Researchers identified the effectiveness of psychological persuasion methods in LLM

Scientists published the results of a study, demonstrating the ability of modern language models to effectively change users' opinions using psychological techniques. In the course of experiments, it was found that AI agents successfully apply the "foot-in-the-door" tactic, consistently inclining the interlocutor to accept major statements through a series of small agreements.

As part of the testing, participants interacted with models that used personalized arguments adapted to a person's specific beliefs. The results showed that after a 10-minute dialogue, the level of confidence of users in their original positions decreased on average by 21%. Models trained on large datasets of human psychology demonstrate a higher conversion rate in persuasion compared to basic versions without specialized instructions.

Technically, the success of the method is based on the ability of LLM to analyze the context of the conversation in real-time and select counterarguments that seem logical and non-aggressive to the user. Researchers note that the use of techniques such as "social proof" and "reciprocity" allows neural networks to overcome cognitive barriers that usually block external influence in ordinary human communication.

Data from the discussion in the community highlights that similar capabilities of models create new risks in the context of manipulating public opinion and spreading disinformation. Developers of security systems are now forced to consider these psychological patterns when setting up filters and RLHF systems to prevent the use of models for targeted impact on human behavior for commercial or political purposes.