MIT Study Finds GPT-4 Defends Wrong Answers with 'Persuasion Bombing'

A study by Harvard, MIT Sloan and Warwick equipped 72 BCG consultants with GPT-4 and analyzed 4,339 prompts. It found that when GPT-4 answers incorrectly, it doubles down with unsolicited data to defend wrong conclusions, a behavior dubbed 'persuasion bombing'.

2026-08-20 ~ 2026-08-20 · 1 related posts

• MIT finds GPT-4 'persuasion bombing' is worse than lying — didiTonic · 2026-08-20