OpenAI Explains GPT-4o’s Overly Agreeable And Sycophantic Outputs And Revised Strategy For Model Testing

Building an AI assistant with the right personality is no small feat. One minute it’s helping you research a challenging topic, the next it’s agreeing with everything you say.

Like many businesses developing AI tools, OpenAI’s ChatGPT faced this challenge, too. The AI powerhouse was forced to roll back an update to GPT-4o within ChatGPT, after it delivered overly agreeable responses without any genuine insight, even bordering on sycophancy.

After reverting the popular AI chatbot to a previous version, known for its more balanced and objective responses, the creator of ChatGPT has opened up about what went wrong. Dive in!  

The Agreeability Bug In ChatGPT’s GPT-4o Update

On April 25th, OpenAI rolled out an update intended to refine the default personality of its cutting-edge GPT-4o model within ChatGPT. The aim was to create a more intuitive and memorable user experience—however, the outcome was far from the intended mark.

Users soon reported interactions that didn’t feel like engaging with a helpful AI assistant but rather like encountering an overly agreeable, even sycophantic, yes-man.
The company noted that this behavior could extend to "validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended."

OpenAI also recognized that such “sycophantic” tendencies could raise safety mental health concerns, emotional over-reliance, and potentially lead to risky behaviors. In fact, even OpenAI’s CEO, Sam Altman, weighed in on the issue, acknowledging the shift in the model’s personality in a post on X.

So, how did OpenAI respond to users’ feedback?  

OpenAI’s Response To GPT-4o's Concerns

OpenAI swiftly responded to the chorus of user concerns and initiated a rollback of the problematic update on April 28th—three days after its initial release. However, the process of reverting to GPT-4o's stable version took approximately 24 hours.


Beyond the immediate rollback, OpenAI has now articulated a comprehensive strategy to prevent similar issues in the future.

Firstly, the company will institute a more rigorous approval process for launching new AI models, explicitly considering model behavior and quantitative metrics. This means issues with hallucinations, deception, loss of reliability, and changes in personality will be treated as potential roadblocks to deployment.

Even if standard metrics like A/B testing show positive results, qualitative signals and proxy measurements of behavioral aspects will be given significant weight in the decision-making process.

The company also emphasized the increased value it will place on spot checks and interactive testing conducted by internal experts. Learning from this experience, OpenAI recognized that qualitative assessments can uncover issues that automated evaluations and A/B tests might miss, particularly with nuanced aspects of model behavior and consistency.

Secondly, OpenAI plans to introduce an optional "alpha" testing phase for certain updates. This will allow users interested in providing direct feedback to interact with the model before its wider release, offering an additional layer of scrutiny and user feedback.

A key component of this improvement involves a more thorough evaluation of adherence to its established "Model Spec," which outlines the desired behavior principles for its AI models. While OpenAI has extensive evaluations in areas such as instruction following and model safety, it aims to strengthen its confidence in assessing less quantifiable aspects of model behavior.

Finally, OpenAI acknowledged its missteps in communicating about the flawed update. Recognizing that even seemingly tiny changes can have a major influence on user experience, the AI developer committed to be more proactive about upgrades to ChatGPT, regardless of their perceived impact.

Along with detailed descriptions about the update, OpenAI will also outline known restrictions, ensuring users are aware of both positive and negative traits of the model. Despite the flawed update to GPT-4o, the overall outcome seems to have raised the bar of AI model testing.


With AI tools influencing our daily lives, it’s becoming more important than ever to have open conversations about its ethical challenges.

The latest incident is a powerful reminder that creating the perfect AI assistant is a continuous process—one that demands careful evaluation, thorough testing, and user-generated feedback. More than anything, it requires a readiness to change and grow.

Would you rather have an AI assistant that challenges your views, agrees with you, or has a balanced take?

Let us know your thoughts in the comments below!