AI safety commitments are hard to enforce when the people responsible for a company’s direction also face pressure to grow it. Helen Toner, executive director of Georgetown University’s Center for Security and Emerging Technology, says the answer cannot be to simply trust one person to make the right decisions.
Why OpenAI’s board mattered
Toner served on OpenAI’s nonprofit board, one of four members who voted to fire Sam Altman. He returned as CEO, Toner resigned, and the company appointed a new board of directors. The episode, sometimes called “the blip,” ended with a broader restructuring of OpenAI.
The board’s role was meant to protect the company’s mission to build safe AI. Its purpose was to guard against profit and growth taking priority over safety. When board members tried to use that authority, the accountability arrangement broke down.
That history matters because governance is only useful if it can work under pressure. A safety promise may sound clear on paper, but the people tasked with enforcing it need enough independence and authority to act when business priorities pull in another direction.
Safety promises face commercial pressure
The interview describes a wider debate about how AI companies handle risk. The CEOs of frontier AI labs have called for a slowdown and asked Washington for help. The White House responded with a voluntary accord that relies in part on board accountability, even though OpenAI’s experience showed how fragile that mechanism can be.
Anthropic, OpenAI’s biggest competitor, is also about to go public. That step could bring new market pressures to the company. The interview does not establish how those pressures will affect Anthropic’s safety work, but it raises a practical question: how will companies maintain commitments when they are also answering to commercial demands?
Toner’s central concern is institutional. If a safety plan depends on any one executive being trustworthy, then the plan has no reliable backstop. Accountability needs to function even when leaders face incentives to move quickly or prioritize growth.
What policymakers need from AI research
Toner leads CSET, a team at Georgetown that began in 2019. Its purpose is to give policymakers rigorous, actionable analysis about emerging technologies, including AI, with attention to both the technology and national security. The group produces long reports, briefs busy policymakers, and uses a data team to support its research.
That work takes place in a political environment where technology companies spend money to influence candidates and public debate. Toner acknowledges that money can overwhelm some policy areas, but argues that it has limits. In the election discussed in the interview, efforts to push AI governance, regulation, and safety out of the debate did not work very well.
She points to growing voter interest and bipartisan attention as reasons. Policymakers across Congress and the executive branch want high-quality information as AI becomes more relevant to their work. Research organizations and industry lobbyists both help explain complex policy issues, although they may approach them from different positions.
Regulation needs more than good intentions
The conversation also covers whether AI alignment can be solved, what AI labs are trying to build, and how regulation should account for China. These questions are connected: policymakers need a grounded understanding of the technology, while safety systems need rules and oversight that do not depend solely on a company’s internal assurances.
The source points to a gap between commitments and enforcement. Voluntary agreements and board oversight can contribute to safety, but their value depends on whether they remain effective when incentives conflict. The OpenAI episode gives that concern a concrete example: a board created to protect the mission acted, yet the accountability process did not hold.
For Toner, the durable approach is to build systems that can withstand those conflicts. That means giving policymakers useful expertise and ensuring that safety oversight can operate without relying on personal trust alone.