AI Ethics Checklist
Purpose: Ensure responsible AI development and deployment.
Overview
As AI becomes embedded in products, PMs and BAs must consider ethical implications beyond just technical feasibility. This checklist helps identify and mitigate potential harms.
When to use: Before building any feature that uses AI/ML (recommendations, predictions, automation, personalization).
Checklist
1. Fairness & Bias
Question: Could this AI system discriminate against protected groups?
- [ ] Training data is representative of all user groups (not skewed toward one demographic)
- [ ] Model has been tested for bias across race, gender, age, disability, etc.
- [ ] Outcomes audited: Do different groups experience different error rates?
- [ ] Mitigation plan exists if bias is detected (e.g., rebalance training data)
Example:
❌ Bad: Resume screening AI trained only on profiles of past hires (who are 80% male) → Systematically ranks women lower
✅ Good: Audit model performance by gender, adjust training data to balance representation
2. Transparency & Explainability
Question: Can users understand how the AI made a decision?
- [ ] Users are informed when AI is being used (not hidden)
- [ ] Key factors influencing decisions are explainable (not a "black box")
- [ ] Users can challenge or appeal AI decisions
- [ ] Documentation exists explaining model logic (for regulators, auditors)
Example:
❌ Bad: Loan application denied by AI with no explanation
✅ Good: "Loan denied due to: debt-to-income ratio (40%), recent missed payment (30%), short credit history (30%)"
3. Privacy & Data Protection
Question: Are we respecting user privacy?
- [ ] Minimum data collection (only what's necessary for the AI to function)
- [ ] User consent obtained before using personal data for AI training
- [ ] Data anonymized where possible (remove PII from training sets)
- [ ] Compliance with GDPR, CCPA, and other privacy regulations
- [ ] Users can opt out of AI features (or request data deletion)
Example:
❌ Bad: Using customer chat transcripts to train AI without informing users
✅ Good: "We use your anonymized chat data to improve our chatbot. [Opt Out]"
4. Accountability
Question: Who is responsible if the AI causes harm?
- [ ] Clear ownership: Who monitors AI performance post-launch?
- [ ] Incident response plan: What happens if AI makes a harmful decision?
- [ ] Human-in-the-loop for high-stakes decisions (medical, legal, financial)
- [ ] AI decisions are auditable (logs retained for review)
Example:
❌ Bad: Fully automated hiring decisions with no human review
✅ Good: AI shortlists candidates → Human recruiter makes final decision
5. Safety & Robustness
Question: Could the AI fail in unexpected ways?
- [ ] Adversarial testing: Can users game or manipulate the system?
- [ ] Edge case testing: What happens with unusual inputs?
- [ ] Graceful degradation: If AI fails, does the system fall back safely?
- [ ] Model monitoring: Alert if accuracy drops below threshold
Example:
❌ Bad: Restaurant recommendation AI always suggests closed restaurants (stale data)
✅ Good: Real-time validation → If restaurant closed, exclude from recommendations
6. User Autonomy
Question: Does the AI empower or manipulate users?
- [ ] Users can override AI suggestions (AI assists, doesn't dictate)
- [ ] Dark patterns avoided (e.g., AI nudging users to spend more money)
- [ ] Addiction mechanisms avoided (e.g., infinite scroll optimized by AI)
- [ ] Users understand consequences of accepting AI recommendations
Example:
❌ Bad: AI auto-subscribes users to premium plan "because it predicts they'll want it"
✅ Good: AI suggests premium plan → User consciously chooses to upgrade
7. Environmental Impact
Question: What is the carbon footprint of this AI?
- [ ] Model size justified (do we need GPT-4 or will a smaller model work?)
- [ ] Training done on renewable energy infrastructure (when possible)
- [ ] Inference optimized (reduce compute per prediction)
- [ ] Carbon cost documented and considered in ROI
Example:
❌ Bad: Retrain massive model daily for marginal accuracy gain (+0.1%)
✅ Good: Retrain quarterly, use smaller model for 95% of use cases
8. Accessibility
Question: Can people with disabilities use this AI feature?
- [ ] AI outputs are screen-reader friendly
- [ ] Voice-based AI includes text alternatives (for deaf users)
- [ ] Visual AI includes audio descriptions (for blind users)
- [ ] AI doesn't rely solely on one modality (e.g., facial recognition excludes people with face coverings)
Example:
❌ Bad: Voice-only chatbot (excludes deaf/hard-of-hearing users)
✅ Good: Chatbot supports both voice and text input/output
9. Truthfulness & Misinformation
Question: Could this AI spread false information?
- [ ] AI-generated content is labeled as such (not passed off as human-created)
- [ ] Fact-checking mechanisms in place (for news, health, legal advice)
- [ ] Sources cited (if AI makes factual claims)
- [ ] Harmful misinformation filtered (e.g., medical advice that could cause harm)
Example:
❌ Bad: AI generates medical advice with no disclaimers or source citations
✅ Good: "This information is AI-generated and not a substitute for professional medical advice. Consult a doctor."
10. Job Displacement & Economic Impact
Question: Will this AI eliminate jobs without a transition plan?
- [ ] Impact assessment: How many jobs could be affected?
- [ ] Reskilling programs offered (if automating roles)
- [ ] Augmentation prioritized over replacement (AI assists workers, doesn't replace them)
- [ ] Stakeholder communication: Workers informed early about automation plans
Example:
❌ Bad: Replace all customer service reps with chatbots overnight
✅ Good: Chatbot handles simple FAQs → Reps focus on complex issues (upskill to "escalation specialists")
Case Study: ReserveEasy AI Features
Use Case: Personalized Restaurant Recommendations
Fairness:
- [x] Training data includes diverse cuisines (not just Italian, French)
- [x] Model tested across zip codes (doesn't only recommend expensive restaurants to wealthy neighborhoods)
Transparency:
- [x] Users see why restaurants are recommended: "Based on your past bookings (Italian, seafood) and high ratings"
Privacy:
- [x] User consent: "ReserveEasy uses your booking history to personalize recommendations. [Manage Privacy Settings]"
Autonomy:
- [x] Recommendations are suggestions, not forced (users can search/browse freely)
Red Flags (When to Pause Development)
⚠️ Stop and reassess if:
- Model shows >10% accuracy difference across demographic groups
- Users can't understand why AI made a decision
- AI makes decisions that could significantly harm users (financially, emotionally, physically)
- No human oversight for high-stakes outcomes
- Privacy regulators raise concerns
Resources
Learn More: