Privacy & Responsible AI Skill

Privacy by design, data protection, and responsible AI principles.

⚠️ Staleness Warning

Privacy regulations and AI ethics guidelines evolve continuously.

Refresh triggers:

• New privacy laws (state, country, region)
• AI regulation updates (EU AI Act, etc.)
• Industry standard changes
• Major incident learnings
• Annual transparency reports (Microsoft, Google)

Last validated: February 2026

Check current state: Microsoft RAI, Google AI Principles, GDPR, CCPA

---

Privacy by Design Principles

Principle	Implementation
Minimize	Collect only what's needed
Purpose	Use data only for stated purpose
Consent	Get explicit permission
Access	Let users see their data
Deletion	Let users delete their data
Security	Protect data at rest and in transit
Transparency	Explain what you collect and why

Data Classification

Level	Examples	Handling
Public	Marketing content	No restrictions
Internal	Employee directory	Internal only
Confidential	Customer data, PII	Encrypted, access-controlled
Restricted	Credentials, health data	Maximum protection

PII Checklist

Personal Identifiable Information includes:

• Names
• Email addresses
• Phone numbers
• Physical addresses
• IP addresses
• Device IDs
• Location data
• Financial data
• Health data
• Biometric data

---

Responsible AI Principles

Microsoft's 6 Principles (2025 RAI Transparency Report)

Principle	Question to Ask	Implementation
Fairness	Does it treat all groups equitably?	Bias testing, diverse datasets, fairness metrics
Reliability & Safety	Does it work consistently and safely?	Testing, monitoring, failure modes, guardrails
Privacy & Security	Does it protect user data?	Data minimization, encryption, access controls
Inclusiveness	Does it work for everyone?	Accessibility, diverse user testing, edge cases
Transparency	Can users understand how it works?	Explainability, documentation, model cards
Accountability	Who is responsible for outcomes?	Human oversight, audit trails, governance

Google's 3 Pillars (2024 AI Responsibility Report)

Pillar	Description
Bold Innovation	Deploy AI where benefits substantially outweigh risks
Responsible Development	Human oversight, safety research, bias mitigation, privacy
Collaborative Progress	Enable ecosystem, share learnings, engage stakeholders

Key RAI Tools & Frameworks

Tool	Purpose	Source
HAX Workbook	Human-AI interaction best practices	Microsoft
Responsible AI Dashboard	End-to-end RAI experience	Microsoft/Azure
Model Cards	Structured model documentation	Google
People + AI Guidebook	Design guidance for AI products	Google PAIR
Frontier Safety Framework	Advanced model risk management	Google

---

Bias Detection

Ask:
1. What data was the model trained on?
2. Are there underrepresented groups?
3. What are the failure modes?
4. Who might be harmed by errors?
5. Have we tested with diverse inputs?
6. What demographic slices show performance gaps?
7. Are there proxy variables that encode bias?

Bias Categories

Type	Description	Example
Selection Bias	Training data not representative	Hiring model trained only on past hires
Measurement Bias	Flawed data collection	Self-reported data with social desirability
Algorithmic Bias	Model amplifies patterns	Recommendation loops
Presentation Bias	UI choices influence perception	Image ordering in search results

---

AI Transparency & Documentation

Model Card Template

## Model Card: [Model Name]

### Model Details
- **Developer**: [Organization]
- **Version**: [Version number]
- **Type**: [Classification/Generation/etc.]
- **License**: [License terms]

### Intended Use
- **Primary use cases**: [Description]
- **Out-of-scope uses**: [What NOT to use it for]
- **Users**: [Target users]

### Training Data
- **Sources**: [Data sources]
- **Size**: [Dataset size]
- **Known limitations**: [Data gaps]

### Performance
- **Metrics**: [Evaluation metrics]
- **Sliced analysis**: [Performance by demographic groups]
- **Failure modes**: [Known failure patterns]

### Ethical Considerations
- **Risks**: [Potential harms]
- **Mitigations**: [Steps taken]
- **Human oversight**: [Review processes]

AI Feature Transparency (User-Facing)

## How This AI Works

**What it does**: [Clear description]
**What it doesn't do**: [Limitations]
**Data used**: [What inputs, how stored]
**Human oversight**: [When humans review]
**How to appeal**: [Process for disputes]
**Confidence indicators**: [How certainty is communicated]

---

Human-AI Collaboration

Appropriate Reliance Framework

State	Description	Signal
Over-reliance	Blind acceptance	User never questions AI
Appropriate reliance	Calibrated trust	User verifies when uncertain
Under-reliance	Excessive skepticism	User ignores useful AI output

Design for Appropriate Reliance

1. Show confidence levels — Don't present all outputs as equally certain
2. Explain reasoning — Help users evaluate AI logic
3. Enable challenge — Make it easy to question or override
4. Provide alternatives — Show multiple options when available
5. Track calibration — Monitor if users trust appropriately

Code Patterns

Don't Log PII

// Bad
console.log(`User ${email} logged in`);

// Good
console.log(`User ${hashUserId(userId)} logged in`);

Consent Before Collection

// Get explicit consent
const consent = await showConsentDialog({
    purpose: 'Improve recommendations',
    data: ['usage patterns', 'preferences'],
    retention: '90 days',
    optOut: 'Settings > Privacy'
});

if (!consent.granted) {
    return fallbackBehavior();
}

Data Minimization

// Bad: Store everything
saveUser({ ...fullUserObject });

// Good: Store only what's needed
saveUser({
    id: user.id,
    preferences: user.preferences
    // Don't store: email, name, location
});

Right to Deletion

async function deleteUserData(userId: string) {
    await db.users.delete(userId);
    await db.userPreferences.delete(userId);
    await db.userHistory.delete(userId);
    await analytics.purge(userId);
    await logs.redact(userId);

    return { deleted: true, timestamp: new Date() };
}

Regulatory Quick Reference

Regulation	Region	Key Requirements
GDPR	EU	Consent, access, deletion, breach notification
CCPA/CPRA	California	Disclosure, opt-out, deletion
LGPD	Brazil	Similar to GDPR
PIPL	China	Data localization, consent
HIPAA	US Healthcare	Health data protection

AI Incident Response

When AI causes harm:

1. Stop — Disable the feature immediately
2. Assess — Who was affected, how severely
3. Notify — Inform affected users
4. Fix — Root cause + prevention
5. Document — Post-mortem for learning

Synapses

See synapses.json for connections.