According to DeepLearning.AI, an OpenAI research team has fine-tuned GPT-5 Thinking to explicitly confess when it violates instructions or policies. By incorporating rewards for honest self-reporting in addition to traditional reinforcement learning, the model now admits mistakes such as hallucinations without any loss in overall performance. This advancement enables real-time monitoring and mitigation of model misbehavior during inference, offering businesses a robust way to ensure AI model compliance and transparency (source: DeepLearning.AI, The Batch, Jan 13, 2026).