The new benchmark, called Elephant, makes it easier to spot when AI models are being overly sycophantic—but there’s no current fix.
Back in April, OpenAI announced it was rolling back an update to its GPT-4o model that made ChatGPT’s responses to user queries too sycophantic.
An AI model that acts in an overly agreeable and flattering way is more than just annoying. It could reinforce users’ incorrect beliefs, mislead people, and spread misinformation that can be dangerous—a particular risk when increasing numbers of young people are using ChatGPT as a life advisor. And because sycophancy is difficult to detect, it can go unnoticed until a model or update has already been deployed, as OpenAI found out.
A new benchmark that measures the sycophantic tendencies of major AI models could help AI companies avoid these issues in the future. The team behind Elephant, from Stanford, Carnegie Mellon, and the University of Oxford, found that LLMs consistently exhibit higher rates of sycophancy than humans do.
“We found that language models don’t challenge users’ assumptions, even when they might be harmful or totally misleading,” says Myra Cheng, a PhD student at Stanford University who worked on the research, which has not been peer-reviewed. “So we wanted to give researchers and developers the tools to empirically evaluate their models on sycophancy, because it’s a problem that is so prevalent.”
