Helen Toner on AI Safety, Governance, and Why Trust Should Not Rest on One Person
Helen Toner, executive director of the Center for Security and Emerging Technology (CSET) at Georgetown University, is the subject of a Decoder podcast interview published by The Verge on 14 November 2026. The conversation covers AI safety, AI alignment, AI regulation, and China.
Toner previously held a seat on the OpenAI board. She was one of four board members who voted to fire Sam Altman. After that vote, Altman returned as CEO, Toner resigned, and OpenAI got a new board of directors.
In the interview, Toner said that if hopes for the future depend on any one person being trustworthy, "we're in big trouble."
The quote is short, but it addresses a structural question that the rest of the conversation touches: how much of the outcome of frontier AI development (the most capable models, built by a small number of labs) rests on the judgment of a few leaders, and how much rests on institutions that can operate without them. The episode's subjects bear on that question from different directions. Alignment research, which studies whether a model's behavior tracks the goals its developers intend, asks whether those goals are being met. Regulation asks which external bodies can check those developers. The China thread asks how both questions look when the frontier is contested between national systems.
The broader context here is the one the quote points to. For much of the industry's recent history, governance of leading AI labs has been arranged around the people running them, through founder control, board composition, and informal norms. The 2023 board confrontation that Toner took part in showed how quickly that arrangement can become a question of personalities.
In my view, the more useful reading of Toner's remark is as a design critique rather than a verdict on any individual. A system that works only when a particular person behaves well has a single point of failure, and such failures tend to become visible only after they have occurred.
That reading does not require assuming bad faith. Most people who build these systems are, by every public account, trying to do the right thing. Distributed-systems engineers already work on this assumption. They do not design databases to depend on one administrator never making a mistake. They add replication, access controls, audit trails, and failover, so that trust is spread across mechanisms rather than concentrated in a person. Applied to AI governance, that means defining who can review a model release, who can halt one, and what evidence those decisions rest on, in terms that outlast any single executive or board.
Where the technology is heading, the interview suggests that these questions have moved into mainstream policy discussion rather than remaining a niche concern for safety researchers. CSET's remit, as a Georgetown research center, sits squarely in that policy space. Whether the frameworks that emerge can survive personnel changes is the question the quote leaves open. The more encouraging possibility is that a field which has spent years learning to make software resilient to individual error can apply the same habit to its own leadership.
