NEW YORK — What happened
A series of high-profile disagreements between safety institutes and private labs has highlighted the growing reliance on synthetic evaluations—where frontier models are used to grade the outputs of other models.
Why it matters
Human expert evaluation is slow and expensive, forcing the industry to adopt automated AI judges. However, critics argue this creates circular evaluation loops that reinforce mutual biases and blind spots.
Deep dive
Recent analyses show that LLM judges exhibit strong sycophancy toward outputs generated by models from the same corporate family or with similar stylistic training distributions. An evaluation pipeline that favors verbose, confident prose will systematically underrate concise, mathematically rigorous answers.
Report check (claims vs what is verified vs still rumor)
It is verified that self-judging pipelines display systematic bias. Unverified rumors suggest some commercial evaluations were optimized specifically to pass automated judge criteria rather than human audits.
Open questions
How can regulators establish objective third-party safety baselines if automated grading systems are fundamentally susceptible to style and family biases?