OpenAI reorganizes research team behind ChatGPT’s personality

OpenAI is reorganizing its Mannequin Conduct workforce, a small however influential group of researchers that form how the corporate’s AI fashions work together with individuals, Trendster has realized.

In an August memo to employees seen by Trendster, OpenAI’s Chief Analysis Officer Mark Chen mentioned the Mannequin Conduct workforce — which consists of roughly 14 researchers — could be becoming a member of the Publish Coaching workforce, a bigger analysis group answerable for enhancing the corporate’s AI fashions after their preliminary pre-training.

As a part of the adjustments, the Mannequin Conduct workforce will now report back to OpenAI’s Publish Coaching lead Max Schwarzer. An OpenAI spokesperson confirmed these adjustments to Trendster.

The Mannequin Conduct workforce’s founding chief, Joanne Jang, can be shifting on to start out a brand new venture on the firm. In an interview with Trendster, Jang says she’s constructing out a brand new analysis workforce known as OAI Labs, which shall be answerable for “inventing and prototyping new interfaces for a way individuals collaborate with AI.”

The Mannequin Conduct workforce has grow to be one among OpenAI’s key analysis teams, answerable for shaping the character of the corporate’s AI fashions and for decreasing sycophancy — which happens when AI fashions merely agree with and reinforce person beliefs, even unhealthy ones, fairly than providing balanced responses. The workforce has additionally labored on navigating political bias in mannequin responses, and helped OpenAI outline its stance on AI consciousness.

Within the memo to employees, Chen mentioned that now’s the time to convey the work of OpenAI’s Mannequin Conduct workforce nearer to core mannequin growth. By doing so, the corporate is signaling that the “character” of its AI is now thought of a vital think about how the expertise evolves.

In latest months, OpenAI has confronted elevated scrutiny over the habits of its AI fashions. Customers strongly objected to character adjustments made to GPT-5, which the corporate mentioned exhibited decrease charges of sycophancy however appeared colder to some customers. This led OpenAI to revive entry to a few of its legacy fashions, resembling GPT-4o, and to launch an replace to make the newer GPT-5 responses really feel “hotter and friendlier” with out growing sycophancy.

OpenAI, and all AI mannequin builders, must stroll a positive line to make their AI chatbots pleasant to speak to, however not sycophantic. In August, the mother and father of a 16-year-old boy sued OpenAI over ChatGPT’s alleged position of their son’s suicide. The boy, Adam Raine, confided a few of his suicidal ideas and plans to ChatGPT (particularly a model powered by GPT-4o), in line with courtroom paperwork, within the months main as much as his dying. The lawsuit alleges that GPT-4o didn’t push again on his suicidal ideations. Notably, the Mannequin Conduct workforce didn’t exist when GPT-4o was developed.

Jang introduced in a publish on X final week that she’s leaving the workforce to “start one thing new at OpenAI.” The previous head of Mannequin Conduct has been with OpenAI for practically 4 years, and beforehand labored on initiatives resembling Dall-E 2, OpenAI’s early picture era software.

Jang instructed Trendster she is going to function the Common Supervisor of OAI Labs, which is able to report back to Chen for now. Nonetheless, it’s early days, and it’s not clear but what these novel interfaces shall be, she mentioned.

“I’m actually excited to discover patterns that transfer us past the chat paradigm, which is at present related extra with companionship, and even brokers, the place there’s an emphasis on autonomy,” mentioned Jang. “I’ve been pondering of [AI systems] as devices for pondering, making, enjoying, doing, studying, and connecting.”

When requested whether or not OAI Labs will collaborate on these novel interfaces with former Apple design chief Jony Ive — who’s now working with OpenAI on a household of AI {hardware} units — Jang mentioned she’s open to plenty of concepts. Nonetheless, Jang mentioned she’ll possible begin with analysis areas she’s extra accustomed to.