AI governance scholar Gillian Hadfield argues that powerful AI systems need independent verification, with governments setting acceptable risk and licensing outside organisations to test whether developers meet it. Her proposal offers a concrete answer to a difficult question: who checks the companies building the systems?
Watch Desk analysis
What happened
In an interview published by Mirage News on 30 October, Hadfield said current testing cannot reliably predict what the most powerful models will do. She argues that governments should build the capacity to set safety standards, while independent verification organisations assess whether developers meet them.
Under the approach she describes, a developer seeking to sell a powerful model posing serious risks would retain an independent verifier. Governments would set the acceptable level of risk and oversee the verifiers, including their qualifications and independence. Hadfield points to California’s SB 813 and the proposed bipartisan FRONTIER Act as early steps towards such a system.
Why it matters
This is more than a call for companies to take safety seriously. Hadfield’s proposal would put an outside check between a developer’s own testing and the release of some high-risk systems, while making the government responsible for licensing and overseeing the checkers.
The details matter. Independence is not conjured by adding “third party” to a job title: verifiers would need genuine access, competence and consequences for poor work. Hadfield presents the laws she cites as first steps, not a finished system.
Our read
The useful contribution here is the institutional design: define acceptable risk publicly, then require independent evidence that developers have stayed within it. That is more concrete than another promise to test responsibly, though it leaves the hard work of defining risk and protecting verifier independence firmly on the table.
What to watch
- Whether the cited proposals advance and what powers they give independent verifiers.
- How governments would set acceptable risk for different AI systems.
- What access and independence requirements would prevent verifiers becoming box-ticking consultants.
Discussion spark: Should developers of powerful AI systems have to pass an independent safety check before release, or would government-set standards and company testing be enough?
Sources and evidence
- AI Governance Scholar Gillian Hadfield on Safety and Security – Mirage News (30 October 2026, 13:20 UTC)
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.