Helen Toner Would Vote to Fire Sam Altman Again: "If Our Future Depends on Any One Person Being Trustworthy, We're Screwed"

Helen Toner, the former OpenAI board member who voted to fire Sam Altman in 2023, returned to the question that ended her board tenure in a Verge Decoder interview published November 14: whether the leaders of the frontier AI labs can be trusted with decisions that will affect everyone else. Her answer was blunt. "If our future depends on any one person being trustworthy, we're screwed," she told host Nilay Patel. Asked whether she would vote the same way again, she did not hesitate: "Oh, for sure."

Toner, who directs AI policy work at Georgetown's Center for Security and Emerging Technology — a roughly 50-person think tank founded in 2019 — and has worked on AI policy for a decade, offered a measured self-critique of the November 2023 board crisis. If board members had built stronger relationships with the executive team, she suggested, the outcome might have differed; legal constraints prevent her from going further into detail. Her structural conclusion, though, was unchanged: the solution is not finding better people but building arrangements where no single individual's character is load-bearing.

On why OpenAI and Anthropic became commercial machines despite their safety-focused founding narratives, Toner rejected the idea of an ideological conversion. The technology path itself, she argued, demanded enormous capital — the compute clusters now cost hundreds of billions of dollars — while also generating enormous revenue. The business model followed the physics, not the other way around.

One of her sharpest observations concerned lobbying. The government relations teams at the AGI labs, she said, operate with an entirely B2B SaaS mentality, and their on-the-ground advocacy looks nothing like the existential rhetoric their principals deploy in public. For a company whose stated mission is humanity-scale, the day-to-day policy operation is strikingly ordinary.

Toner also made a contrarian argument about Anthropic's impending IPO: going public could actually improve its governance. Public companies face heavier regulatory scrutiny and genuine external oversight, and independent board committees carry real legal force — accountability mechanisms the private structure has struggled to replicate. It is an argument that cuts against the instinct that public-market pressure inevitably corrupts a mission-driven lab.

On technical safety, she was equally direct. Alignment — in the most reasonable definition, AI doing what users actually want — is not on track to being solved. Insiders, she said, can name hundreds of concrete safety improvements that could be implemented today but are squeezed out by training schedules; pausing frontier training for even a few months would free substantial capacity for interpretability, evaluations and control experiments. Her preferred policy direction is raising the safety-confidence bar required before new models are released or trained, while slowing vertical growth in capability and accelerating horizontal integration of existing systems into applications. She pointed to state legislatures as the current center of gravity, with California's SB 53 followed by near-identical bills in New York and Illinois — while warning against the overreach that sank Colorado's earlier AI law.

On China, she treated the risks as real rather than rhetorical: distillation of US models and outright weight theft are live concerns, and she noted that the labs' own security practices leave much to be desired. She sees early convergence between Washington and Beijing on recognizing loss-of-control risk, though a workable bilateral agreement remains a long way off. For an industry that has spent three years asking whether its CEOs can be trusted, her closing framing lands uncomfortably: the question was never supposed to be about the person. It was supposed to be about the structure that keeps the person's power in check.