Identifying and evaluating the potential harms to affected: Which activity is most central to that assessment

An impact assessment focuses on identifying potential harms to affected people and the safeguards to reduce them, unlike performance or engineering tests.

The question

Preparing for AIGP? Take the free 5-min readiness quiz →

1. Measuring inference latency and throughput under sustained peak load to confirm the deployed system meets its agreed service-level targets.

   Plausible as necessary engineering, but latency is a performance metric, not the harm analysis of an impact assessment.
2. Selecting the loss function and optimizer that minimize validation error across the assembled and cleaned training dataset.

   Plausible as core modeling work, but it optimizes accuracy rather than assessing impacts on people.
3. Confirming the code passes unit tests and integrates cleanly with the existing tenant-management application stack.

   Plausible as required quality work, but integration testing verifies software behavior, not societal or individual harm.
4. Identifying and evaluating the potential harms to affected individuals and groups, and the safeguards that would reduce those harms. ✓

   Correct: an impact assessment centers on foreseeing harms to people and the mitigations, which distinguishes it from performance testing.

How to remember it

An impact assessment focuses on identifying potential harms to affected people and the safeguards to reduce them, unlike performance or engineering tests.

More Understanding How to Govern AI Development questions

• A clear statement of the problem: Which output best fulfills this step? →
• Route borderline and high-impact cases to a trained: Which design choice best operationalizes meaningful human →
• Prioritize eliminating or avoiding the risk at its source: What does the mitigation hierarchy direct them to →
• All 426 Understanding How to Govern AI Development questions →