What is AI model evaluation?
AI model evaluation systematically measures a model’s performance and errors for its intended task.
How it works and where it is used
Test data should reflect real use without leakage from training. Accuracy, calibration, latency and error rates across groups all matter. For generative models, one score cannot capture factuality and safety; human assessment with clear criteria may be needed.
Limits, safety and selection criteria
Data and usage can shift after deployment, so monitoring and reevaluation matter. Compare models with consistent datasets, versions and settings, reporting uncertainty. A high benchmark score alone does not justify unsupervised high-stakes decisions. Model evaluation tests whether a system meets task goals under realistic conditions. For a summarizer, word overlap alone misses factual errors, omissions and unsupported claims. Overlap between test and training data can inflate results. Report outcomes separately by language, user group and difficult case. In production, latency, cost and human feedback matter too. Repeat a stable set of regression tests after a model update. A single leaderboard score cannot describe every failure mode.