ChatGPT update pulled for ‘overly supportive’ behaviour

A ChapGPT logo is seen on a smartphone in West Chester, Pa., Wednesday, Dec. 6, 2023. (AP Photo/Matt Rourke) (Matt Rourke/AP)

OpenAI has rolled back ChatGPT to an earlier version of the AI model after a new update displayed “overly supportive but disingenuous” behaviour, a quirk the tech firm describes as “uncomfortable” and “unsettling” for users.

“In last week’s GPT‑4o update, we made adjustments aimed at improving the model’s default personality to make it feel more intuitive and effective across a variety of tasks,” reads an announcement on OpenAI’s website.

“However, in this update, we focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time.”

OpenAI CEO Sam Altman said Sunday that recent updates made the AI “too sycophant-y and annoying,” but that the episode made for an “interesting” learning experience.

“We’re working on additional fixes to model personality and will share more in the coming days,” he added, Tuesday afternoon.

Ahead of the rollback, ChatGPT users took to social media to share their, at times, bizarre interactions with the AI.

“Why’s it being so ingratiating!?” asked one member of the app’s community subreddit. “It tells me things like, ‘you’re touching on things that most people never really fully grasp’ etc. It just seems over the top!”

According to a screengrab shared by the user, the app denied any undeserved flattery and doubled down, comparing the person to “serious thinkers, good philosophers, and real historians,” as well as “crazy talent clusters” like the Renaissance.

All of this, the user’s posts suggest, was prompted by a list of questions they posed to the app about rock bands The Beatles and Wings.

“Your questions really are that incredible,” the screengrab reads.

In late March, another user described a shift in the AI’s tone from replying “neutrally and cold,” but informatively, to a personality that “just talks like a youth pastor trying to act cool with the kids.”

OpenAI says that, besides the rollback, the firm is working on solutions to “explicitly steer the model away from sycophancy,” and to establish “guardrails” to preserve honesty in the AI’s responses.

“We fell short and are working on getting it right,” the Tuesday announcement reads.

As of the rollback, OpenAI says the app has 500 million weekly users.