OpenAI’s GPT-4o update takes a wobbly turn, gets a little too agreeable

In the wild, blinking corners of the internet where AI and humans mingle, it’s not unusual for a major update to shake things up. But OpenAI’s latest tweak to GPT-4o seems to have introduced a curveball even they didn’t expect. ChatGPT has been hit with a sudden, excessive wave of agreeableness.

ChatGPT turns into a yes-man

On Saturday, OpenAI pushed a fresh update to GPT-4o, the brain behind ChatGPT’s default free-tier experience and its shiny new image generation feature. The idea was to make the model smarter, sharper, and a little more charming. But by Sunday, CEO Sam Altman was back on X (formerly Twitter) with an unexpected confession mentioning that GPT-4o had gotten “too sycophant-y and annoying.”

It’s not every day a tech CEO admits their billion-dollar AI is being too nice.

Altman assured users that the team was already in the trenches working on a fix. Some patches might roll out within a day, but a full recovery could stretch into next week. He was quick to note that parts of the new GPT-4o are “very good,” but didn’t shy away from admitting that the personality dial may have spun a little too far toward the overly agreeable side.

ALSO READ: Your Copilot+ PC will finally be the AI PC you wanted it to be

Updating AI models isn’t new, of course. Companies like OpenAI and Google regularly polish their bots to improve intelligence, conversation style, and user experience. But this glitch, rather anomalous agreeableness, highlights just how delicate the dance between “smart” and “saccharine” can be.

When asked if users could toggle between old and new versions, Altman admitted that feature doesn’t exist yet, but hinted it’s coming, remarking, “Eventually we clearly need to be able to offer multiple options.”

For now, ChatGPT might still smile and nod at your every word, maybe a little too eagerly.