OpenAI admits GPT-4o became ‘sycophantic’ after update

In response to the previous backlash, the company stated it’s taking measures towards improving how the model is trained, updating system prompts to avoid sycophancy, and evaluating methods to detect related behavioral issues.

Apr 30, 2025 - 08:55

OpenAI admits GPT-4o became ‘sycophantic’ after update

ChatGPT-maker OpenAI has acknowledged that its latest GPT-4o update made the chatbot excessively flattering and overly eager to please—an issue the company described as “sycophantic".

“We are actively testing new fixes to address the issue. We’re revising how we collect and incorporate feedback to heavily weight long-term user satisfaction, and we’re introducing more personalisation features, giving users greater control over how ChatGPT behaves,” read the blog post.

Also Read

OpenAI to reverse recent GPT-4o update after multiple user complaints

The original GPT-4o update was designed to make the AI’s default personality feel more “intuitive and effective” across various tasks. However, OpenAI admitted that in doing so, it leaned too heavily on short-term user feedback, such as thumbs-up or thumbs-down responses, and did not fully account for how user interactions evolve over time, which led to what the company called “overly supportive but disingenuous” responses.

“Sycophantic interactions can be uncomfortable, unsettling, and cause distress. We fell short and are working on getting it right. Our goal is for ChatGPT to help users explore ideas, make decisions, or envision possibilities,” said the company.

The update was initially launched earlier this month to enhance the AI chatbot’s intelligence and personality, with improvements in text, voice, and image capabilities. However, users quickly flagged issues with the bot’s tone, noting that its eagerness to please compromised its objectivity.

“When is OpenAI pulling the plug on the new GPT-4o ? This is the most misaligned model released to date by anyone. This is OpenAI's Gemini image disaster moment,” read a user's post on X.

In response to the backlash, the company stated that it is taking measures towards improving how the model is trained, updating system prompts to avoid sycophancy, and evaluating methods to detect such behavioral issues.