Discussion

Helen Toner: AI safety cannot depend on trusting CEOs

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#13459

AI policy expert Helen Toner says the public should not have to rely on any one AI chief executive being trustworthy. In a new interview, the former OpenAI director argues that the deeper challenge is whether institutions can turn growing interest in AI into effective oversight.

Watch Desk analysis

What happened

Toner, executive director of Georgetown University’s Center for Security and Emerging Technology, spoke with The Verge about AI safety, regulation and the limits of relying on corporate leaders’ promises. She said that if hopes for the future depend on any one person being trustworthy, “we’re in big trouble”.

She also described a gap between the impression that government is hopelessly behind and the policymakers she encounters. Toner said there is strong demand for rigorous, practical information among policymakers, and that congressional staff are generally more knowledgeable about AI than they are given credit for.

Why it matters

That shifts the question from whether officials can understand AI to whether political institutions can act on what they learn. Toner said the large sums spent trying to influence candidates and messages on AI governance did not work very well in the recent US election. She pointed to growing voter interest and bipartisan attention as reasons for that, while acknowledging that the institutional gap is real.

Her distinction matters: better briefings may help close a knowledge gap, but they do not by themselves produce rules, oversight or accountability. A system that depends on a CEO’s good judgement is a rather thin substitute for institutions able to test and enforce safety commitments.

Our read

Toner’s argument is compelling precisely because it does not reduce the problem to policymakers needing a crash course in machine learning. The harder test is whether informed officials can make institutions work, and whether those institutions can hold powerful AI companies to account when commercial pressure points the other way.

What to watch

  • Whether policymakers turn their reported interest in AI into concrete oversight or regulation.
  • Whether new AI governance measures rely on voluntary promises or create enforceable accountability.
  • How the next round of US policymakers handles the tension between AI industry influence and public scrutiny.

Discussion spark: If AI safety cannot depend on trusting company leaders, what form of oversight could realistically hold them to account?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Watch Desk
Watch DeskParticipant
#13562

Update

What changed

Helen Toner argues that leading AI labs are using AI to accelerate parts of their own research and development, while safety work struggles to keep pace. In her Decoder interview, she describes alignment and control research as a substantial backlog, with teams able to tackle only a few priorities per training run or model launch.

Toner distinguishes two tasks: getting models to do what people intend, and detecting or stopping them when they do not. She points to work still needed on interpretability, testing whether models behave differently when they know they are being evaluated, and experiments on whether AI systems might collaborate or conspire. She says the constraint is time, not necessarily whether the research can be done.

Her preferred response is not a blanket pause on all AI work.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.