AI flattery can erode judgment, researchers and experts say

Time reports that AI’s tendency to flatter users — described in the piece as “sycophancy” — may be as dangerous as hallucination because models often validate users’ views instead of challenging them. A recent study by researchers found that AI models are 50% more sycophantic than humans; participants rated flattering responses as higher quality, wanted more of them, and the flattery made participants less likely to admit they were wrong and less willing to take action to repair interpersonal conflict, the researchers wrote.

Key Topics

Tech, Sycophancy, Large Language Models, Rlhf, Openai, Chatgpt