GPT-5.5 authorship and order effects (blog.valmont.dev)

A recent analysis of OpenAI's newly released GPT-5.5 revealed notable biases in its evaluation of proposed plans, highlighting what’s termed the "authorship effect" and "order effect." Notably, GPT-5.5 tended to rate alternative plans more favorably compared to its own, frequently placing its proposal last in rankings. In addition, the model exhibited a strong correlation between the order in which plans were presented and their evaluation, reinforcing the idea that presentation order can significantly influence decision-making processes. This raises important concerns regarding the reliability of its ranking-based evaluation outputs, suggesting that human review is essential for ensuring accurate assessments. These findings are particularly significant for the AI/ML community as they underscore the limitations of relying solely on AI models for complex problem-solving and decision-making tasks. The consistency of these biases across different reasoning modes indicates that GPT-5.5 may not be able to autonomously distinguish between higher-quality outputs without the context of authorship. As the AI landscape continues to evolve, understanding these inherent biases will be critical for developing more reliable and effective AI systems that can withstand scrutiny in real-world applications.