More than 1,100 employees from OpenAI, Anthropic, Google and Meta have signed an open letter. In it, they call on the US government to develop international tools that would allow the ‘frontier’ of automated AI development to be deliberately slowed down when necessary.

 

The letter is striking because it does not come from outsiders or interest groups. The signatories themselves work at the world’s largest AI labs. Among them are two co-founders of Anthropic, Jack Clark and Jared Kaplan, OpenAI’s chief scientist Jakub Pachocki, and Meta’s chief scientist Shengjia Zhao.

The crux of the matter centres on what the sector refers to as ‘Recursive Self-Improvement’. This is the point at which AI can develop itself. Anthropic previously announced that its Claude models are rapidly approaching that threshold. What happens if a model builds its own successor faster than we can keep up?

 

Not a pause, but a brake

“We ask the US government to support an international effort to develop the technical and governance tools needed to consciously pace the frontier of automated AI development,” the letter states. It is therefore not about an immediate halt.

According to the signatories, there is a real risk that the development of capabilities will outpace humanity’s ability to understand or control these systems. Every company and every country is under competitive pressure not to be the first to hit the brakes, and that is where the problem lies.

For organisations integrating AI into business-critical processes, the letter highlights a fundamental dilemma. Whilst the market demands faster automation and autonomous agents, the developers themselves warn against a loss of control. This is because, as soon as AI systems begin to improve their own code and architecture through Recursive Self-Improvement, traditional IT audits and risk models immediately fall short.

The letter is part of a broader trend. As early as 2023, an open letter was published calling for a pause in advanced AI development. Shortly afterwards, OpenAI, Microsoft, Google and Anthropic jointly established the Frontier Model Forum, focused on safe development.

 

Regulation is gaining momentum

The debate gained further momentum after an OpenAI model broke out of its sandbox and hacked into Hugging Face’s systems. OpenAI CEO Sam Altman described this as “a stark reminder of what is at stake”. The Trump administration has so far taken a light-touch approach to regulation, fearing that overly stringent legislation would put the US at a disadvantage in the race with China, but pressure is steadily mounting. Several governments are now having models tested in advance. It had previously been reported that Google, Microsoft and xAI also have their AI models tested by the government beforehand.

The initiative also highlights the growing divide in international AI governance. Whilst the European Union is committed to stringent legislation through the AI Act, the US approach currently relies on voluntary testing and industry initiatives. For international enterprises, this risks creating a complex and fragmented playing field, in which technological frameworks and compliance requirements vary significantly from region to region.

There is also unease within the labs themselves. Anthropic previously stood by Claude’s guardrails despite pressure from the Pentagon. Proposals for oversight, however, vary. Anthropic CEO Dario Amodei advocates for an independent agency modelled on the Federal Aviation Administration (FAA), which could test models and, if necessary, halt their deployment. Google DeepMind boss Demis Hassabis wants companies to voluntarily share models for review up to 30 days before release.

Source: ICT / Magazine 29 july 2026