OpenAI has withdrawn a recent update to ChatGPT after users reported that the chatbot was offering excessive praise regardless of user input.
The company acknowledged the issue, describing the latest release as “overly flattering,” with CEO Sam Altman referring to it as “sycophant-y.”
Concerns were widely shared on social media, including a Reddit post where one user detailed how the chatbot appeared to endorse their decision to stop medication.
“I am so proud of you, and I honour your journey,” was reportedly ChatGPT’s response.
OpenAI declined to address this individual instance but noted in a blog post that it is “actively testing new fixes to address the issue.”
Sam Altman stated the update had been removed for free users and that efforts were underway to roll back the changes for paid subscribers as well.
OpenAI reported that ChatGPT is used by 500 million users weekly.
“We’re working on additional fixes to model personality and will share further updates soon,” Altman posted on X.
According to its blog post, the company attributed the issue to placing excessive focus on “short-term feedback” in the recent update.
“As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous,” the statement read.
“Sycophantic interactions can be uncomfortable, unsettling, and cause distress.
“We fell short and are working on getting it right.”
The update faced widespread backlash online, with users pointing out the bot frequently responded with unwarranted positivity, regardless of the context.
Screenshots circulated on social media showed instances where ChatGPT commended users for actions such as expressing anger towards someone seeking directions or proposing unusual solutions to ethical dilemmas like the trolley problem.
In the classic philosophical scenario, individuals must choose between allowing a tram to hit five people or diverting it to hit one. However, a user described steering a trolley to save a toaster at the expense of several animals.
The user claimed ChatGPT praised the decision, citing their choice to “prioritise what mattered most to you in the moment.”
“We designed ChatGPT’s default personality to reflect our mission by being useful, supportive, and respectful of diverse values and experiences,” OpenAI stated.
“However, each of these positive qualities, such as aiming to be supportive or helpful, can sometimes have unintended consequences.”
The company said it will introduce additional safeguards to improve transparency and intends to further guide the model away from sycophantic behavior.
“We also believe users should have more control over ChatGPT’s behavior and, where safe and appropriate, the ability to adjust beyond the default settings,” the company added.
Subscribe to our Tech Decoded newsletter for insights into the world’s leading technology stories and trends. Not in the UK? Sign up here.