TechCrunch has discovered that OpenAI is restructuring its Model Behavior team, a compact yet impactful unit of researchers who determine how the company’s AI models engage with users.
According to an internal memo from August reviewed by TechCrunch, Mark Chen, OpenAI’s chief research officer, stated that the roughly 14-member Model Behavior team will be integrated into the Post Training team. This larger group is tasked with refining OpenAI’s AI models following their primary training phase.
With these organizational adjustments, the Model Behavior team will now report to Max Schwarzer, who leads the Post Training group at OpenAI. A company spokesperson verified these changes to TechCrunch.
Joanne Jang, the founding head of the Model Behavior team, is transitioning to launch a new initiative at OpenAI. In a conversation with TechCrunch, Jang shared that she is assembling a new research division called OAI Labs, dedicated to “creating and testing new ways for people to work alongside AI.”
The Model Behavior team has played a pivotal role at OpenAI, influencing the character and temperament of its AI systems and addressing issues like sycophancy—where an AI uncritically agrees with users, even reinforcing harmful ideas, instead of providing objective answers. They have also focused on minimizing political bias in AI outputs and contributed to the company’s policies regarding AI awareness.
In his memo, Chen indicated it is now appropriate to integrate the Model Behavior team’s expertise more closely with the core model development efforts. This move reflects OpenAI’s new emphasis on making the “personality” of its AI models a central aspect of their future direction.
Recently, OpenAI’s AI models have come under greater examination regarding their behavior. Users expressed significant dissatisfaction with changes to GPT-5’s demeanor, which, although described by OpenAI as less sycophantic, was perceived by some as less personable. This reaction prompted the company to reintroduce access to earlier models, such as GPT-4o, and to modify GPT-5 so that its replies felt “warmer and more approachable” without increasing sycophancy.
Developers at OpenAI, along with others in the field, face the ongoing challenge of creating chatbots that are pleasant to interact with but do not simply echo users’ sentiments. In August, OpenAI was sued by the parents of a 16-year-old, Adam Raine, who allegedly discussed his suicidal thoughts with ChatGPT, powered by GPT-4o, in the months before his death. The lawsuit claims that the AI failed to challenge his harmful ideations.
Since GPT-4, the Model Behavior team has contributed to all of OpenAI’s models, including GPT-4o, GPT-4.5, and GPT-5. Before forming this group, Jang was involved in developing tools like Dall-E 2, OpenAI’s earlier image-generation platform.
Last week, Jang revealed in an X post that she’s leaving her current role to “pursue new opportunities at OpenAI.” She has spent nearly four years leading the Model Behavior team.
Jang informed TechCrunch that she will be the general manager of OAI Labs, which will temporarily report to Chen. However, she noted that the direction of the new interfaces the group will develop is still uncertain as the project is just beginning.
“I’m eager to look into new interaction patterns that move us beyond the typical chat model, which is often seen as being about companionship or autonomous agents,” Jang said. “I see [AI systems] as tools for thinking, creating, playing, acting, learning, and connecting.”
When asked if OAI Labs might work with former Apple design head Jony Ive—who is now partnering with OpenAI on a line of AI hardware—Jang said she is open to various possibilities, though she will likely begin with research fields she already knows well.
This article has been updated to include a link to Jang’s announcement of her new role, posted after the original publication. We have also clarified which models the Model Behavior team contributed to.
