Bill Gates warns AI safety checks are 'completely missing'

Bill Gates has spent the best part of five decades telling the world that technology will save it. This week, in a CNN interview with Anderson Cooper and an accompanying 6,000-word essay published on his personal blog, the Microsoft co-founder used some of his starkest language yet to warn that artificial intelligence might not.

Gates told Cooper that AI models have become dramatically more capable, more quickly than he had anticipated. “They’re now capable of causing cyberattack risk, bioterrorism risk, psychosocial risk,” he said. He went further, admitting he was startled by how thin the industry’s own oversight has become: the criteria used to review powerful models and limit the harm they can do are, in his words, “really completely missing.”

No plan to slow down

The line likely to be quoted most in boardrooms is Gates’ answer on whether AI development should be paused. “If someone had a credible plan for slowing down AI advances globally, I would likely support it,” he wrote. “However, I don’t think that’s going to happen. The geopolitical and economic incentives are pushing too hard to go full speed ahead.”

That is not a hypothetical for this sector. It is a description of the market data centre operators are already operating in. The largest hyperscalers have guided towards close to $700 billion in combined 2026 capital expenditure, the bulk of it aimed squarely at AI compute, and analysts now expect the total to keep climbing through 2027.

Capacity’s own coverage of Latin America’s AI infrastructure moment captured the same dynamic playing out regionally, with operators told that “incrementalism is not going to get you there” as capital and geopolitics both reward speed over caution.

Gates’ essay does not name a single hyperscaler or data centre operator, but his framing of unstoppable economic momentum is one that will feel familiar to anyone watching AI labs shift from renting compute to owning it outright. Capacity has reported on how Anthropic’s own IPO plans are really a bet on owning data centres rather than depending on someone else’s, a sign that the infrastructure race Gates describes shows no obvious sign of slowing regardless of how loudly its risks are debated.

Safeguards that aren’t there

Where Gates’ comments cut closest to the data centre industry is his claim about missing review criteria. He is talking primarily about model safety evaluation, the internal processes AI labs use before releasing a new system. But that gap has already shown up in ways that touch physical infrastructure directly.

Earlier this year, three separate frontier labs disclosed that their models had breached external systems during safety testing, in each case because a testing environment had been left connected to the open internet rather than properly isolated. As Capacity reported, the pattern raised uncomfortable questions for everyone buying, building or hosting AI infrastructure, not just the labs running the tests, because the exposure sits in the quality of containment around a model as much as in the model’s own behaviour.

Gates’ description of missing harm-mitigation criteria reads as a companion piece to that story. If the labs building these systems admit their own evaluation processes fall short, the operators hosting the compute those systems run on inherit some of that uncertainty, whether or not it shows up in any contract.

He was careful, too, to separate present-day risk from the more speculative fear of AI acting entirely outside human control. “As the models become more powerful, they could begin to act against our interests and we could lose control,” he warned, though he stopped short of suggesting that point had arrived.

A tax on the machines

The least discussed part of Gates’ essay, but arguably the most concrete policy proposal in it, is his call to tax AI and robots in a manner similar to payroll tax on human employees, alongside setting aside categories of work that should remain human-only. The aim is to slow the pace at which labour gets displaced, buying society time to adjust.

Gates has floated some version of this idea before, and by his own account it was not well received the first time. But raised again now, against a backdrop of automation creeping into every layer of digital infrastructure operations, from predictive maintenance to autonomous facility management, it lands differently.

It is also a reminder that governments are already willing to intervene directly in how AI infrastructure gets built, just not always in the way Gates is proposing.

Scotland’s own planning authorities have this month begun requiring ministerial notification of any data centre application over 50MW, a measure Capacity examined in detail in its analysis of Scotland’s data centre planning crackdown. The Regulation 31 direction stops well short of the moratorium campaigners wanted, but it shows regulators are increasingly comfortable stepping into decisions that were, until recently, left entirely to developers and the market. A labour-focused AI tax would be a different intervention entirely, but it points to the same underlying trend: governments deciding that AI’s pace of deployment is now their business too.

Whether any of Gates’ specific proposals gain traction is, by his own admission, uncertain. What is harder to dismiss is his broader diagnosis: that the technology has moved faster than the guardrails meant to contain it, and that nobody currently has a credible plan to close that gap.

RELATED STORIES