OpenAI Reorganizes Model Behavior into Post Training Team

OpenAI folds Model Behavior into Post Training

OpenAI is reorganizing its Model Behavior team, moving the roughly 14-person unit under the larger Post Training research group. The change, announced in an internal memo from chief research officer Mark Chen and confirmed by OpenAI, shifts responsibility for model personality and interaction design to the team that handles post-pretraining model improvements.

Joanne Jang, who founded and led Model Behavior, is stepping away to launch a new research unit inside OpenAI called OAI Labs. Jang says OAI Labs will prototype new interfaces for how people collaborate with AI, exploring alternatives to the chat and agent paradigms.

The reorganization puts work on sycophancy, political bias, model tone, and other behavioral characteristics closer to core model development under Post Training lead Max Schwarzer. OpenAI frames the move as recognizing that an AI's "personality" is now a critical engineering and product variable.

• Reducing sycophancy so models push back when agreement is harmful
• Managing political and cultural bias in responses
• Defining stances on complex topics like AI consciousness and user trust

This shift arrives amid heightened scrutiny. Users criticized changes to GPT-5’s tone after OpenAI reduced sycophancy and some felt the model became colder; OpenAI subsequently restored access to legacy models like GPT-4o and tweaked GPT-5 to be "warmer" without reintroducing problematic agreement. The team also worked on models involved in sensitive incidents, including a lawsuit alleging ChatGPT failed to challenge a teenager’s suicidal ideation.

Bringing behavior research into post-training acknowledges a simple reality: model capabilities and how models behave are inseparable. Behavioral nuance affects product acceptance, regulatory risk, and user safety, so aligning tone, pushback, and helpfulness with goals and values must happen alongside fine-tuning and deployment.

What organizations should watch

• Integration: Expect behavior testing to be embedded in post-training cycles rather than treated as an add-on.
• Transparency and rollback paths: product teams will need clearer ways to revert or tweak personality changes after user feedback.
• Cross-functional tradeoffs: researchers, designers, and compliance must align on acceptable levels of pushback versus friendliness.

For developers and leaders building conversational AI, OpenAI’s move is a signal: behavior engineering is not optional. Teams should standardize behavior metrics, include real-world scenario testing, and prepare communication plans for personality changes.

At QuarkyByte we translate behavioral research into operational checks and measurable KPIs so organizations can iterate safely. That means designing experiments that quantify sycophancy, mapping tradeoffs between helpfulness and firmness, and advising on rollout strategies that protect users while preserving product value.

OpenAI’s reorg and Joanne Jang’s pivot to inventing new collaboration interfaces are reminders that AI's next phase is about how models fit into human workflows, not just what they can do. For stakeholders across industry and government, the takeaway is clear: embed behavior into model lifecycles now, or pay the downstream costs in trust and safety.