AI models tested as judges of their own work, show limited self-preference

A researcher tested 10 AI models' ability to judge homework answers, including their own previously submitted work. The experiment used 90 programmatically generated questions with verified correct answers to create a pool of right and wrong responses. Most models did not show bias toward passing their own incorrect answers, with only two smaller OpenAI models exhibiting this self-preference. A key finding was that a model's grading reliability depended more on its own ability to solve the problem correctly than on bias. The test also assessed if factors like answer order, author attribution, or confident phrasing influenced grading.