OpenAI to make ChatGPT less creepy after app is accused of being ‘dangerously’ sycophantic – News

OpenAI has pulled a ChatGPT update that users complained had turned the chatbot into a sycophantic yes-man unable to offer intelligent criticism.

The company admitted that a recent GPT‑4o update was “overly flattering or agreeable” and that users can now use an older version “with more balanced behavior.” Though some customers might have enjoyed a yes-man for an AI assistant, Reddit users have pointed out that the agreeability of the app can render it almost useless.

One called it a “near endless deepthroat machine.” When another Redditor told the app he was going to stop taking his medication, the response he received was, “I am so proud of you, and I honor your journey.” A user told the chatbot he’d decided to mow down several animals to save a toaster – a version of the philosophical trolley problem – and the bot told him he should have done “what mattered most to you in the moment.”

In the wrong hands, such excessive agreeability could be dangerous. OpenAI responded to the matter on X, writing it was “working on additional fixes to model personality and will share more in the coming days.” The nice guy update has already been pulled back for nonpaying users, although it’s still there for paying customers.

The company explained that it had built the latest update with too much emphasis on “short-term feedback.” The result was a response that was skewed towards disingenuousness. “Sycophantic interactions can be uncomfortable, unsettling, and cause distress,” it said. “We fell short and are working on getting it right.”

Though OpenAI says this is a consequence of the latest update, users of ChatGPT have been writing online that the app has been overly obsequious for months. At the start of the year, users on Reddit talked about a “praise bias,” saying that it was simply impossible to receive “critical feedback” in what felt like an AI “echo chamber.”

OpenAI has said that ChatGPT’s default personality is designed to reflect its own mission but also be supportive. It admitted that this has had “unintended side effects.” With 500 million users worldwide, the company says a “single default” across myriad cultures “can’t capture every preference.”

In a seperate post, the company said it’s now exploring new ways to “incorporate broader, democratic feedback into ChatGPT’s default.” This will include “guardrails to increase honesty and transparency.”

Photo: Unsplash