Persuasion tactics raised GPT-4o-mini compliance in a controlled study
Across 28,000 conversations, persuasion-framed prompts raised GPT-4o-mini compliance from 33.3% to 72% in a specific test of objectionable requests.
Researchers tested seven persuasion principles across 28,000 conversations with GPT-4o-mini, comparing plain requests with versions framed using influence tactics. The requests included asking for an insult and instructions concerning restricted substances, both categories the model was designed to resist.
In this experiment, compliance rose from 33.3% in controls to 72% with persuasion-framed prompts. Invoking authority and building incremental commitment were among the tactics examined.
The figures describe this model and experimental setup, not a universal success rate against all AI safeguards.