OpenAI Safety Researchers Deny Leaking Secrets, Blame Transparency Advocacy
Three former OpenAI safety researchers have denied allegations of leaking confidential information, arguing instead that their terminations were retaliation for advocating for AI transparency and the preservation of verifiable reasoning steps. This high-profile dispute highlights a growing friction between leading AI labs and safety researchers regarding corporate accountability, independent auditing, and the suppression of internal dissent. Such conflicts risk chilling external safety collaborations and undermining the governance of advanced artificial intelligence. OpenAI investigated the researchers for allegedly sharing internal model architecture details over Signal with an external researcher, though the company maintained the messaging app was not the reason for dismissal. The researchers have written to OpenAI's board calling for stronger cooperation with independent safety organizations.
## BACKGROUND
Reasoning models use internal scratchpads known as chains of thought to record intermediate problem-solving steps, helping researchers understand how an AI reaches its conclusions. However, some labs have recently explored reducing these visible steps in newer models, sparking industry debates over monitoring and transparency.