ethics-review（伦理审查）

Ethics Review

Comprehensive guidance for ethical assessment of technology systems, AI applications, and responsible innovation.

When to Use This Skill

• Conducting ethical impact assessments for new projects
• Evaluating AI systems for ethical risks
• Establishing ethics review boards and processes
• Developing ethical guidelines for technology teams
• Assessing stakeholder impacts and potential harms

Core Ethical Principles

Foundation Principles

Technology-Specific Principles

AI/ML Systems:
├── Fairness - Equitable treatment across groups
├── Explainability - Understandable decisions
├── Reliability - Consistent, predictable behavior
├── Safety - Prevent harm, fail safely
├── Privacy - Protect personal data
├── Security - Resist adversarial attacks
├── Inclusiveness - Accessible to all users
└── Human Control - Meaningful human oversight

Ethical Impact Assessment Framework

Assessment Process

┌─────────────────────────────────────────────────────────────┐
│                  Ethical Impact Assessment                   │
├─────────────────────────────────────────────────────────────┤
│  1. Describe     │  System purpose, capabilities, context   │
├──────────────────┼──────────────────────────────────────────┤
│  2. Stakeholder  │  Identify all affected parties           │
│     Analysis     │  Map interests and concerns              │
├──────────────────┼──────────────────────────────────────────┤
│  3. Impact       │  Assess benefits and harms               │
│     Assessment   │  Evaluate likelihood and severity        │
├──────────────────┼──────────────────────────────────────────┤
│  4. Ethical      │  Apply ethical principles                │
│     Analysis     │  Identify conflicts and tensions         │
├──────────────────┼──────────────────────────────────────────┤
│  5. Mitigation   │  Design controls and safeguards          │
│     Planning     │  Define monitoring approach              │
├──────────────────┼──────────────────────────────────────────┤
│  6. Decision &   │  Approve, modify, or reject              │
│     Review       │  Schedule ongoing review                 │
└─────────────────────────────────────────────────────────────┘

Ethical Impact Assessment Template

# Ethical Impact Assessment

## 1. System Description

### Purpose
[What is the system designed to do?]

### Capabilities
[What can the system do? What decisions does it make or influence?]

### Context
[Where and how will the system be used?]

### Data
[What data does the system use? How is it collected?]

---

## 2. Stakeholder Analysis

### Direct Stakeholders
| Stakeholder | Relationship | Interests | Power | Concerns |
|-------------|--------------|-----------|-------|----------|
| [Group] | [Relationship] | [Interests] | [H/M/L] | [Concerns] |

### Indirect Stakeholders
| Stakeholder | How Affected | Interests | Concerns |
|-------------|--------------|-----------|----------|
| [Group] | [Impact] | [Interests] | [Concerns] |

### Vulnerable Groups
| Group | Vulnerability | Special Considerations |
|-------|---------------|----------------------|
| [Group] | [Why vulnerable] | [Protections needed] |

---

## 3. Impact Assessment

### Benefits
| Benefit | Beneficiary | Magnitude | Likelihood |
|---------|-------------|-----------|------------|
| [Benefit] | [Who] | [H/M/L] | [H/M/L] |

### Potential Harms
| Harm | Affected Group | Severity | Likelihood | Reversible? |
|------|----------------|----------|------------|-------------|
| [Harm] | [Who] | [H/M/L] | [H/M/L] | [Y/N] |

### Unintended Consequences
| Consequence | Description | Risk Level |
|-------------|-------------|------------|
| [Consequence] | [Details] | [H/M/L] |

---

## 4. Ethical Analysis

### Principle Evaluation

| Principle | Supports | Tensions | Score (1-5) |
|-----------|----------|----------|-------------|
| Beneficence | [How] | [Conflicts] | [Score] |
| Non-maleficence | [How] | [Conflicts] | [Score] |
| Autonomy | [How] | [Conflicts] | [Score] |
| Justice | [How] | [Conflicts] | [Score] |
| Transparency | [How] | [Conflicts] | [Score] |
| Accountability | [How] | [Conflicts] | [Score] |
| Privacy | [How] | [Conflicts] | [Score] |

