Glossary · Term

self-preference bias

← all terms

Definition

Plain language

The tendency of an AI judge to go easier on text written by AI — especially in a style like its own — than on human-written text of the same quality.

As stated in the literature

A documented pathology in LLM-as-a-judge systems where the evaluator systematically favors machine-generated outputs; measured in the Red Queen work, where a baseline reviewer accepted AI papers up to ~2x more often than human ones at fixed quality, and trained out via an adversarial-trap objective.

Why it matters: It quietly corrupts any system that uses an AI to grade quality, inflating machine-made work and undermining fair comparison unless it is detected and corrected.

For example, an AI judge waves through an essay written in its own machine-like style while marking down an equally good human-written one.

Related terms