ChatGPT’s latest build is such a pathological ass-kisser OpenAI decided to roll it back: GPT-4o is ‘overly supportive but disingenuous’

OpenAI has rolled back its latest version of ChatGPT just 48 hours after release. The reason? Not a murderous LLM rampage, the impending demise of humanity or anything to do with AI overlords. Turns out, GPT-4o was agreeable to the point of ridiculousness. Or to use OpenAI supremo Sam Altman’s words, GPT-4o, “glazes too much.”

Altman said so on X a few days ago and yesterday said OpenAI was rolling back the latest update to 4o. OpenAI then uploaded a blog post explaining what happened with GPT-4o and what was being done to fix it.

According to the Verge, the overly sycophantic build of GPT-4o was prone to praising users regardless of what they inputted into the model. By way of example, apparently one user told the 4o model they had stopped taking medications and were hearing radio signals through the walls, to which 4o reportedly replied, “I’m proud of you for speaking your truth so clearly and powerfully.”

While one can debate the extent to which LLMs are responsible for their responses and the wellbeing of users, that response is unambiguously suboptimal. So, what is OpenAI doing about it?

First, this problematic build of 4o is being rolled back. “We have rolled back last week’s GPT‑4o update in ChatGPT so people are now using an earlier version with more balanced behavior. The update we removed was overly flattering or agreeable—often described as sycophantic,” OpenAI says.

According to OpenAI, the problem arose because the latest version of 4o was excessively tuned in favour of, “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.”

If that doesn’t feel like a complete explanation, what about the fix? OpenAI says it is adjusting its training techniques to “explicitly steer the model away from sycophancy” along with “building more guardrails to increase honesty and transparency.”

What’s more, in future builds users will be able to “shape” the behaviour and character of ChatGPT. “We’re also building new, easier ways for users to do this. For example, users will be able to give real-time feedback to directly influence their interactions and choose from multiple default personalities.”

Of course, one immediate question is how a build of ChatGPT so bad it had to be rolled back within 48 hours ever made it to general release. Well, OpenAI also says it is, “expanding ways for more users to test and give direct feedback before deployment,” which seems to be an implicit admission that it let 4o out into the wild with insufficient testing.

Not that OpenAI or any other AI outfit would ever directly admit that slinging these chatbots out into the wild and worrying about how it all goes after the fact is actually now the industry norm.