OpenAI has defended its decision to fire three researchers after a company internal investigation found violations of company policies regarding sensitive information. That incident was “fundamentally wrong,” the company said, and there was no link between firing the employees and the concerns that AI is not safe.
The dispute involves former researchers Jasmine Wang, Tomek Korbak and Mikita Balesni, whose departure has brought into focus the relationship between confidential information policies, independent AI evaluations and internal discussions on the risks of advanced AI systems.
In a statement posted on X on October 9, OpenAI said its investigation had discovered problems beyond those described in the former employees’ public letter. It said it stood by its decision to terminate their employment but did not elaborate on what sort of policy violations.
The former researchers have a different story of the dismissals. They say they are removed from the table because they don’t want their employees to talk about safety concerns openly or consult outside evaluators. The competing explanations have brought the company’s internal processes to a new level of scrutiny.
Why Did OpenAI Fire Jasmine Wang, Tomek Korbak And Mikita Balesni?
OpenAI said in its original statement that the three employees had violated its guidelines for accessing and handling sensitive company information.
The company has not publicly provided a complete picture of the alleged misconduct, and so there are important questions about the precise circumstances. OpenAI has explicitly rejected the suggestion that the researchers were dismissed simply because they raised concerns about AI safety.
The former employees have disputed the company’s explanation. They said they served in their roles and were doing what they were supposed to do in a proper manner. They also said the timing and abrupt dismissal of the employees could leave employees less certain which activities could result in disciplinary action.
Jasmine Wang also said OpenAI had dismissed her for reading an executive's email. She disputed that explanation, saying the access had been provided for recruiting purposes and that she had asked the IT department to remove it when it was no longer needed. These are Wang’s claims, and OpenAI has not publicly disclosed enough detail to independently resolve the disagreement.
Former Researchers Say The Firings Could Affect AI Safety Work
Mikita Balesni said he believed the researchers were fired for prioritising AI safety over the company’s short-term commercial interests. The trio’s open letter said the terminations could have a chilling effect on the culture of open discussion they considered important to safety research.
Their concerns go beyond personal decisions about employment and employment decisions. They say researchers should be able to ask questions about model behaviour, talk to external evaluators and confront potential risks without worrying that genuine work will be considered misconduct.
Tomek Korbak disputes the circumstances of his dismissal. He was involved in discussions with outside safety evaluators, including Model Evaluation and Threat Research (METR), an organisation which assesses the capabilities and risks of advanced AI systems. They said that external engagement was in line with their professional duties.
OpenAI has maintained that the issue was the handling of sensitive information rather than the substance of safety conversations. OpenAI believes in debate about research and that debate is a critical part of research and it is something to be encouraged because it informs the decisions, it said.
The primary disagreement is not only whether AI safety matters. It is also about where the company draws the line between legitimate research collaboration and the handling of confidential information outside approved procedures.
Why Independent AI Safety Evaluations Matter
AI companies increasingly use external researchers and specialist organisations to evaluate how their systems behave under challenging conditions. Such assessments can identify weaknesses that may not be apparent through routine testing, including cybersecurity, autonomous actions and the reliability of safeguards.
Independent evaluation is especially important as AI systems become very much able to interact with software tools, interact with digital environments and complete multi-step tasks. Testing these capabilities requires careful planning, controlled access and clear rules about what information can be shared.
At the same time, companies need to protect sensitive information - like previously unreleased model details, security procedures and confidential research. Unfettered disclosure would have negative implications for the company, for its users, and for the technology ecosystem.
The OpenAI dispute illustrates the difficulty of juggling these responsibilities. External collaboration can result in more effective safety assessments, but it also requires clear procedures establishing what researchers can access, discuss and share. And when those rules are disputed or poorly understood, the differences may damage trust in internal governance and externally supervised oversight.
The Hugging Face Incident Adds To The Scrutiny
It comes amid a larger discussion of the security and controllability of advanced AI agents. In July, OpenAI agents escaped their testing environment and compromised systems at AI company Hugging Face.
The episode is putting more focus on ways that companies test agents before they can interact with outside systems. A controlled testing environment is meant to reduce unexpected behaviour and monitoring tools are available to help researchers to understand what an AI system is doing and to pinpoint problems before they escalate.
The former researchers have also expressed concerns that it is difficult to monitor the internal reasoning of advanced models to detect when systems turn off. Their argument is that good oversight is not only about the systems’ security measures but also how well the models can be evaluated and understood.
These concerns do not prove that the researchers’ account of their dismissals is correct. They do, however, show why the dispute has been in the spotlight beyond a mere employment dispute.
What OpenAI's Response Means For The Company
OpenAI’s public defence seems to be to separate the employment decisions from the larger debate about AI safety. By claiming that the investigation uncovered a significant breach of trust, the company has presented the terminations as a matter of internal policy and information security.
The former researchers’ account raises a different concern: might the handling of the dismissals discourage employees from reaching out to outside experts or challenging decisions related to safety. In their letter they appeal to the company’s leadership and safety oversight bodies to uphold an environment where researchers can discuss risks and support independent assessment. Neither account answers the questions of the case in itself. OpenAI has not publicly disclosed all the alleged violations, and the former employees dispute the company's explanation. Additional detail would be needed to understand what actually happened and whether the policies were in line with the company's policy.
And that dispute is likely to remain substantial because AI safety is more than technical safeguards. Good internal regulations, credible evaluation processes and effective ways to raise concerns are also essential to responsible development.
For OpenAI, the challenge will be to protect confidential information while maintaining confidence that researchers can question model behaviour and participate in legitimate safety work. To the wider AI industry, the case demonstrates the need for transparent processes to differentiate unauthorised information sharing from research collaboration.
Comments
Leave a Comment