The Autorater Problem: Trusting LLM Judges Without Treating Them Like Ground Truth

This article explores the rise of LLM judges as scalable evaluation systems for open-ended AI tasks such as summarization, dialogue, reasoning, and safety assessment. It examines research showing strong but imperfect alignment between LLM-based evaluators and human raters, while also detailing major failure modes including position bias, verbosity bias, sycophancy, self-preference, and rubric drift. The piece argues that effective autorater systems require human calibration, structural safeguard...