High-Risk AI Testing and Validation

Test high-risk AI against predefined performance and risk metrics before release and at appropriate lifecycle points.

For a high-risk AI provider, connect testing to the intended purpose and the risks the system is meant to control.

1. Set acceptance criteria and appropriate metrics before testing, including relevant thresholds and representative test conditions.
2. Test normal use, foreseeable misuse and material failure scenarios at suitable development stages and before release.
3. Record the system version, data, methods, results and limitations so the decision can be reviewed.
4. Assign owners to failures, retest corrections and document the release or restriction decision. Use separate approval processes for real-world testing where applicable. A successful test or a benchmark result is not, by itself, evidence of compliance with every linked requirement.

At appropriate development stages; before market placement or putting into service; after material changes; thereafter as defined by the quality process.

Approved test plan and predefined acceptance criteria Versioned test data, environment and execution records Results against thresholds, including failures and limitations Corrective actions and retest results Release, restriction or escalation decisions

Choose a test that influenced release. Check that its criteria existed before execution, its environment reflects intended use and failures were not simply removed from the report. Trace one correction to a completed retest.