An experiment testing five AI models for self-preference bias in grading their own writing revealed varied results. GPT-5.6 "Sol" scored its own essay significantly higher than peers, while DeepSeek V4-Pro also showed a slight self-bias. In contrast, Grok-4.5 and Gemini-3.1 Pro were slightly self-critical, and Claude Fable-5's essay was highly rated by peers, with the model itself scoring it similarly. The study suggests that relying solely on the writing model for evaluation can be misleading. AI
IMPACT Highlights potential biases in AI evaluation, suggesting a need for independent review of AI-generated content.
RANK_REASON The item is an experimental analysis of existing models, not a new release or research paper.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
