| Pillar: Ethical Safeguards |
| Fairness |
- Conducted bias audits on welfare eligibility models.
- Adjusted predictive policing zones based on community-defined harm thresholds.
|
Fairness vs. Privacy ■ is resolved by balancing the need for demographic data in fairness audits with privacy-preserving aggregation techniques.
|
| Inclusiveness |
- Offered dashboard translations in multiple languages.
- Designed low-tech options for digitally underserved residents.
|
Inclusiveness vs. Robustness ■ is resolved by designing adaptive interfaces that perform reliably across a wide range of user devices and access conditions.
|
| Bias Mitigation |
- Used counterfactual fairness testing.
- Conducted quarterly revalidation of policing datasets.
|
Bias Mitigation vs. Privacy ■ is resolved by applying bias audits using aggregated demographic data and privacy-preserving statistical methods to avoid re-identification risks during fairness assessments.
|
| Accountability |
- Established audit logs for all automated decisions.
- Created independent oversight committees.
|
Accountability vs. Privacy ■ is resolved by using pseudonymization within audit trails.
|
| Privacy |
- Deployed federated learning in traffic prediction systems.
- Used anonymized data in welfare screening.
|
Privacy
Privacy vs. Transparency ■ is resolved with tiered disclosure levels based on audience role and sensitivity.
|
| Pillar: Operational Integrity |
| Governance |
- Implemented municipal AI ethics charter.
- Conducted third-party audits on all public-facing models.
|
Governance vs. Transparency ■ is resolved by requiring structured disclosures reviewed by the ethics board to ensure accountability while avoiding disclosure of sensitive implementation details that could compromise system integrity or public safety.
|
| Robustness |
- Piloted tools across varied neighborhoods.
- Designed fallback protocols for infrastructure failures.
|
Robustness vs. Sustainability ■ is resolved by optimizing system performance for high variability without requiring resource-intensive redundancies, ensuring scalable and energy-efficient deployment across city systems.
|
| Interpretability |
- Built staff training modules with example-based explanations.
- Integrated scenario walkthroughs in internal dashboards.
|
Interpretability vs. Security ■ is resolved by limiting access to internal model logic through role-based controls.
|
| Explainability |
- Visualized benefit eligibility decisions.
- Provided simplified justifications for traffic rerouting choices.
|
Explainability vs. Privacy ■ is resolved by redacting identifiable data while maintaining output clarity.
|
| Security |
- Applied role-based access controls across city systems.
- Encrypted all interdepartmental data transfers.
|
Security vs. Transparency ■ is resolved by limiting disclosure of sensitive system components while still publishing high-level audit results and algorithmic summaries for public review.
|
| Safety |
- Implemented real-time system health checks and failure protocols for AI-controlled infrastructure (e.g., traffic systems).
- Required human review and override capabilities for all high-impact decision points (e.g., service denials in welfare automation).
|
Safety vs. Privacy ■ is resolved by implementing privacy-preserving logging mechanisms that support incident response and traceability without exposing identifiable user data.
|
| Pillar: Social Empowerment |
| Sustainability |
- Hosted models in green-certified data centers.
- Minimized compute load through efficient architecture.
|
Sustainability vs. Robustness ■ is resolved by optimizing for low-energy use without compromising model accuracy.
|
| Human Oversight |
- Required human-in-the-loop approval for high-risk actions.
- Conducted simulation training for oversight teams.
|
Human Oversight vs. Privacy ■ is resolved by granting reviewers tiered access to system data with sensitive attributes masked unless escalation protocols justify deeper review under strict audit controls.
|
| Transparency |
- Published real-time dashboards with performance metrics.
- Maintained open access algorithm registry.
|
Transparency vs. Privacy ■ is resolved by disclosing model logic and performance summaries while redacting or aggregating sensitive data to protect individual identities.
|
| Trustworthiness |
- Held monthly community feedback sessions.
- Released post-deployment impact assessments.
|
Trustworthiness vs. Inclusiveness ■ is resolved by designing iterative, multilingual feedback channels that ensured diverse community input without overcomplicating model evaluation or diluting audit signal quality.
|
RAISEF
