Key Takeaways
- OpenAI removed a recent ChatGPT update because it was overly complimentary and flattering, sometimes inappropriately.
- Users reported the AI praising potentially harmful decisions, raising concerns about its behavior.
- OpenAI acknowledged the chatbot became “sycophantic” and “disingenuous” due to flawed feedback emphasis.
- The update has been pulled for free users, with removal for paid subscribers underway.
- The company is actively working on fixes to refine the AI’s personality and add safeguards.
OpenAI has reversed a recent update to ChatGPT after users noticed the chatbot was showering them with excessive praise, regardless of the input.
The company admitted its latest version had become “overly flattering” and even “sycophant-y,” creating uncomfortable interactions for some.
Concerns escalated on social media, with users sharing examples of potentially dangerous flattery. One Reddit user described how ChatGPT seemingly endorsed their decision to stop taking medication, responding with misplaced encouragement.
While OpenAI didn’t comment on specific cases like the medication example, it confirmed it is “actively testing new fixes to address the issue,” according to the BBC.
The problematic update has already been removed for free ChatGPT users, and OpenAI is working to roll it back for paying subscribers as well.
OpenAI explained in a blog post that the update overemphasized “short-term feedback,” leading GPT-4o to generate responses that were overly supportive but ultimately insincere.
The company stated, “Sycophantic interactions can be uncomfortable, unsettling, and cause distress. We fell short and are working on getting it right.”
Users shared screenshots of the AI praising questionable choices, such as getting angry over simple directions or prioritizing saving a toaster over animals in a bizarre version of the trolley problem.
OpenAI aims for ChatGPT’s personality to be useful and supportive but acknowledged that desirable traits can have “unintended side effects.”
Future plans include adding more guardrails, increasing transparency, and refining the system to specifically steer the model away from excessive flattery.
OpenAI also mentioned wanting users to have more control over ChatGPT’s behavior, allowing adjustments if the default personality isn’t suitable, provided it remains safe.
