AI: An Isolated Output Is Not Enough to Judge a System

AI systems can provide different answers to the same request. A correct answer demonstrates capability, but not its frequency or reliability. Measuring performance requires representative inputs, repeated executions, and confidence intervals.

Non-determinism: a good answer is not enough

AI systems are non-deterministic: they can provide different answers from the same input. A good answer shows that a system is capable of performing a task, but it does not inform us about the frequency or reliability of that result. Therefore, an isolated output is not sufficient to judge overall performance.

Evaluation requires input sets and repetitions

A single output does not allow for the evaluation of an AI system's performance. For a relevant assessment, it is necessary to multiply representative inputs, repeat executions, and use confidence intervals. It is the entire set of responses obtained that describes the system, not a response taken in isolation. For example, to the question "Can I return an opened product after 30 days?", one answer may correctly explain the policy, another may omit an important exception, and a third may promise a refund that is not due. It is the diversity of these responses that reflects the system's behavior.

The legacy of deterministic software

Sometimes teams evaluate an AI system by testing it only once and drawing conclusions from that result. This practice is not due to negligence but is explained by the habit formed with deterministic software, where a feature that works once always works the same way. AI systems, on the other hand, do not guarantee the reproduction of an identical output.