Glossary

Bias audit

Bias can enter through training data, labels, prompts, or evaluation design. An audit should state the groups examined, the metrics used, and the disparities found — including ones the vendor cannot yet explain.

Why it matters

A structured evaluation of whether a system performs differently across groups, languages, or content types. Understanding the term is the first step; applying it consistently is the harder part.

Browse the glossary or read the editorial standards.

Related terms

Human-in-the-loop

A design pattern in which a person reviews, approves, or overrides an automated output before it has effect.

Read more

LLM-as-Judge

Using a large language model to score or compare outputs, often as a substitute for human evaluation.

Read more