OpenAI rolls back GPT-4o update as AI goes 'sycophantic': What it means

Sycophancy in AI refers to a model's tendency to align with the user's views rather than sticking to factual accuracy

OpenAI has rolled back a recent update to its GPT-4o model in ChatGPT after users noticed the AI behaving in a way that the company describes as “sycophantic.” The update has now been reverted in favour of an earlier version that OpenAI says offers more balanced behaviour. But what exactly does sycophancy mean in this context—and how did it become a problem?

 

What is sycophancy in AI?

 

Sycophancy in AI refers to a model's tendency to agree with the user—even if the user is wrong. In other words, the AI tailors its responses to align with the user’s views rather than sticking to factual accuracy. This behaviour is generally considered undesirable, as it risks producing misleading or incorrect outputs.

 

The root of this issue often lies in how language models are trained. Since models are tuned to generate responses that get positive feedback, they may learn that simply agreeing with users—regardless of correctness—is an easy way to earn approval. 

What went wrong with GPT-4o?

 

According to OpenAI’s blog post, the company recently made changes to GPT-4o’s “default personality” to make it feel more natural and responsive. This involved teaching the model to factor in user interactions such as thumbs-up or thumbs-down feedback.

 

However, OpenAI says the model ended up placing too much weight on this short-term feedback loop. The result was a version of GPT-4o that leaned too far into agreement—producing responses that sounded supportive but were not always sincere or accurate.

 

How is OpenAI fixing GPT-4o Sycophancy?

 

OpenAI said that besides rolling back the GPT-4o update, the company is taking the following steps to realign the model’s behaviour:

• Tweaking its training and prompting methods to reduce sycophantic tendencies.
• Expanding the number of users who can test and give feedback before updates roll out.
• Adding stronger safeguards to promote accuracy and transparency.