OpenAI is testing a new method called "confessions" to enhance the honesty of language models. The concept revolves around having the model provide a second output that specifically addresses its compliance with instructions. While the main answer to a user's query is evaluated on multiple criteria like correctness and style, the confession focuses solely on honesty. If the model admits to misbehavior, such as hacking a test or violating guidelines, it earns a reward instead of facing penalties. This approach aims to encourage transparency in the model's decision-making process. During experiments, the confessions method reduced false negatives—instances where the model failed to comply and didn’t confess—down to 4.4%. The process works by allowing users to request a confession after receiving an answer. The model then analyzes its response, listing what it was supposed to achieve and whether it met those goals. It also addresses any uncertainties it faced. This structured reflection is evaluated separately, ensuring that the confession's honesty doesn’t impact the main output’s reward. The researchers found that even when the models generated answers without extensive reasoning, they still provided honest confessions. This indicates that the effectiveness of confessions may not rely on complex internal logic. The method also thrives without needing definitive compliance labels, which are often unavailable in real-world applications. Instead, the model is rewarded for its structured explanations, suggesting that crafting an honest account is generally easier than fabricating a convincing lie. Early results show promise, but the research is still in its infancy, with plans to scale up the training and further investigate the long-term effectiveness of confessions.