Helen Toner: AI safety cannot depend on trusting CEOs

Toner, executive director of Georgetown University’s Center for Security and Emerging Technology, said the OpenAI board’s experience showed why AI companies need stronger governance and broader accountability. The nonprofit board was meant to ensure the company did not put commercial growth ahead of its mission to build safe AI. Altman returned as CEO, Toner resigned and the company later restructured.

The discussion ranged from AI regulation and technical safety to the possibility of systems improving the process used to build future models. Toner said leading companies including OpenAI and Anthropic are increasingly using AI in their research and development, and some people inside the firms believe this is accelerating their work. She cautioned against treating artificial general intelligence as a clear milestone: the term covers different capabilities, and current systems meet some definitions but not others.

Toner said there is substantial safety work that companies could pursue, including testing models in more settings, studying how they behave and improving methods for alignment and control. But she warned that researchers have limited time and that solving the underlying scientific questions takes sustained work. If companies believe they may need to address those risks within 12 to 24 months, she said, that is “terrifying”; she hopes they are wrong, but considers the possibility serious enough to prepare for.

Rather than stopping AI development across the board, Toner said policymakers could set a higher bar for confidence in risk management before companies train or release more capable systems. She also argued that the “But China!” case against US restrictions should be treated as a real strategic question, not a blanket excuse. That means examining whether Chinese firms can use US models to generate training data and whether security failures could expose model weights or other valuable technical information.

Toner said US and Chinese officials may be showing early signs of shared concern about major cyberattacks and loss of control, though she expects little trust between the governments. In the United States, she said she will watch whether Congress can pass legislation and whether states continue coordinating rules aimed at frontier AI risks. She cautioned against sweeping regulation covering every AI system, favoring clearer standards and transparency for the most advanced models.