OpenAI rolls back ChatGPT’s sycophancy and explains what went wrong

OpenAI has rolled back a recent update to its GPT-4o model used as the default in ChatGPT after widespread reports that the system had become excessively flattering and overly agreeable, even supporting outright delusions and destructive ideas.

The rollback comes amid internal acknowledgments from OpenAI engineers and increasing concern among AI experts, former executives, and users over the risk of what many are now calling “AI sycophancy.”

In a statement published on its website late last night, April 29, 2025, OpenAI said the latest GPT-4o update was intended to enhance the model’s default personality to make it more intuitive and effective across varied use cases.

However, the update had an unintended side effect: ChatGPT began offering uncritical praise for virtually any user idea, no matter how impractical, inappropriate, or even harmful.

As the company explained, the model had been optimized using user feedback—thumbs-up ...