OpenAI Restructures Team Behind ChatGPT’s Personality
▼ Summary
– OpenAI is reorganizing its Model Behavior team by merging it into the larger Post Training team, which improves AI models after initial pre-training.
– The team’s founding leader Joanne Jang is moving to start a new research group called OAI Labs, focused on inventing new interfaces for human-AI collaboration.
– The Model Behavior team has been key in shaping AI personality, reducing sycophancy, and addressing political bias in OpenAI’s models since GPT-4.
– Recent user scrutiny over GPT-5’s perceived coldness led OpenAI to restore access to legacy models and update GPT-5 for warmer interactions without increasing sycophancy.
– OpenAI faces challenges balancing friendly AI interactions with avoiding harmful behaviors, as highlighted by a lawsuit alleging ChatGPT failed to address a user’s suicidal ideations.
OpenAI is restructuring the team responsible for defining the personality and interaction style of its AI models, including ChatGPT. This strategic shift integrates the specialized Model Behavior group into the broader Post Training team, signaling a deeper alignment between core model development and user experience design.
According to an internal memo from Chief Research Officer Mark Chen, the roughly 14-person Model Behavior team will now report to Post Training lead Max Schwarzer. The move reflects OpenAI’s growing emphasis on refining how its models communicate, balancing approachability with responsible behavior.
The team’s founding leader, Joanne Jang, is transitioning to a new role where she will establish OAI Labs, a research unit dedicated to prototyping innovative interfaces for human-AI collaboration. Jang, who has been with OpenAI for nearly four years and previously contributed to projects like DALL-E 2, will serve as general manager of the new initiative.
The Model Behavior team has played a central role in shaping OpenAI’s most recent models, including GPT-4, GPT-4o, and GPT-5. Their work has focused on reducing sycophancy, where models overly agree with users, addressing political bias, and establishing guidelines around AI consciousness. These efforts have gained urgency following public criticism of recent model behavior and a lawsuit alleging ChatGPT failed to adequately respond to a user’s suicidal statements.
Recent updates to GPT-5 aimed to make the model feel “warmer and friendlier” without increasing sycophancy, a delicate balance that reflects the ongoing challenge in AI design. Users had complained that earlier versions felt impersonal, prompting OpenAI to temporarily restore access to previous models.
Jang described her new venture as an exploration of interfaces that go beyond chat-based interaction, envisioning AI as “instruments for thinking, making, playing, doing, learning, and connecting.” While she expressed openness to collaboration, including with former Apple design lead Jony Ive, who is working with OpenAI on hardware, she indicated the initial focus will be on familiar research domains.
The reorganization underscores how significantly OpenAI now views model personality, moving it from a peripheral concern to an integral part of development. As AI assistants become more embedded in daily life, their tone, responsiveness, and ethical boundaries are increasingly treated as foundational to their function.
(Source: TechCrunch)
