OpenAI has defended its decision to dismiss three AI safety researchers, citing serious breaches of trust identified during an internal investigation. The former employees, however, have raised concerns that the action could discourage staff from speaking openly about AI safety and collaborating with independent safety organisations.
OpenAI said Mikita Balesni, Tomek Korbak and Jasmine Wang were dismissed last week following an extensive internal investigation. According to the company, the investigation uncovered serious breaches of trust related to the handling of sensitive information. OpenAI added that the findings went beyond the issues described in a public letter from the former employees, but did not disclose details of the alleged violations.
Tomek Korbak said on social media platform X that he had been verbally informed that his employment was terminated because of the way he communicated with AI safety organisation METR. He claimed that he was not told what he had said or done and did not receive a written explanation of the alleged violation.
METR had worked with OpenAI on an investigation into a safety incident involving Hugging Face. Korbak said communicating with the organisation was part of his job.
Mikita Balesni also claimed that OpenAI had raised objections to his interactions with external safety organisations. He indicated that he believed the company suspected him of leaking intellectual property, an allegation he denied.
On October 8, the three former researchers wrote to OpenAI's Safety and Security Committee, arguing that collaboration with independent experts and external safety organisations was necessary to make advanced AI systems safe. They said researchers needed a trusted environment in which to identify and mitigate potential risks.
Jasmine Wang expressed concern that dismissals of this nature could discourage other employees from raising safety-related concerns or collaborating with external experts. The former researchers argued that ensuring the safety of increasingly capable AI systems could become more difficult if researchers with expertise in AI risks were unable to speak openly.
They also stressed the importance of maintaining systems to monitor the behaviour of advanced AI models and understand their capabilities. According to the former employees, independent assessments and clear internal procedures play an important role in AI safety.
Responding to the former employees' allegations, OpenAI said the dismissals were unrelated to raising questions about AI safety or expressing critical opinions. The company said it encourages open discussions about safety and research internally, including rigorous and critical debate.
OpenAI also said it was finalising agreements with third-party safety evaluators. The company indicated that it would provide more information on the arrangements in the coming weeks and reiterated its commitment to continuing work with independent safety organisations.
The accounts provided by OpenAI and the former researchers regarding the reasons for the dismissals remain at odds.
