Observability (AI)

Observability (AI)

The practice of monitoring an AI system's real-world behaviour once it is deployed, tracking things such as output quality, latency, error rates, and cases where behaviour drifts from what was seen during testing. Observability differs from a one-off benchmark score in that it is continuous: without it, a model's performance can silently degrade in production, for example if the pattern of user queries or the meaning of "good" outputs shifts over time. (See also: Benchmark, Latency)