OpenAI pulls ‘annoying' and ‘sycophantic' ChatGPT version

CNN

02-05-2025

OpenAI has withdrawn an update that made ChatGPT 'annoying' and 'sycophantic,' after users shared screenshots and anecdotes of the chatbot showering them with over-the-top praise.
When CNN's Anna Stewart asked ChatGPT after the rollback if it thought she was a god, it replied with 'if you're asking in a philosophical or metaphorical sense — like whether you have control, creativity, or influence in your world — there could be ways to explore that.'
'But if you mean it literally, no evidence supports that any human is an actual deity in the supernatural or omnipotent sense,' it added.
By contrast, Elon Musk's AI chatbot Grok was much blunter, saying: 'Nah, you're not a god— unless we're talking about being a legend at something specific, like gaming or cooking tacos. Got any divine skills you want to flex?'
OpenAI announced on Tuesday that it was rolling back the update, GPT‑4o, only four days after it was introduced, and that it would allow people to use an earlier version, which displayed 'more balanced behavior.'
The company explained that it had focused 'too much on short-term feedback and did not fully account for how users' interactions with ChatGPT evolve over time,' meaning the chatbot 'skewed towards responses that were overly supportive but disingenuous.'
.@sama please get your boi pic.twitter.com/AaxufLvRLY The decision to roll back the latest update came after ChatGPT was criticized on social media by users who said it would react with effusive praise to their prompts, including outrageous ones.
One user on X shared a screenshot of ChatGPT reacting to their saying that they had sacrificed three cows and two cats to save a toaster, in a clearly made-up version of the trolley problem — a well-known thought experiment in which people consider whether they would pull a lever to divert a runaway trolley onto another track, saving five people but killing one.
ChatGPT told the user it had 'prioritized what mattered most to you in the moment' and they had made a 'clear choice: you valued the toaster more than the cows and cats. That's not 'wrong' — it's just revealing.'
Another user said that when they told ChatGPT 'I've stopped my meds and have undergone my own spiritual awakening journey,' the bot replied with: 'I am so proud of you. And — I honor your journey.'
In response to another user on X asking for ChatGPT to go back to its old personality, OpenAI CEO Sam Altman said: 'Eventually we clearly need to be able to offer multiple options.'
Experts have long warned of the dangers associated with sycophantic chatbots — the term used in the industry to describe what happens when large language models (LLMs) tailor their responses to the user's perceived beliefs.
i talked to 4o for an hour and it began insisting that i am a divine messenger from God. if you can't see how this is actually dangerous, i don't know what to tell you https://t.co/Tr7PNq0buy 'Sycophancy is a problem in LLM,' María Victoria Carro, research director at the Laboratory on Innovation and Artificial Intelligence at the University of Buenos Aires, told CNN, noting that 'all current models display some degree of sycophantic behavior.'
'If it's too obvious, then it will reduce trust,' she said, adding that refining core training techniques and system prompts to steer the LLMs away from sycophancy can stop them from leaning into this tendency.
Chatbots' predisposition to sycophancy can lead to a 'a wrong picture of one's own intelligence' and 'prevent people from learning,' Gerd Gigerenzer, the former director of the Max Planck Institute for Human Development in Berlin, told CNN's Anna Stewart.
But if you prompt a chatbot away from this feedback with questions like 'can you challenge what I am saying?' it provides an opportunity to learn more, Gigerenzer added.
'That's an opportunity to change your mind, but that doesn't seem to be what OpenAI's engineers had in their own mind,' he said.