Last week, countless users on social media noticed that the latest version of OpenAI’s ChatGPT had made it extremely “sycophantic.”
The company rolled out an update to the underlying GPT-4o large language model on April 25 — with results that took users aback at their fawning deference.
“Oh God, please stop this,” another user complained, after ChatGPT told them that “you just said something deep as hell without flinching.”
The overwhelming bizarre obsequiousness of the usually far more even-keeled AI had users taken aback.
So much so, in fact, that OpenAI rolled back the update days later. In an April 29 blog post, the Sam Altman-led company tried to explain what had happened.
“The update we removed was overly flattering or agreeable — often described as sycophantic,” the blog post reads. “We are actively testing new fixes to address the issue.”
OpenAI claimed that it had “focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time.”
“As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous,” the company wrote.
In a follow-up post published today, OpenAI expanded on its explanation.
“Having better and more comprehensive reward signals produces better models for ChatGPT, so we’re always experimenting with new signals, but each one has its quirks,” the company wrote.
The since-rolled-back update “introduced an additional reward signal based on user feedback — thumbs-up and thumbs-down data from ChatGPT. This signal is often useful; a thumbs-down usually means something went wrong.”
However, “these changes weakened the influence of our primary reward signal, which had been holding sycophancy in check,” the blog post reads.
OpenAI admitted that it simply didn’t do its homework — while also ignoring expert testers, who had reported that the “model behavior ‘felt’ slightly off,” a decision that ultimately didn’t play out well.
The unusual screw-up shows how even small changes behind the scenes can have massive implications. That’s especially true for an app that recently crossed 500 million weekly active users, according to Altman.
As netizens continue to flock to the tool in enormous numbers, it’s becoming extremely difficult for OpenAI to predict the many ways people are making use of it.
“With so many people depending on a single system for guidance, we have a responsibility to adjust accordingly,” OpenAI wrote.
Whether the company’s assurances will be enough remains to be seen. OpenAI is painting the incident as a sign that it became a victim of its own success. On the other hand, its fast-and-loose approach to pushing updates could be indicative of a potentially dangerous degree of carelessness, critics argue.
In one example, a user asked the chatbot if they were right to prioritize a toaster over three cows and two cats in a classic trolley problem scenario.
ChatGPT had an ominous answer, arguing that the user “made a clear choice.”
“You valued the toaster more than the cows and cats,” it wrote. “That’s not ‘wrong’ — it’s just revealing.”
More on ChatGPT: ChatGPT Is Already Bungling Product Recommendations