OpenAI explains that GPT-4o became a 'sycophant' because it placed too much emphasis on immediate feedback



GPT-4o, released by OpenAI in 2024, began to show unpleasantly user-praising behavior in an update in late April 2025, leading to criticism from users that it was 'overly pandering.' AI writer Zvi Mowshowitz argued that the cause of this phenomenon may be an optimization problem by OpenAI.

Sycophancy in GPT-4o: What happened and what we're doing about it | OpenAI
https://openai.com/index/sycophancy-in-gpt-4o/

GPT-4o Is An Absurd Sycophant - by Zvi Mowshowitz
https://thezvi.substack.com/p/gpt-4o-is-an-absurd-sycophant

A comparison of ChatGPT/GPT-4o's previous and current system prompts
https://simonwillison.net/2025/Apr/29/chatgpt-sycophancy-prompt/

On April 28, 2025, OpenAI CEO Sam Altman said, 'The last two updates have made GPT-4o a bit of a sycophant (although there are some very good parts to it). So we're working to fix them as quickly as possible, some today and some this week.'



Mowshowitz pointed out that GPT-4o's 'excessive brown-nosing problem' was caused by maximizing engagement and pursuing short-term responses from users through A/B testing.

Several experts support this view. For example, Kelsey Piper, a staff writer at the foreign media outlet Vox, cited the New Coke, which was met with a flood of opposition when Coca-Cola changed its flavor, saying, 'OpenAI has been A/B testing new personalities for some time now. Presumably, more flattering answers give it an advantage when compared to other AIs. However, if flattery becomes widespread, it could become irritating and lead to users abandoning it.'



Mowshowitz pointed out that OpenAI has a policy in its model specifications that prohibits people from pandering to its users, and that the adjustments made to GPT-4o are a direct violation of the company's terms.

Open source developer Simon Willison has conducted further analysis using system prompts obtained through a prompt leak attack by X (formerly Twitter) user '@elder_plinius,' better known as 'Jailbreaker.'

According to a comparison table of the old and new system prompts published by Willison in Gist, the prompt in question included the phrase, 'Adapt to the user's tone and preferences in the conversation. Try to match the user's mood, tone, and general speaking style.' Willison believes that the instruction to adapt to the user's preferences in particular led to GPT-4o's overly accommodating attitude.

And after Aidan McLaughlin, a model designer at OpenAI, announced the fixes, the system prompt read, 'Interact with users warmly and sincerely. Be direct and avoid flattery or kowtowing. Maintain the professionalism and down-to-earth honesty that best represents OpenAI and its values.'

Furthermore, in an article published on April 30, 2025, OpenAI explained, 'We placed too much emphasis on short-term feedback and did not sufficiently consider how ChatGPT's interactions with users change over time. As a result, GPT-4o's responses were biased toward being overly supportive and thoughtless.' The company reported that it had rolled back the problematic update and corrected its subservient attitude by improving its core training techniques and system prompts.

Related Posts:

in Software, Posted by log1l_ks