Evals / AI evaluations
Definition
Structured tests used to measure how reliably an AI system consistently performs against defined tasks, requirements and failure scenarios.
Put simply
Instead of asking whether the AI “seems good”, evals test how reliably it performs and where it fails.
What boards and leaders should understand
AI testing needs to cover more than a handful of normal examples.
Depending on the consequence of the system, evals may include ambiguous situations, edge cases, adversarial requests, attempts to access confidential information, misleading inputs and situations where the AI should refuse or escalate.
And because models, inputs, data and connected systems can change, evaluation is not necessarily a one-off activity performed before launch.
Questions to consider
- What evidence gives us confidence that this system is reliable enough for the consequence of the work we're asking it to perform?
