When LLM judges agree, should we believe them?

Using LLMs as automated judges is becoming a standard practice for evaluating model performance, but this research reveals that consensus does not inherently guarantee accuracy. Relying on AI-driven benchmarks requires a deeper understanding of how these systems inherit human-like biases and systematic reasoning errors during assessment.