AI safety testing needs access
Reports say the US has restored Britain’s access to unreleased AI models. Outside scrutiny needs access—and clear consequences for what evaluators find.
Britain may regain an important window into unreleased AI systems. Reuters, citing The Telegraph’s October 11 report, says the US has reversed a request that American developers stop sharing models with Britain before release.
The Telegraph reports that lobbying by UK AI minister Kanishka Narayan helped restore access for the AI Security Institute. Reuters says the White House, OpenAI and Anthropic did not immediately respond to its requests for comment. This remains an attributed report: Reuters’s account draws on the Telegraph, rather than providing a separate confirmation. We have not located a published government agreement setting out the restored access.
What access makes possible
The AI Security Institute describes its work as testing leading systems before and after public release, investigating risks, and informing government decisions. Its research spans cyber misuse, safeguards, autonomy, human influence, and whether developers can maintain control. Access matters because an outside evaluator cannot meaningfully examine a system it cannot use.
That does not mean every evaluation proves a model safe. AISI describes a government research organization, not a blanket approval for products. The reported reversal also does not tell us how much time evaluators receive, which systems they can inspect, or how their findings affect a release.
The questions to keep asking
Our view is that scrutiny needs consequences. Can evaluators examine the system that will actually be deployed, with its tools and permissions? Can they publish important findings, including unfavorable ones? Who can require a developer to address an unresolved risk before proceeding?
Access is a necessary starting point for those questions, not an answer to all of them. A company’s willingness to cooperate is useful; lasting accountability needs arrangements that survive commercial pressure and political changes.
There are competing proposals for getting there. In his September essay, Anthropic CEO Dario Amodei proposes embedded outside evaluators and coordinated pacing. That is a developer’s proposal, not evidence that the full system is operating across the industry.
We support independent scrutiny, meaningful safeguards, and public accountability alongside useful AI. Our Take action guide explains the superintelligence statement we support and other ways to join the policy conversation. Participation is optional; enthusiasm for ChatGPT should leave room for difficult questions.