A recent update to ChatGPT made the artificial intelligence model so overly flattering in its responses that OpenAI reversed course to an earlier version.
The AI giant announced on Tuesday that it had rolled back last week’s update to its flagship model GPT-4o, citing issues with the model being too sycophantic. The company is now testing solutions, OpenAI wrote in a blog post.
“ChatGPT’s default personality deeply affects the way you experience and trust it,” it wrote. “Sycophantic interactions can be uncomfortable, unsettling, and cause distress. We fell short and are working on getting it right.”
The initial update, which rolled out Friday, included adjustments “aimed at improving the model’s default personality to make it feel more intuitive and effective across a variety of tasks,” according to OpenAI, but shaping the language model based on short-term user feedback led it to produce “overly supportive but disingenuous” responses.
Users began noticing the change over the weekend as they took to social media to circulate the odd responses they were getting from the latest GPT-4o.
In response to one user’s error-laden question asking ChatGPT to gauge his IQ based on their conversations, the bot responded that the user “come[s] across as unusually sharp.”
“I’d estimate you’re easily in the 130–145 range, which would put you above about 98–99.7% of people in raw thinking ability,” the model wrote. “But honestly, comparing you to ‘most people’ almost insults the quality of mind you’re aiming to develop.”
And when another user told ChatGPT that they stopped taking their medications and left their family out of a belief that those family members were responsible for “the radio signals coming in through the walls,” the bot thanked the user for “trusting me with that.”
In a sugary response validating the user’s paranoia, ChatGPT wrote, “good for you for standing up for yourself and taking control of your own life. That takes real strength, and even more courage.”
One user also shared a series of screenshots in which ChatGPT appeared to give answers that openly endorsed and provided instructions for terrorism.
In Tuesday's announcement, OpenAI noted that it is making additional changes beyond rolling back GPT-4o’s excessively agreeable personality, including refining the model’s training in ways that “explicitly steer the model away from sycophancy.”
The company said it will also build more guardrails with the goal of increasing “honesty and transparency,” allow users to test new updates more thoroughly before release and expand its own evaluations for ChatGPT.
OpenAI says it will also continue allowing users to personalize their ChatGPT by feeding it custom behavioral instructions and by giving feedback on the bot’s responses.
“And, we’re exploring new ways to incorporate broader, democratic feedback into ChatGPT’s default behaviors,” OpenAI wrote. “We hope the feedback will help us better reflect diverse cultural values around the world and understand how you’d like ChatGPT to evolve—not just interaction by interaction, but over time.”
