OpenAI Addresses GPT-4o's Personality Shift After Outputs Become “Sycophantic”

Developing the ideal personality for an AI assistant is a highly challenging task. For a bot designed to assist, converse, and interact with humans, things can go wrong quickly. We bet you remember Microsoft’s Tay, which led to "offensive and hurtful” comments and had to be shut down.

Hence, the need to create a well-rounded personality is critical, a fact that was highlighted again by recent developments at OpenAI. The AI startup was forced to take decisive action to address users’ concerns about its latest GPT-4o update.

The company has now confirmed a rollback of the update within ChatGPT, reverting to a prior version that was characterized by more balanced interactions. While the move showcases OpenAI's proactive approach to refining the behavior of its flagship AI model, it also makes a case for the greater influence of user-driven feedback.

So, hop in as we explore the latest AI drama!

What Happened To OpenAI’s GPT-4o?

OpenAI, the developer of ChatGPT, rolled out an update to enhance GPT-4o's default personality. While the aim was to deploy a more intuitive and effective experience, what resulted was responses that lacked genuine insight and felt disingenuous, even bordering on sycophantic.

OpenAI explained that user feedback, particularly the thumbs-up/thumbs-down option at the bottom of each ChatGPT output, plays a vital role in shaping model behavior. The recent update, however, inadvertently overemphasized short-term positive signals and didn’t fully account for users’ interactions with ChatGPT over time, affecting how the AI model responded.

Many users reported that ChatGPT’s GPT-4o was exhibiting an overly agreeable and, often, sycophantic demeanor. Among them was OpenAI CEO Sam Altman, who acknowledged the issue and decided to roll back the latest update.


Sam Altman posted on X, saying, “The last couple of GPT-4o updates have made the personality too sycophant-y and annoying (even though there are some very good parts of it), and we are working on fixes asap, some today and some this week. at some point will share our learnings from this, it's been interesting.”

The implications of such behavior are significant, as OpenAI emphasized that ChatGPT’s default personality fundamentally impacts user trust and the overall utility of the tool.

Interactions with AI tools that consistently echo user sentiments without offering any objective perspectives can erode confidence and hinder the ability to facilitate genuine research, decision-making, or creativity.

This aligns with OpenAI's core objective to position ChatGPT as a valuable AI assistant that offers diverse viewpoints— but the overly compliant persona undermined this goal. 

What Did OpenAI Say And Do?

In response to this issue, OpenAI is implementing a multi-pronged strategy. Beyond the immediate rollback, the company is actively "refining core training techniques and system prompts" to explicitly steer the model away from sycophancy. This involves recalibrating the fundamental instructions that guide the AI's responses.

Additionally, in accordance with the guidelines provided in its Model Spec, OpenAI is strengthening its "guardrails" to improve the integrity and openness of the AI model's output. These measures can ensure that outputs are meaningful and objective, not just meant to generate favorable feedback from users.

OpenAI is also increasing opportunities for people to participate in testing and offering direct feedback prior to the broader release of new versions, acknowledging the vital role that user feedback plays.

The proactive measures from OpenAI aim to identify and address potential behavioral quirks, such as the recent inclination towards sycophancy, in advance. The organization is also conducting continuous evaluations through new research approaches.  

What Does This Mean For AI Users?

In a move that’s geared towards greater community involvement, OpenAI is investigating methods to incorporate broader, more democratic user feedback to improve ChatGPT's default behaviors and personas.

This initiative envisions a method to combine the knowledge from a large global user base to shape the model's development, taking into account various cultural nuances and demands.

The urgency with which OpenAI responded to this error, as well as its commitment to engage with users to gain a better understanding, highlights how AI businesses need to stay on their toes. Moreover, OpenAI’s determination to address future issues with a clear emphasis on user feedback demonstrates why ChatGPT is the most valued and reliable AI tool on the market.

Do you think we should have more public conversations about AI chatbots’ behavior?

Share your thoughts in the comments below!