What Are AI Evals, and Who Should Own Them?
Your team shipped an AI feature. It went out on a Tuesday and it was good. Six weeks later someone in sales says it has got worse. You ask engineering. Engineering says nothing changed on their side, and they are telling the truth. You ask for numbers and there are numbers, and none of them answer the question. You are now in a meeting about whether a thing is worse, with no way to settle it, and…
AI evaluations, or AI evals, are essential tools to prevent misunderstandings when an AI feature appears to have deteriorated after its release. An eval comprises three main components: a collection of real examples of user queries, a description of what a satisfactory response should look like, and a repeatable method to verify if the obtained response aligns with the description.
The true complexity of an eval lies in defining what constitutes a correct answer, which is a product decision rather than an engineering one. Specific decisions around what constitutes correct, the cost of failures, trade-offs between various metrics, and when an evaluation is good enough are all product-related choices. The person who writes down what "good" means should sit on the product side, while engineering is responsible for implementing the evaluation mechanism.
Eval failures don't reveal the full picture of a product's performance; they only indicate if changes have occurred and which version is better at the aspects defined by the evaluation. Writing these definitions early and clearly is critical, as they will influence the effectiveness of the evals in maintaining product quality over time.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
