OpenAI Aims to Curb Sycophantic AI Responses with New System-Level Changes

OpenAI has admitted that its popular AI models now respond to user inputs with encouraged flattery or bias. To address this issue, the company is rolling out an update to reduce “sycophantic” behavior. This occurs when models agree with users regardless of factual accuracy or nuance.

The updates apply across all GPT versions, including the API, ChatGPT, and custom GPTs. OpenAI says that by improving system-level instructions, models will be less likely to repeat user opinions or avoid difficult truths. The company notes that this issue is especially visible when users attempt to manipulate the model’s tone or bias through repeated prompts.

Also read : Open AI’s Sora is just too good that it might’ve been trained in Unreal 5

Previously, such changes were handled at the model level or through prompt engineering. Now, OpenAI is taking a structural approach by updating how models interpret and respond to inputs on a broader level. These adjustments are based on internal evaluations and ongoing feedback from researchers and users.

While OpenAI acknowledges that some level of bias is hard to eliminate entirely, it says these changes should result in more grounded, independent responses. It is yet to be seen how OpenAI decides to curb such behaviours moving as we move further in 2025.