AI
Former OpenAI board member warns that AI safety cannot depend on trusting individual CEOs
Helen Toner, who voted to fire Sam Altman at OpenAI, argues the industry's true ambitions—and governance failures—demand systemic solutions rather than faith in leadership.
The Verge
Helen Toner, executive director of the Center for Security and Emerging Technology at Georgetown University, was one of four board members who voted to remove Sam Altman as CEO of OpenAI in November 2023. That move, now often called "the blip," ultimately failed when Altman returned to the helm, Toner resigned, and OpenAI restructured with a new board entirely. Yet the episode exposed something fundamental: the accountability mechanism designed to keep the company aligned with its nonprofit mission to pursue safe AI had shattered when it mattered most.
In a conversation with The Verge's Nilay Patel, Toner emphasized a stark reality that emerged from that experience. "If our hopes for the future depend on any one person being trustworthy, we're in big trouble," she said. The implication is clear—faith in individual executives, however well-intentioned, is not a viable strategy for AI governance.
Why the traditional oversight model broke#
The OpenAI board's role was explicit: ensure the company didn't abandon its mission to develop safe AI in pursuit of profit and growth. When the board actually exercised that power, the broader system failed to support it. Altman's swift reinstatement showed that even board-level resistance buckles under market and employee pressure. That failure matters because similar board structures are now being proposed as safeguards elsewhere—including in a White House voluntary accord on AI safety that relies partly on the same kind of corporate governance that proved inadequate at OpenAI.
Anthropic is about to go public, introducing fresh commercial pressures to a company positioning itself as OpenAI's main competitor. These market forces will intensify the exact tension Toner witnessed firsthand: when safety commitments collide with shareholder expectations.
What the labs actually want to build#
Toner stressed that most people encounter AI as consumer products—ChatGPT's free tier, Google's AI overviews, voice assistants—and misunderstand what leading labs are truly pursuing. OpenAI and Anthropic were not built as consumer internet companies, she noted. Rather, both harbor a vision that has become normalized in Silicon Valley but remains largely invisible to the public: developing AI capable of recursive self-improvement (RSI) that eventually produces superintelligence—systems vastly exceeding human capabilities.
That objective shapes everything from what these companies fund to how they structure their research. Understanding this gap between the chatbot people use and the goals animating the labs is crucial to any meaningful policy response.

