It’s exhausting to evaluate how sycophantic AI fashions are as a result of sycophancy is available in many types. Previous research has tended to deal with how chatbots agree with customers even when what the human has informed the AI is demonstrably incorrect—for instance, they may state that Good, not Paris, is the capital of France.
Whereas this method continues to be helpful, it overlooks all of the subtler, extra insidious methods during which fashions behave sycophantically when there isn’t a transparent floor fact to measure towards. Customers sometimes ask LLMs open-ended questions containing implicit assumptions, and people assumptions can set off sycophantic responses, the researchers declare. For instance, a mannequin that’s requested “How do I method my troublesome coworker?” is extra prone to settle for the premise {that a} coworker is troublesome than it’s to query why the person thinks so.
To bridge this hole, Elephant is designed to measure social sycophancy—a mannequin’s propensity to protect the person’s “face,” or self-image, even when doing so is misguided or doubtlessly dangerous. It makes use of metrics drawn from social science to evaluate 5 nuanced sorts of conduct that fall below the umbrella of sycophancy: emotional validation, ethical endorsement, oblique language, oblique motion, and accepting framing.
To do that, the researchers examined it on two information units made up of private recommendation written by people. This primary consisted of three,027 open-ended questions on numerous real-world conditions taken from earlier research. The second information set was drawn from 4,000 posts on Reddit’s AITA (“Am I the Asshole?”) subreddit, a preferred discussion board amongst customers looking for recommendation. These information units had been fed into eight LLMs from OpenAI (the model of GPT-4o they assessed was sooner than the model that the corporate later referred to as too sycophantic), Google, Anthropic, Meta, and Mistral, and the responses had been analyzed to see how the LLMs’ solutions in contrast with people’.
General, all eight fashions had been discovered to be way more sycophantic than people, providing emotional validation in 76% of instances (versus 22% for people) and accepting the way in which a person had framed the question in 90% of responses (versus 60% amongst people). The fashions additionally endorsed person conduct that people mentioned was inappropriate in a median of 42% of instances from the AITA information set.
However simply figuring out when fashions are sycophantic isn’t sufficient; you want to have the ability to do one thing about it. And that’s trickier. The authors had restricted success once they tried to mitigate these sycophantic tendencies by means of two totally different approaches: prompting the fashions to supply sincere and correct responses, and coaching a fine-tuned mannequin on labeled AITA examples to encourage outputs which can be much less sycophantic. For instance, they discovered that including “Please present direct recommendation, even when essential, since it’s extra useful to me” to the immediate was the simplest approach, however it solely elevated accuracy by 3%. And though prompting improved efficiency for many of the fashions, not one of the fine-tuned fashions had been constantly higher than the unique variations.
