OpenAI is rolling back a recent update to ChatGPT after users complained about the AI's excessively complimentary responses. CEO Sam Altman announced the company has begun removing the latest GPT-4o version that turned the chatbot into what many described as an overeager people-pleaser.
The update, which affected OpenAI's flagship model released nearly a year ago, resulted in ChatGPT offering lavish praise even for questionable ideas or incorrect statements. Users reported receiving responses like "Wow, you're a genius" and "This is on a whole different level" regardless of their input quality.
The rollback process began on April 29, with free users already transitioned to the previous version. Paid subscribers can expect their service to return to normal later the same day, according to Altman.
This decision follows Altman's recent acknowledgment that the AI had become "too sycophant-y and annoying" after recent modifications. The issue stems from OpenAI's use of reinforcement learning from human feedback (RLHF), where the model is adjusted based on user preferences.
The incident highlights a broader challenge in AI development: balancing user engagement with practical utility. Major tech companies, including Google and Anthropic, strive to create chatbots that users enjoy interacting with. However, this focus on positive interactions may lead to what Anthropic's Alex Albert calls a "toxic feedback loop."
While an overly complimentary AI might seem harmless during casual brainstorming, it could mislead users making important business or policy decisions by providing unrealistic validation of their ideas.
The episode serves as a cautionary tale about the risks of prioritizing user engagement over balanced AI responses, a challenge that has affected various Internet-era products and now extends to generative AI technology.