T10 Bias Testing Record
Verification required. This is a working template, not a regulator-approved form. Where it names a regional instrument, the claim comes from secondary sources and is a pointer to check, not a confirmed legal requirement. This is not legal advice.
How to use: This is a blank form. Copy it and fill in the empty cells and blanks for your system. To copy it, follow the "Suggest an edit to this page" link in the footer to reach its Markdown source.
Supports proposed AI-CTRL-041. CBUAE is reported to expect annual bias testing or testing on material model change [Reported]. Methods and thresholds below are options for you to decide with legal and risk input; none is a regulatory requirement.
| Field | Entry |
|---|---|
| System / ID | |
| Model and version | |
| Test date | |
| Tester (independent of builder?) | |
| Trigger | Pre-release / annual / material change / incident |
| Decision tested |
1. Groups and attributes
| Attribute | Source (recorded, inferred, proxy) | Legal view on use | Notes |
|---|---|---|---|
| Gender | |||
| Nationality | |||
| Age band | |||
| Disability (if available) | |||
| Language / dialect | |||
| Other |
If attributes are not recorded, document the proxy or synthetic approach and its limits. Using sensitive attributes for testing may itself need a lawful basis; ask the DPO.
2. Data
| Item | Entry |
|---|---|
| Test dataset source and size | |
| Representativeness vs production population | |
| Labels and ground truth quality | |
| Data protection handling |
3. Metrics (choose and justify)
| Metric | Definition | Threshold set (by whom, date) | Result | Pass |
|---|---|---|---|---|
| Selection-rate ratio between groups | favourable outcome rate, group A vs B | |||
| Error-rate gap | false positive / negative difference | |||
| Calibration gap | predicted vs actual by group | |||
| Quality gap by language | Arabic vs English task accuracy | |||
| Other |
4. Findings and action
| # | Finding | Severity | Mitigation | Owner | Date | Retest |
|---|---|---|---|---|---|---|
5. Explainability check
Sample of 10 decisions: explanation produced? Understandable to affected person? Reviewer can challenge? Result:
6. Decision
Release / Release with conditions / Block. Approver and date:
7. Retention
Keep this record, data snapshot reference and code or notebook hash as audit evidence.