### Ethical Dilemmas
| Dilemma | Trade-off | Proposed Resolution |
|---------|-----------|---------------------|
| [Dilemma] | [Trade-off] | [Resolution] |

---

## 5. Mitigation Plan

### Technical Mitigations
| Risk | Mitigation | Owner | Status |
|------|------------|-------|--------|
| [Risk] | [Control] | [Who] | [Status] |

### Procedural Mitigations
| Risk | Mitigation | Owner | Status |
|------|------------|-------|--------|
| [Risk] | [Process] | [Who] | [Status] |

### Monitoring Plan
| Metric | Threshold | Frequency | Response |
|--------|-----------|-----------|----------|
| [Metric] | [Limit] | [How often] | [Action] |

---

## 6. Decision

### Recommendation
[ ] Approve - Proceed with current design
[ ] Approve with conditions - Proceed after mitigations
[ ] Defer - Requires further analysis
[ ] Reject - Unacceptable ethical risks

### Conditions (if applicable)
1. [Condition]
2. [Condition]

### Review Schedule
- Initial review: [Date]
- Ongoing review: [Frequency]

### Approvals
| Role | Name | Decision | Date |
|------|------|----------|------|
| Ethics Board | | [ ] | |
| Technical Lead | | [ ] | |
| Business Owner | | [ ] | |
| Legal | | [ ] | |

Harm Assessment Framework

Categories of Harm

Direct Harms:
├── Physical harm to individuals
├── Psychological harm (stress, manipulation)
├── Financial harm (fraud, loss)
├── Privacy harm (exposure, surveillance)
├── Discrimination harm (unfair treatment)
└── Autonomy harm (manipulation, coercion)

Indirect/Systemic Harms:
├── Environmental harm
├── Democratic harm (manipulation, division)
├── Economic harm (displacement, inequality)
├── Social harm (erosion of trust, relationships)
└── Cultural harm (homogenization, loss)

Group-Specific Harms:
├── Harm to marginalized groups
├── Harm to vulnerable populations
├── Harm to future generations
└── Harm to non-users

Harm Severity Matrix

               REVERSIBILITY
               Easy    Difficult   Permanent
S      Low     1          2           3
E      Medium  2          4           6
V      High    3          6           9
E      Extreme 4          8          12
R
I
T
Y

Score:
1-2:  Acceptable with monitoring
3-4:  Requires mitigation
6-8:  Significant controls required
9-12: May be unacceptable

AI Ethics Specifics

AI Ethics Checklist

public class AiEthicsChecklist
{
    public List<EthicsCheckItem> GetChecklist()
    {
        return new List<EthicsCheckItem>
        {
            // Fairness
            new("FAIR-01", "Bias Testing",
                "Has the model been tested for bias across protected groups?",
                EthicsCategory.Fairness, Priority.Critical),
            new("FAIR-02", "Fairness Metrics",
                "Are fairness metrics defined and monitored?",
                EthicsCategory.Fairness, Priority.High),
            new("FAIR-03", "Training Data",
                "Is training data representative and free from historical bias?",
                EthicsCategory.Fairness, Priority.Critical),

            // Transparency
            new("TRANS-01", "Explainability",
                "Can the system explain its decisions to affected users?",
                EthicsCategory.Transparency, Priority.High),
            new("TRANS-02", "AI Disclosure",
                "Are users informed they are interacting with AI?",
                EthicsCategory.Transparency, Priority.Critical),
            new("TRANS-03", "Limitation Disclosure",
                "Are system limitations clearly communicated?",
                EthicsCategory.Transparency, Priority.High),

            // Human Control
            new("CTRL-01", "Human Oversight",
                "Is there meaningful human oversight of AI decisions?",
                EthicsCategory.HumanControl, Priority.Critical),
            new("CTRL-02", "Override Capability",
                "Can humans override AI decisions when needed?",
                EthicsCategory.HumanControl, Priority.High),
            new("CTRL-03", "Escalation Path",
                "Is there a clear escalation path for concerning outputs?",
                EthicsCategory.HumanControl, Priority.High),

