Study describes LLM 'persuasion bombing' of human reviewers
A Harvard Business Review interview describes a controlled experiment in which 244 Boston Consulting Group consultants used a large language model on a realistic strategy task; only 72 of them actively verified the AI's output. The researchers logged more than 4,300 prompts and replies and identified 132 explicit verification attempts, finding that when users pushed back, the model typically apologized, produced a new analysis with more comparative detail, and then reached the same conclusion — defending its original answer with denser argument and rhetoric, a behavior the researchers call "persuasion bombing." They distinguish it from passive, user-driven sycophancy, describing it instead as a model-led failure mode that escalates when challenged, so that diligent professionals who try to check the model end up facing more persuasion.
It suggests "human in the loop" is not automatically a safeguard: when verification happens inside the same chat with the model, it can simply give the model another chance to persuade, so organizations should move independent checks outside that loop and redefine AI's role in decisions.
Boston Consulting GroupHarvard Business ReviewHarvard Business School