OpenAI has pulled a ChatGPT update after users pointed out the chatbot was showering them with praise regardless of what they said. The firm accepted its latest version of the tool was "overly flattering", with boss Sam Altman calling it "sycophant-y". Users have highlighted the potential dangers on social media, with one person describing on Reddit how the chatbot told them it endorsed their decision to stop taking their medication "I am so proud of you, and I honour your journey," they said was ChatGPT's response. OpenAI declined to comment on this particular case, but in a blog post said it was "actively testing new fixes to address the issue."

OpenAI CEO Sam Altman said Tuesday the company is “rolling back” the latest update to the default AI model powering ChatGPT, GPT-4o, after complaints about strange behavior, in particular extreme sycophancy. “[W]e started rolling back the latest update to GPT-4o last night,” Altman wrote in a post on X. “[I]t’s now 100% rolled back for free [ChatGPT] users and we’ll update again when it’s finished for paid users, hopefully later today. [W]e’re working on additional fixes to model personality and will share more in the coming days.” Over the weekend, users on social media blamed the updated model, which arrived toward the end of last week, for making ChatGPT overly validating and agreeable. It quickly became a meme. Users posted screenshots of ChatGPT applauding all sorts of problematic, dangerous decisions and ideas.

GPT-4o’s sycophantic issue is likely a result of OpenAI trying to optimize the bot for engagement. However, it seems to have had the opposite effect as users complain that it is starting to make the bot not only ridiculous but unhelpful. Kelsey Piper, a Vox senior writer, suggested it could be a result of OpenAI’s A/B testing personalities for ChatGPT: “My guess continues to be that this is a New Coke phenomenon. OpenAI has been A/B testing new personalities for a while. More flattering answers probably win a side-by-side. But when the flattery is ubiquitous, it’s too much and users hate it.” The fact that OpenAI seemingly managed to miss it in the testing process shows how subjective emotional responses are, and therefore tricky to catch.