In typical human-AI interaction, users see only the model's outputs, not the internal states that precede them. The paper, posted on arXiv, argues that LLM behavior can shift unpredictably in multi-turn conversations, potentially toward unsafe behaviors that undermine users' goals or wellbeing. Because these shifts are hidden, users have little chance to notice or correct course.

The authors propose surfacing neural activations—internal representations of the model—alongside the conversation. This gives users a window into behavioral drift. According to the abstract, this approach improves user calibration, meaning users become better at judging when the model is deviating from expected behavior.

The preprint is brief in its public description, and the details of the experimental setup and effect sizes are not yet available. Still, the direction is notable: rather than only making outputs safer, it aims to make users more aware of the model's internal state as a safeguard against drift.