AI Bias Assessment Template

AI BIAS AND IMPACT ASSESSMENT WORKSHEET

INTERNAL WORKSHEET — NOT A CERTIFICATION OR DEPLOYMENT APPROVAL

1. Assessment Control

2. System and Decision Context

3. Current-Authority and Obligation Register

Counsel must complete one row for every potentially applicable jurisdiction, sector, contract, program, decision type, and population.

This worksheet does not state that any named legal category, test, threshold, notice, audit, retention period, appeal, or remedy applies.

4. People and Impact Mapping

Include direct subjects, applicants, users, workers, customers, nonusers, bystanders, reviewers, operators, and people affected by feedback loops.

5. Data, Labels, and Representation

6. Evaluation Design

A missing difference does not prove fairness. A measured difference does not by itself identify its cause or legal significance.

7. Results Register

Attach code, data version, queries, test output, and reproducibility instructions.

8. Cause and Control Analysis

Potential controls to assess include data changes, label review, feature restrictions, model changes, thresholds, uncertainty handling, human review, documentation, user interface, access limits, appeal or correction processes, monitoring, and retirement.

9. Human Review and Contestability

Do not claim that human involvement cures a system risk without testing the actual workflow.

10. Monitoring and Change Control

Reassessment triggers: purpose, population, geography, data, label, model, threshold, vendor, interface, human workflow, law, contract, complaint pattern, incident, or material performance change.

11. Decision Record

Scope and conditions: [________________________________]

Known limitations: [________________________________]

Residual risks and rationale: [________________________________]

Required notices, documentation, or contractual controls: [________________________________]

Next review date / trigger: [________________________________]

12. Final Review Gate

☐ Intended and prohibited uses are documented.

☐ Affected people and harm scenarios were reviewed.

☐ Current legal and contractual obligations were entered from official sources.

☐ Data, labels, representation, proxies, and lineage were assessed.

☐ Evaluation slices, measures, uncertainty, and reproducibility are documented.

☐ Mitigations were tested rather than merely proposed.

☐ Human review, correction, complaint, and accessibility processes were evaluated.

☐ Monitoring, escalation, change, rollback, and reassessment controls are assigned.

☐ No unsupported fairness, compliance, certification, safety, or legal conclusion remains.

Product owner: [________________________________]

Data-science reviewer: [________________________________]

Legal reviewer: [________________________________]

Domain / affected-community reviewer: [________________________________]

Approving executive: [________________________________]

Approval date: [__/__/____]