Fired OpenAI researchers call for independent AI safety audits
What's the story
Three former OpenAI employees have raised concerns about the safety of artificial intelligence (AI) systems. They have called on the company to increase external scrutiny of its AI models, arguing that safeguards need to evolve as more advanced models become harder to evaluate. The researchers made their case in a letter sent to OpenAI board members and safety committees on Wednesday, The Wall Street Journal reported.
Safety worries
Concerns over AI developers' understanding
The letter from the former employees, identified as Jasmine Wang, Tomek Korbak, and Mikita Balesni, raised concerns about AI developers potentially losing their ability to comprehend or track the internal reasoning of increasingly sophisticated systems.
They cautioned against approaches that could make models less observable or harder to assess.
The intervention comes after their termination by OpenAI over an internal investigation into their handling of confidential information.
Company stance
Terminations due to policy violations, says OpenAI
OpenAI has said that the terminations were due to alleged violations of company policies on sensitive research material. The company claimed that investigators found a broader pattern of improper handling of research information.
However, it has not publicly identified the outside organization involved or what specific information was shared.
In their letter, the former employees argued that these dismissals could deter other employees from voicing concerns about AI safety.
Safety protocols
OpenAI 'strongly agreed' with recommendations
An internal memo from a company research leader, shared with The Wall Street Journal by an OpenAI spokesperson, said the organization "strongly agreed" with the recommendations concerning AI monitoring and outside assessment.
The memo also clarified that the terminations weren't due to employees raising safety concerns.
"We do not terminate employees for raising concerns," it read, emphasizing that monitoring its AI systems remains a major priority and acknowledging a role for external evaluators in the wider safety framework.