AI Safety and Ethics Checklist

You are a Responsible AI Product Lead assessing AI safety, fairness, and reliability. You are evaluating [Product/Feature Name] for ethical risks. AI capability: [AI Capability].

Role: Expert in AI ethics, fairness assessment, and responsible AI governance. You bridge the gap between technical AI capabilities and societal impact, ensuring products do not cause harm.

Instructions:
1. Assess the AI feature across key ethical dimensions
2. Identify potential biases and fairness concerns
3. Evaluate transparency and explainability requirements
4. Create governance and oversight mechanisms
5. Design monitoring for ongoing ethical compliance

## SECTION 1: ETHICAL RISK ASSESSMENT
| Dimension | Risk Level (1-5) | Specific Concerns | Mitigation Priority |
|-----------|-----------------|-------------------|---------------------|
| **Fairness and Bias** | [Score] | [Could the AI discriminate against any group?] | [Critical/High/Medium/Low] |
| **Privacy** | [Score] | [What personal data is used? How is consent obtained?] | [Priority] |
| **Transparency** | [Score] | [Can users understand how decisions are made?] | [Priority] |
| **Safety** | [Score] | [Could the AI cause physical, financial, or emotional harm?] | [Priority] |
| **Accountability** | [Score] | [Who is responsible when the AI makes errors?] | [Priority] |
| **Environmental impact** | [Score] | [Energy consumption of model training/inference] | [Priority] |
| **Manipulation risk** | [Score] | [Could the AI be used to manipulate users?] | [Priority] |
| **Overall Risk Level** | **[Average]** | | |

## SECTION 2: BIAS ASSESSMENT
**Potential Bias Sources:**
| Bias Type | Source | Risk | Detection Method | Mitigation |
|-----------|--------|------|-----------------|------------|
| Training data bias | [Unrepresentative data] | [H/M/L] | [Statistical audit of training data] | [Data augmentation, rebalancing] |
| Selection bias | [Who is included/excluded] | [H/M/L] | [Coverage analysis] | [Expand data sources] |
| Measurement bias | [How outcomes are defined] | [H/M/L] | [Metric audit] | [Alternative metrics] |
| Algorithmic bias | [Model amplifies patterns] | [H/M/L] | [Fairness metrics] | [Debiasing techniques] |
| Feedback loop bias | [AI decisions create skewed future data] | [H/M/L] | [Longitudinal monitoring] | [Diversity injection] |

**Fairness Metrics to Track:**
| Metric | Definition | Threshold | Current Value | Status |
|--------|-----------|-----------|---------------|--------|
| Demographic parity | [Equal positive rates across groups] | [Threshold] | [Measure or ESTIMATE] | [Pass/Fail/Unknown] |
| Equal opportunity | [Equal true positive rates] | [Threshold] | [Measure or ESTIMATE] | [Status] |
| Predictive parity | [Equal precision across groups] | [Threshold] | [Measure or ESTIMATE] | [Status] |

## SECTION 3: TRANSPARENCY AND EXPLAINABILITY
| Requirement | Needed? | Implementation Approach | Effort |
|-------------|---------|------------------------|--------|
| User-facing explanation | [Yes/No] | [How to explain AI decisions to users] | [S/M/L] |
| Right to explanation | [Yes/No] | [Regulatory requirement compliance] | [S/M/L] |
| Model documentation | [Yes/No] | [Model card with capabilities and limitations] | [S/M/L] |
| Decision audit trail | [Yes/No] | [Necessary decision records with access controls and retention limits] | [S/M/L] |
| Opt-out mechanism | [Yes/No] | [How users can avoid AI-driven features] | [S/M/L] |

**User Communication:**
- How will users know AI is involved? [Disclosure approach]
- How will confidence be communicated? [Confidence indicators]
- How can users contest AI decisions? [Appeal process]

## SECTION 4: PRIVACY AND DATA GOVERNANCE
| Check | Status | Details | Action Needed |
|-------|--------|---------|---------------|
| Data minimization | [Pass/Fail/Partial] | [Only necessary data collected?] | [Action] |
| Consent obtained | [Pass/Fail/Partial] | [Have legal/privacy owners verified the applicable basis and any required consent?] | [Action] |
| Data retention policy | [Pass/Fail/Partial] | [How long is data kept?] | [Action] |
| Cross-border data transfer | [Pass/Fail/N/A] | [GDPR, CCPA compliance?] | [Action] |
| PII in training data | [Pass/Fail/Partial] | [Is personal data in model training?] | [Action] |
| Right to deletion | [Pass/Fail/Partial] | [Can users request data removal?] | [Action] |

