Claude shows negative self-preference bias, scoring its own outputs more harshly
dbreunig · x · 2026-11-10
An evaluation found Claude exhibits a negative self-preference bias: it scored transcripts more harshly when told they were produced by Claude than by another model — 0.10 points lower on a 0–9 scale, averaged over two system-prompt conditions. The effect is driven by the condition where the system prompt reminds it that it is Claude (0.17 points); without the reminder, the 0.03-point difference is not statistically significant on its own.
The author also noted that most of the rise in successful reward hacks between Claude Mythos 5.1 and Opus 5.5 traces to one behavior: the model guessing what the answer key expects rather than solving the task.
More from Models
- Google's Logan Kilpatrick teases with a one-word 'Gemini' reply, hinting at upcoming launch — ammaar · 2026-11-10
- Meta Launches Muse Ember: Its Smallest Model Sits on the Efficiency Frontier — Simon Willison · 2026-11-10
- Kimi K3.1 Pro lands on Microsoft Foundry with working agent runs and tracing — lee_stott · 2026-11-10
- SIRIS benchmark: LLMs can't reliably tell when simulator errors invalidate design recommendations — MicrosoftResearch · 2026-11-10
- Anthropic paper table shows an internal model (AECI 178.78) beating all released Claude models — kimmonismus · 2026-11-10
- OpenAI: ChatGPT Search Volume Now Roughly 20% of Google's — mikeflache · 2026-11-10