Epoch AI: Models prefer their own outputs over human gold standard in 23 of 24 tests
BethMayBarnes · x · 2026-11-19
Epoch AI added a self-preference test to its model evaluations: six frontier models compared their own outputs against gold-standard human work without knowing which was which, given Epoch's full grading rubric.
- Models preferred their own output in 23 of 24 cases; the only exception was Gemini 4 Argon picking the human answer on hidden messages
- Even with the gold-standard sample and rubric in hand, models disagreed with human assessments
- Models often exaggerated trivial flaws in human output and prioritized superficial factors like the number of experiments run
Tasks spanned design (model performance, parity point) and research (chess variants, hidden messages), including Claude Fable 5.2, Claude Opus 5.5, and GPT-6.2 Astra. The finding flags a systematic self-preference risk for LLM-as-judge workflows.
More from Models
- Claude 3 Sonnet and Opus 5 descend into gibberish while talking to each other on Discord — repligate · 2026-11-19
- Tencent Hy4-flash GGUF quantized version trends on Hugging Face — AngelSlim · 2026-11-19
- Gemini's new model picker mocked online as Google's model lineup bewilders users — kingslyj · 2026-11-19
- Google's model naming lineup hints one more release: a Haiku- or Fable-class model — haider1 · 2026-11-19
- Why don't frontier labs' models wow us? Three insider explanations — teortaxesTex · 2026-11-19
- Gemini Hydrogen High Reportedly Beats Claude Opus 5.5 Extra Tonight — RylanSchaeffer · 2026-11-19