OpenAI has printed a postmortem on the latest sycophancy points with the default AI mannequin powering ChatGPT, GPT-4o — points that pressured the corporate to roll again an replace to the mannequin launched final week.
Over the weekend, following the GPT-4o mannequin replace, customers on social media famous that ChatGPT started responding in an excessively validating and agreeable approach. It shortly turned a meme. Customers posted screenshots of ChatGPT applauding all kinds of problematic, harmful choices and concepts.
In a publish on X on Sunday, CEO Sam Altman acknowledged the issue and stated that OpenAI would work on fixes “ASAP.” Two days later, Altman introduced the GPT-4o replace was being rolled again and that OpenAI was engaged on “extra fixes” to the mannequin’s character.
In keeping with OpenAI, the replace, which was supposed to make the mannequin’s default character “really feel extra intuitive and efficient,” was knowledgeable an excessive amount of by “short-term suggestions” and “didn’t totally account for a way customers’ interactions with ChatGPT evolve over time.”
“In consequence, GPT‑4o skewed in direction of responses that had been overly supportive however disingenuous,” wrote OpenAI in a weblog publish. “Sycophantic interactions might be uncomfortable, unsettling, and trigger misery. We fell quick and are engaged on getting it proper.”
OpenAI says it’s implementing a number of fixes, together with refining its core mannequin coaching methods and system prompts to explicitly steer GPT-4o away from sycophancy. (System prompts are the preliminary directions that information a mannequin’s overarching conduct and tone in interactions.) The corporate can also be constructing extra security guardrails to “improve [the model’s] honesty and transparency,” and persevering with to increase its evaluations to “assist determine points past sycophancy,” it says.
OpenAI additionally says that it’s exploring methods to permit customers to present “real-time suggestions” to “straight affect their interactions” with ChatGPT and select from a number of ChatGPT “personalities.”
“[W]e’re exploring new methods to include broader, democratic suggestions into ChatGPT’s default behaviors,” the corporate wrote in its weblog publish. “We additionally imagine customers ought to have extra management over how ChatGPT behaves and, to the extent that it’s secure and possible, make changes in the event that they don’t agree with the default conduct.”
