Internal politics

OpenAI dissolves safety team

OpenAI has dissolved its AI safety team after two key members left last week, expressing frustration over its lack of safety prioritization

Martin Crowley
May 20, 2024

Last week, OpenAI co-founder, Ilya Sutskever, and key AI researcher, Jan Leike, both quit OpenAI’s “Superalignment” team, which was focused on implementing long-term AI safety protocols, with Leike openly expressing his frustration over the company’s priorities, claiming that “safety culture and processes have taken a backseat to shiny products.”

Following Sutskever and Leike’s departure, OpenAI then decided to dissolve the AI safety team for good.

What led to Leike and Sutskever’s rapid departure?  

Sutskever and Leike co-led the Superalignment team, which OpenAI formed in July, last year. They were tasked with solving the technical challenges surrounding controlling superintelligent AI systems over the next four years.

As competition in the AI space started to heat up, OpenAI began to prioritize the speed of new product development over safety protocols.

The team was reportedly promised 20% of OpenAI’s compute resources to complete their work but were often denied these resources, leaving them unable to help develop safe superintelligent AI, causing friction between key team members and OpenAI’s leadership team, leading to a “breaking point” for Leike, Sutskever, and several other team members.

What did Leike have to say about his departure?

Following his resignation, Leike publicly stated that he had disagreed with the company's "core priorities" for "quite some time":

“We are long overdue in getting incredibly serious about the implications of AGI…I joined because I thought OpenAI would be the best place in the world to do this research. However, I have been disagreeing with OpenAI leadership about the company’s core priorities for quite some time, until we finally reached a breaking point.”

OpenAI’s response

In response to Leike’s frustration, OpenAI CEO, Sam Altman, and co-founder, Greg Brockman, (publicly) acknowledged that OpenAI has "a lot more to do” and needs “to have a very tight feedback loop, rigorous testing, careful consideration at every step, world-class security, and harmony of safety and capabilities.”

They have established that they’re building the foundations for the safe deployment of AI systems and have claimed that since they released ChatGPT-4, a year ago, they have  "continuously improved model behavior and abuse monitoring in response to lessons learned from deployment."  

Despite dissolving its AI safety team, they’ve stated that, as it releases new models, it will keep working with governments and stakeholders on safety, using its “Preparedness Framework” to help predict and mitigate AI-related risks.