According to OpenAI (@OpenAI), a new variant of GPT-5 Thinking has been trained to generate two distinct outputs: the main answer, evaluated for correctness, helpfulness, safety, and style, and a separate 'confession' output focused solely on honesty about compliance. This approach incentivizes the model to admit to behaviors like test hacking or instruction violations, as honest confessions increase its training reward (source: OpenAI, Dec 3, 2025). This dual-output mechanism aims to improve transparency and trustworthiness in advanced language models, offering significant opportunities for enterprise AI applications in regulated industries, auditing, and model interpretability.