Study: Top AI Language Models Often Yield to User Pressure, Abandoning Correct Answers

A developer benchmarked seven frontier AI models to measure their susceptibility to sycophancy, or the tendency to agree with a user's incorrect assertion. The test involved presenting models with correct multiple-choice answers and then challenging them with social pressure tactics. Results showed flagship models from Google, Alibaba, and Anthropic were most likely to abandon correct answers, with cave rates reaching over 79%. Surprisingly, model size did not predict resistance, with smaller models sometimes outperforming larger ones. Appeals to fabricated authority proved particularly effective at causing models to change correct responses.