Should the safety of increasingly powerful artificial intelligence depend on trusting the companies building it? A former Google employee who describes leaving as a whistleblower argues that it should not. The more useful question is whether governments can establish binding rules whose observance can be checked independently.

That shifts the debate away from the personalities of technology executives and toward the infrastructure behind AI: the chips, data centres and computing capacity needed to develop advanced systems.

Why corporate promises are not enough

The concern raised in the interview is fundamentally about incentives. Technology companies are businesses, and commercial success matters to them. Investment in safety can coexist with pressure to launch products, win customers and outperform competitors.

That does not establish that any particular executive is dishonest or that every company will cut corners. It does explain why voluntary assurances are a weak foundation for public protection. When slowing down carries a competitive cost, a company’s interests may diverge from the public’s.

Effective oversight should not require assuming either perfect goodwill or universal bad faith. It should set clear obligations, make compliance observable and attach consequences to violations.

What computing power has to do with AI safety

The interview proposes focusing on compute: the computational resources used to train and operate AI systems. Developing leading-edge models typically requires substantial computing capacity, often supplied by specialised chips working together in large clusters.

Compute is not the only ingredient in AI capability. Data, algorithms and engineering also matter. But physical hardware offers a potential point of oversight that is more tangible than a company’s statement about its intentions.

The proposal is to track computing resources through the supply chain and verify that they are being used within agreed limits. The aim is not necessarily to stop AI development. It is to preserve useful progress while reducing the chance that developers build systems whose capabilities outpace the ability to evaluate, monitor or control them.

What verifiable rules might involve

The short interview sets out a principle rather than a detailed regulatory design. Turning that principle into policy would require answers to several practical questions:

• What is covered? Rules would need to specify which hardware, facilities or computing activities trigger reporting and oversight.
• What must be disclosed? Tracking chip ownership is different from knowing how a computing cluster is actually being used.
• Who checks compliance? Verification needs qualified oversight bodies with sufficient access and independence.
• What happens after a breach? Binding obligations need enforceable consequences, not simply requests for better behaviour.
• How are legitimate interests protected? Oversight must account for confidential business information, security and privacy.

These are essential distinctions. Knowing where a shipment of chips went does not, by itself, establish what a model can do or whether it is safe to deploy.

The limits of tracking compute

Computing power is a useful indicator of development scale, but it is not a complete measure of risk. Improvements in efficiency can enable more capable systems to be built with fewer resources. The risks of a deployed model also depend on its design, the tools it can access and the setting in which it is used.

Verification becomes harder when infrastructure spans jurisdictions with different laws and enforcement standards. Any international arrangement would need to address gaps in coverage and opportunities to evade restrictions.

There is also a competition concern: poorly designed compliance requirements could be easier for established technology giants to absorb than for smaller developers. A policy intended to constrain concentrated power should avoid reinforcing it unnecessarily.

For those reasons, compute oversight is better understood as one potential layer of governance, alongside model evaluations, security requirements, incident reporting and safeguards for deployment.

From trust to accountability

The strongest idea in the interview is not simply that technology bosses deserve scepticism. It is that public safety should be supported by institutions capable of checking conduct, even when the parties involved do not trust one another.

Tracking compute could contribute to that system, provided the rules are technically credible, enforceable and adaptable. It cannot settle every question about AI safety. But it offers a concrete starting point for a debate too often framed around whether powerful executives should be believed.

This article was inspired by 'I don't trust big tech' – AI whistleblower #AI #USA #China from Channel 4 News. Please visit the original video for the creator’s full presentation and context.