Reported US shift could strengthen Britain’s AI safety checks

Catching flaws before an artificial intelligence (AI) system reaches the public gives developers a chance to fix them. Britain’s researchers could have more room to do that following a reported change in Washington.

On 11 October, Reuters, citing The Telegraph, reported that the US had dropped its opposition to American developers sharing unreleased models with the UK’s AI Security Institute. Such access lets researchers test a system before its wider release.

The change remains a reported policy shift. It does not confirm that particular companies have resumed sharing models, or guarantee that testing will catch every problem.

There is evidence that early access can produce practical results. In a parliamentary answer on 9 October, the government said the institute’s voluntary relationships with leading AI laboratories had helped identify “dozens of vulnerabilities” that developers fixed before release.

The process is straightforward: researchers examine a model, flag weaknesses and give its maker an opportunity to respond. Those reported fixes show why keeping these testing relationships open matters.

Sources

• Reuters — US reverses course on Britain's testing of AI models, Telegraph says
• UK Parliament — Written answer on pre-release AI safety testing (HL3654)