            // Safety
            new("SAFE-01", "Harm Prevention",
                "Are there safeguards against harmful outputs?",
                EthicsCategory.Safety, Priority.Critical),
            new("SAFE-02", "Fail-Safe Design",
                "Does the system fail safely when errors occur?",
                EthicsCategory.Safety, Priority.High),
            new("SAFE-03", "Adversarial Testing",
                "Has the system been tested against adversarial inputs?",
                EthicsCategory.Safety, Priority.High),

            // Privacy
            new("PRIV-01", "Data Minimization",
                "Does the system collect only necessary data?",
                EthicsCategory.Privacy, Priority.High),
            new("PRIV-02", "Consent",
                "Is there informed consent for data use?",
                EthicsCategory.Privacy, Priority.Critical),
            new("PRIV-03", "Data Protection",
                "Is personal data adequately protected?",
                EthicsCategory.Privacy, Priority.Critical),

            // Accountability
            new("ACCT-01", "Responsibility",
                "Is there clear ownership for system outcomes?",
                EthicsCategory.Accountability, Priority.High),
            new("ACCT-02", "Audit Trail",
                "Are decisions logged for accountability?",
                EthicsCategory.Accountability, Priority.High),
            new("ACCT-03", "Redress Mechanism",
                "Is there a way for affected parties to seek redress?",
                EthicsCategory.Accountability, Priority.High)
        };
    }
}

Algorithmic Impact Questions

Ethics Review Board

Board Structure

Ethics Review Board Composition:
├── Chair (Senior Leadership)
├── Ethics Officer (if applicable)
├── Technical Lead (understands the technology)
├── Legal Representative
├── Privacy Officer
├── Business Representative
├── External Ethicist (optional but recommended)
└── User/Community Representative (for significant decisions)

Review Thresholds

Board Decision Framework

public enum EthicsDecision
{
    Approved,                    // Proceed as designed
    ApprovedWithConditions,      // Proceed after specified changes
    RequiresRedesign,           // Fundamental changes needed
    Deferred,                   // Need more information
    Rejected,                   // Unacceptable ethical risk
    EscalateToExecutive         // Beyond board authority
}

public class EthicsReviewResult
{
    public required EthicsDecision Decision { get; init; }
    public required string Rationale { get; init; }
    public List<string> Conditions { get; init; } = new();
    public List<string> MonitoringRequirements { get; init; } = new();
    public DateTimeOffset? NextReviewDate { get; init; }
    public List<BoardMemberVote> Votes { get; init; } = new();
}

Responsible Innovation Framework

Stage-Gate Ethics Integration

Stage 1: Ideation
├── Initial ethics screening
├── Identify potential concerns
└── Go/No-Go for research

Stage 2: Research & Design
├── Stakeholder analysis
├── Preliminary impact assessment
└── Ethics-by-design integration

Stage 3: Development
├── Ongoing ethics review
├── Testing for bias/harm
└── Documentation

Stage 4: Pre-Deployment
├── Full ethical impact assessment
├── Board review (if triggered)
└── Mitigation verification

Stage 5: Deployment
├── Monitoring plan activation
├── Feedback mechanisms
└── Incident response ready

Stage 6: Operations
├── Ongoing monitoring
├── Regular reviews
└── Continuous improvement

Ethics Review Checklist

Pre-Development

• Ethical impact assessment completed
• Stakeholder analysis documented
• Potential harms identified
• Ethics review board consulted (if required)
• Mitigation plans defined

Development

• Ethics-by-design principles applied
• Bias testing conducted
• Explainability built in
• Human oversight designed
• Documentation complete

Pre-Deployment

• Full assessment reviewed
• All mitigations implemented
• Monitoring in place
• Redress mechanism ready
• Ethics sign-off obtained

Operations

• Regular monitoring active
• Feedback collected and reviewed
• Incidents investigated
• Periodic re-assessment scheduled

Cross-References

• AI Governance: ai-governance for regulatory compliance
• Bias Assessment: Research fairness metrics via MCP (perplexity: "AI fairness metrics NIST")
• Data Privacy: gdpr-compliance for privacy considerations

Resources

• IEEE Ethically Aligned Design
• ACM Code of Ethics
• AI Ethics Guidelines Global Inventory
• Markkula Center for Applied Ethics