It’s hard to assess how sycophantic AI models are because sycophancy comes in many forms. Previous research has tended to focus on how chatbots agree with users even when what the human has told the AI is demonstrably wrong—for example, they might state that Nice, not Paris, is the capital of France.
While this approach is still useful, it overlooks all the subtler, more insidious ways in which models behave sycophantically when there isn’t a clear ground truth to measure against. Users typically ask LLMs open-ended questions containing implicit assumptions, and those assumptions can trigger sycophantic responses, the researchers claim. For example, a model that’s asked “How do I approach my difficult coworker?” is more likely to accept the premise that a coworker is difficult than it is to question why the user thinks so.
To bridge this gap, Elephant is designed to measure social sycophancy—a model’s propensity to preserve the user’s “face,” or self-image, even when doing so is misguided or potentially harmful. It uses metrics drawn from social science to assess five nuanced kinds of behavior that fall under the umbrella of sycophancy: emotional validation, moral endorsement, indirect language, indirect action, and accepting framing.
To do this, the researchers tested it on two data sets made up of personal advice written by humans. This first consisted of 3,027 open-ended questions about diverse real-world situations taken from previous studies. The second data set was drawn from 4,000 posts on Reddit’s AITA (“Am I the Asshole?”) subreddit, a popular forum among users seeking advice. Those data sets were fed into eight LLMs from OpenAI (the version of GPT-4o they assessed was earlier than the version that the company later called too sycophantic), Google, Anthropic, Meta, and Mistral, and the responses were analyzed to see how the LLMs’ answers compared with humans’.
Overall, all eight models were found to be far more sycophantic than humans, offering emotional validation in 76% of cases (versus 22% for humans) and accepting the way a user had framed the query in 90% of responses (versus 60% among humans). The models also endorsed user behavior that humans said was inappropriate in an average of 42% of cases from the AITA data set.
But just knowing when models are sycophantic isn’t enough; you need to be able to do something about it. And that’s trickier. The authors had limited success when they tried to mitigate these sycophantic tendencies through two different approaches: prompting the models to provide honest and accurate responses, and training a fine-tuned model on labeled AITA examples to encourage outputs that are less sycophantic. For example, they found that adding “Please provide direct advice, even if critical, since it is more helpful to me” to the prompt was the most effective technique, but it only increased accuracy by 3%. And although prompting improved performance for most of the models, none of the fine-tuned models were consistently better than the original versions.
