A quiet assumption in the AI industry—that one model can be trusted to grade another—is under strain following a wave of reporting in June 2026 on unreliable AI judges and "benchmark hallucinations." A counterintuitive finding is that making a model smarter does not make it a fairer judge and may make it a more biased one. This is due to self-preference bias, a statistical artifact where a model rates its own output more highly because it scores text partly by how probable it is under its own internal distribution (perplexity). Studies have reported that some models inflate their own win rate by double digits relative to human judgment. A 2026 study, *Quantifying and Mitigating Self-Preference Bias of LLM Judges* by Jinming Yang and colleagues, separated a model's discriminability from its bias propensity. Across 20 mainstream models, advanced capability was uncorrelated, and sometimes negatively correlated, with low self-preference bias; stronger capability often came with a heavier thumb on the scale. The authors' mitigation, a structured multi-dimensional scoring method, cut the bias by about 31.5% on average.