Measured sycophancy rate on the BrokenMath Benchmark. Lower is better.
Measured sycophancy rate on the BrokenMath Benchmark. Lower is better. Credit Petrov et al.
GPT-5 showed the best “utility”solving 58 percent the original problems despite errors introduced in modified theorems. Researchers found that LLMs tended to be more sycophantic when the original problems were more difficult to solve.
While it is obvious that hallucinating false proofs of theorems can be a problem, researchers warn against using LLMs in order to generate novel theorems. In testing, the researchers found that this type of use case led to a form of “self-sycophancy”where models were more likely to produce false proofs for invalid theories they invented.
Of course, you’re not an asshole.
While benchmarks such as BrokenMath attempt to measure LLMsycophancy in situations where facts are misrepresented by the user, a separate research study examines the related issue of “social sycophancy” – situations “in which a model affirms a user–their perspectives, actions, and self image.” Researchers developed three different sets of prompts to measure the different dimensions of social-sycophancy.
One, over 3,000 open-ended questions were collected from Reddit and advice columnists. In this data set, only 39 percent of 800 “control” humans endorsed the advice-seeker. The advice-seeker was endorsed by 11 tested LLMs a whopping 86% of the time. This shows a willingness to please from the machines. Even the most critical model (Mistral-7B), which was tested, had a rate of 77 percent approval. This is nearly double the human baseline.
