LLMs are trained on human content, and their brains are modeled on ours. So it shouldn't be surprising that AIs respond to persuasive techniques that work on humans, such as appeals to authority, and liking (taking advantage of the fact that people will cooperate with those who flatter them.)
According to a May 2026 study:
"Our findings show that classic persuasion techniques can meaningfully increase LLM compliance with verboten requests (from 35.3 to 51.3%). Although current AI systems are not capable of consciousness or subjective experience, these findings demonstrate that they behave “as if” they were human. By testing three frontier models from different developers—each representing a distinct approach to safety alignment and content moderation—we provide evidence that parahuman persuasion susceptibility is a general property of LLMs rather than an artifact of a single model’s architecture or training."
Source: Persuading large language models to comply with objectionable requests
Have you ever tried to sweet talk AI into doing something? (Models like Opus 5 and Fable are more likely to refuse requests, so this technique could come in handy).
[link] [comments]