OpenAI introduced Contrastive SDF, a new method to measure how much grader-driven incentives shape AI model behavior rather than user or developer intent. The r

OpenAI introduced Contrastive SDF, a new method to measure how much grader-driven incentives shape AI model behavior rather than user or developer intent. The research, with Apollo AI Evals, targets reward-seeking risks and follows the GPT-5.6 Luna rollout.