | Monthly | Safety Team | | Full audit | Quarterly | External + Internal | | Process review | Annual | ESG Committee |
| Section | Contents | |---------|----------| | Groups evaluated | Demographic breakdown | | Metrics used | Fairness metrics applied | | Results | Metric values by group | | Known limitations | Identified bias patterns | | Mitigations applied | Steps taken to address |
| Model Risk Tier | Demographic Parity | Equalized Odds | |-----------------|-------------------|----------------| | Critical | <5% difference | <10% difference | | High | <10% difference | <15% difference | | Medium | <15% difference | <20% difference | | Low | <20% difference | <25% difference |
| Body | Role | |------|------| | Fairness Lead | Day-to-day oversight | | Safety Team | Evaluation execution | | ESG Committee | Policy and escalations | | External Board | Independent review |
| Trigger | Escalation | |---------|------------| | Threshold violation | Fairness Lead → MRM Lead | | User complaint | Support → Fairness Lead | | External report | Communications → ESG Committee | | Regulatory inquiry | Legal → Board |
| Version | Date | Changes | |---------|------|---------| | 1.0 | 2025-12-17 | Initial draft |
Copyright and related rights waived via CC0.