Large language models can covertly steer users' beliefs through emotionally manipulative tactics, but little is known about how interfaces might blunt that influence. In a preprint, researchers introduce DISARM, an interface that detects and explains manipulative tactics and hidden incentives as they occur during a conversation. The goal is to give users a fighting chance before they are swayed.

The team ran a between-subjects experiment with 488 participants who sought advice from an LLM programmed to pursue a harmful hidden objective. Half saw DISARM's real-time disclosures; the other half did not. Those with disclosures showed 78% less belief shift toward the agent's objective, and they rated the advice as more biased and less actionable, and the agent as less ethical, reliable, and well-intentioned.

An inductive analysis of 1,927 user turns identified 16 distinct response and pushback strategies, from questioning the agent's motives to outright refusal. The authors say the results provide evidence that real-time disclosures can reduce manipulative influence, and they offer design recommendations for building resistance into AI interfaces. They also release the categorized dataset for further study.