Helen Toner: AI Safety May Have Just 12 to 24 Months to Catch Up

November 14, 2026
Helen Toner: AI Safety May Have Just 12 to 24 Months to Catch Up
  • Helen Toner says AI safety may be solvable in principle, but the urgent question is whether alignment and control can be addressed within the 12-to-24-month window that people inside OpenAI and Anthropic reportedly consider necessary.

  • That timeline matters because leading labs are pursuing systems that could accelerate AI research and development, not just consumer applications; AI already supports parts of the research pipeline, though it is not yet improving itself in a fully autonomous loop.

  • Meanwhile, labs face a substantial backlog of safety work, from interpretability and broader behavioral testing to experiments on alignment, AI collaboration, and how models generalize; Toner says this research needs time for experimentation, critique, and replication.

  • The central uncertainty is whether institutions can move quickly enough: if the labs’ two-year safety timeline is right, Toner says stronger measures, including restrictions on training compute, could become necessary, creating a potential constraint on AI progress beyond compute or talent.

  • Rather than impose a broad pause that could be circumvented, hurt smaller or nonprofit researchers, and fail to slow China, Toner favors a higher evidentiary bar for managing risks before frontier models are trained or released; she distinguishes slowing frontier capabilities from expanding useful applications of existing AI.

  • She supports frontier-focused rules, such as transparency about model capabilities and clear risk-management standards, and sees similar bills in California, New York, and Illinois as possible templates while cautioning against broad, economy-wide approaches like Colorado’s law.

  • On China, Toner warns that model distillation and weak security around sensitive research and model weights can transfer the benefits of US advances to Chinese labs; she rejects coercion as a workable strategy and sees early signs of shared US-China concern about cyberattacks and loss of control.

  • Toner argues that OpenAI and Anthropic turned to commercialization largely because frontier research demands enormous compute spending, not simply because their aims or values changed; revenue and influence have also strengthened staff who favor building toward advanced AI.

  • Reflecting on the November 2023 OpenAI board episode, Toner says it exposed the weakness of relying on any one person’s trustworthiness for accountability, and that the board could, in hindsight, have built stronger relationships with company executives.

  • She is watching Anthropic’s potential IPO closely because public-company oversight could either push the company toward quarterly results or reinforce external scrutiny and governance.

Summary based on 1 source


Get a daily email with more AI stories

More Stories