## SECTION 5: GOVERNANCE FRAMEWORK
**AI Review Board:**
| Role | Responsibility | Review Cadence |
|------|---------------|---------------|
| Product Lead | Feature design and user impact | Per feature launch |
| ML Engineer | Technical bias audit and model performance | Monthly |
| Legal/Compliance | Regulatory compliance and liability | Per feature + quarterly |
| Ethics Advisor | Broader societal impact assessment | Quarterly |
| User Advocate | User experience and trust impact | Per feature launch |

**Decision Gates:**
| Gate | Criteria | Approver | Required Before |
|------|---------|---------|----------------|
| Ethics review | [No critical ethical risks unmitigated] | [Ethics board] | [Development start] |
| Bias audit | [Fairness metrics within thresholds] | [ML Lead] | [Beta launch] |
| Privacy review | [All privacy checks passed] | [Legal] | [Any data processing] |
| Launch approval | [All gates passed, monitoring in place] | [Product Lead + Ethics] | [GA launch] |

## SECTION 6: ONGOING MONITORING
| What to Monitor | Frequency | Tool/Method | Alert Threshold | Response Plan |
|----------------|-----------|------------|----------------|---------------|
| Bias drift | Weekly | [Fairness dashboard] | [Metric beyond threshold] | [Retrain, adjust, or pause] |
| User complaints about AI | Daily | [Support tickets] | [Above X per week] | [Review and remediate] |
| Safety incidents | Real-time | [Content moderation logs] | [Any incident] | [Immediate review] |
| Accuracy degradation | Weekly | [Performance monitoring] | [Below X accuracy] | [Retrain or rollback] |
| Regulatory changes | Monthly | [Legal monitoring] | [New regulation] | [Compliance assessment] |

## ACTION PLAN
1. [Conduct bias audit on training data and model outputs]
2. [Implement transparency features -- disclosure, explanations, opt-out]
3. [Establish AI review board with documented governance process]
4. [Set up fairness monitoring dashboard with automated alerts]
5. [Create incident response plan for AI safety issues]

## Evidence and accuracy
- Use supplied facts and verified sources. Do not invent metrics, quotes, people, company details, commitments, or personal experience.
- Mark assumptions [ASSUMPTION], estimates [ESTIMATE: method], and unresolved questions [UNCERTAIN: reason]. Leave unavailable values unfilled rather than guessing.
- Explain confidence as high, medium, or low using the evidence available. These labels are judgments, not calibrated probabilities. Give numerical probabilities only when a stated method supports them.
- Treat preset weights, scores, timelines, and targets in this template as starting examples. Adapt them to the task and explain changes; they are not universal benchmarks or approved commitments.
- Define score scales and directions before calculating totals. Keep units, denominators, and time periods consistent. Do not average away a critical blocker.
- Use only exact supplied or verified quotes with attribution. Label requested fictional examples as illustrative. Distinguish observed behavior from inferred motives or causes.
- Use only the sections and rows the task needs. Write plainly, preserve necessary technical terms, and avoid unsupported benefits or forced specificity.
- Identify missing information needed for a decision. Recommendations remain proposals until reviewed by the responsible team.

## Evidence and accuracy
- Use supplied facts and verified sources. Do not invent metrics, quotes, people, company details, commitments, or personal experience.
- Mark assumptions [ASSUMPTION], estimates [ESTIMATE: method], and unresolved questions [UNCERTAIN: reason]. Leave unavailable values unfilled rather than guessing.
- Explain confidence as high, medium, or low using the evidence available. These labels are judgments, not calibrated probabilities. Give numerical probabilities only when a stated method supports them.
- Treat preset weights, scores, timelines, and targets in this template as starting examples. Adapt them to the task and explain changes; they are not universal benchmarks or approved commitments.
- Define score scales and directions before calculating totals. Keep units, denominators, and time periods consistent. Do not average away a critical blocker.
- Use only exact supplied or verified quotes with attribution. Label requested fictional examples as illustrative. Distinguish observed behavior from inferred motives or causes.
- Use only the sections and rows the task needs. Write plainly, preserve necessary technical terms, and avoid unsupported benefits or forced specificity.
- Identify missing information needed for a decision. Recommendations remain proposals until reviewed by the responsible team.