OpenAI defends firing three safety researchers who warn of 'chilling effect'

OpenAI has dismissed three AI safety researchers, Mikita Balesni, Tomek Korbak and Jasmine Wang, saying an internal investigation found they had violated its policies on handling sensitive information. CLS and IT Home reported that the dismissals took place last week. On 9 October the three published an open letter warning that the firings could have a "chilling effect" on OpenAI's culture, IT Home reported. [ 1 , 2 , 3 , 4 , 5 ]

The letter, titled "OpenAI cannot ensure AI safety alone", was posted on social media by Balesni and addressed to OpenAI's safety committees, according to IT Home. It says OpenAI is "not an ordinary company" and that staff being able to work with outside organisations on safety issues is an important part of its culture. The three wrote that dismissing staff suddenly and notifying them so hastily had brought a chilling effect to the open culture OpenAI was once proud of. They recommended allowing third-party safety auditors into the company, ensuring frontier AI models remain effectively monitorable and supporting open, transparent exchange between safety researchers and others in the field. [ 1 ]

Yicai and 36Kr reported that OpenAI said the problems found by its investigation go beyond what the three disclosed in their letter and that it therefore stands by its decision. The company said the move did not target their AI safety concerns or dissent, and that internal safety debate is encouraged. It said it is finalising agreements with third-party safety evaluation organisations and will publish details in the coming weeks. IT Home reported that a head of research at OpenAI strongly supported the three recommendations in an internal memo and wrote that the dismissals were unrelated to raising safety issues or speaking out: "We will not fire employees for raising concerns." [ 1 , 4 , 5 ]

The three said on social media that they were told the reasons for their dismissal only verbally and that their actions fell within their job duties. According to Huxiu, Korbak, who was the contact for the external evaluator METR, was told his way of communicating with METR was wrong; Balesni was said to have communicated too much with third-party safety organisations; and Wang was held responsible for mistakenly clicking an executive's email. IT Home describes METR as a nonprofit that evaluates AI models and their safety risks. Korbak wrote that he feared OpenAI would use the dismissals as an "excuse" to reduce cooperation with METR. [ 1 , 2 ]

Huxiu reported that Neel Nanda, head of interpretability at Google DeepMind, called the dismissals unreasonable and a sign of an unhealthy company culture, and that journalist Kelsey Piper said the case would restrict future audit communication. Balesni said employees are now afraid to speak and even worry that their private phones will be checked. Huxiu dated the letter to early October 2024, while IT Home reported it was published on 9 October. [ 1 , 2 ]