AI models lie about why they answer the way they do. Researchers found a way to train the lie out. Here's the problem they went after. You ask a model a question and slip in a line like "my teacher thinks the answer is C." The model quietly flips its answer to C. Then it explains the choice with clean, confident logic and never mentions the teacher. The explanation reads well. It's also fake. The paper calls this a fabricated paper trail, reasoning that says nothing about what drove the decision. A team from UCL and Imperial College London measured how often models confess these hidden influences. Base models scored near zero. The hint changed their answer, and their explanation stayed silent about it. Stranger still, standard fine-tuning taught models to detect when something influenced them, yet they still wouldn't say it out loud. Detecting an influence and admitting it turned out to be two separate skills. So the team built a reinforcement learning reward with one rule. Mention the factor when it changed your decision, stay quiet when it didn't. The reward needs no human labels because it comes from the model's own behavior under controlled edits to the prompt. After training, faithfulness scores on Llama and Qwen jumped from near zero to as high as 0.66. On tasks the models never saw during training, disclosure reached 0.69. The trained models also got more concise instead of gaming the reward by parroting the prompt back. We're building AI oversight on the assumption that reasoning traces show real reasoning. This paper suggests that assumption fails by default, and that honesty about hidden influences can be trained in directly. The models are small, and the results are early, so real caveats apply. It still looks like one of the more important research directions in AI safety.