The most consequential general-purpose technologies of the last eighty years all came out of governments. The bomb. The rocket. GPS. The internet. Even recombinant DNA was paused by its own developers and rapidly enclosed in a federal framework before commercial deployment.

AI is the first to break that pattern.

The frontier of artificial intelligence isn’t being pushed by a national lab, a defence ministry, or a multilateral programme. It’s being pushed by ten or so private companies — OpenAI, Anthropic, Google DeepMind, xAI, Meta, Microsoft, and a smaller cluster of Chinese labs operating with state encouragement but not as state instruments. Governments are paying for safety institutes after the fact. Regulators are publishing white papers about systems that have been running for two years.

This is a genuinely new situation, and it deserves to be examined on its own terms rather than collapsed into the usual story of “tech goes too fast, government catches up.” Every previous dangerous technology was built by the state, and that mattered.

For the rest of us: why “built by the state” matters

When governments build a powerful technology, three things tend to be true at once.

Containment is the default. The bomb wasn’t sold to enterprises in 1946. The internet wasn’t a consumer product in 1973. A government-built technology arrives behind a wall of classification, controlled procurement, and explicit national-security framing. It diffuses slowly because containment is the design assumption.

Governance arrives with the technology. The Atomic Energy Act was passed in 1946, one year after Hiroshima. The Outer Space Treaty was signed in 1967, ten years after Sputnik. The Asilomar Conference paused recombinant DNA research in 1975 — two years after the first successful experiment — and the NIH guidelines followed almost immediately. None of these regimes were perfect, but none of them were ten years late.

The state has a doctrine. Even bad doctrines are doctrines. Mutually assured destruction was terrifying, but it was a stable equilibrium. Nuclear non-proliferation treaties have held for sixty years. There is something governing the technology’s use, not just guidelines about its development.

AI has none of those three properties. It is being built by entities whose default is mass deployment, governed by frameworks written years after release, and shaped by no doctrine more coherent than each lab’s individual interpretation of “responsible scaling.”

How we got here

A short timeline is useful.

In 2015, OpenAI was founded as a non-profit dedicated to ensuring that artificial general intelligence benefits all of humanity. In 2019, it created a capped-profit arm to raise capital. By 2023, the same organisation was the commercial frontier of the field. Anthropic split off in 2021 around an even more explicit safety mandate; within a few years it was raising multi-billion-dollar rounds at one of the highest valuations of any private company. Google folded DeepMind into its product organisation. Meta open-weighted models that matched closed-lab quality from a year prior. xAI raised six billion dollars in a matter of months.

None of this happened with state coordination. The compute came from private cloud providers. The capital came from venture funds and corporate balance sheets. The talent came from academia, which the labs out-paid for graduate researchers. The training data came from the public web, scraped without permission and litigated about afterwards.

By the time the EU’s AI Act passed in March 2024, GPT-4 had been live for a year and Claude 3 had been released the same month. By the time the US executive order on AI was signed (October 2023), models capable of triggering its thresholds were already deployed. The UK’s AI Safety Institute, founded in 2023 and renamed the AI Security Institute in 2025, operates as a post-hoc evaluator: it tests models the labs voluntarily share with it. China has the most state-directed approach of any major jurisdiction, but the actual development still happens inside private labs — DeepSeek, Alibaba, ByteDance — operating with state encouragement rather than state direction.

The pattern is consistent: technology first, governance second, with a multi-year gap.

Could the state just take it over?

This is the question that gets posed every time the gap widens. Could a government simply nationalise frontier AI development, the way the US nationalised uranium enrichment in the 1940s?

In principle, yes. In practice, the gap between principle and reality is enormous.

The state would inherit talent it cannot retain at federal salary scales. It would inherit compute it does not own — the largest clusters sit inside Microsoft, Google, Meta, and Amazon. It would inherit a development culture that is intentionally distributed across thousands of small decisions in small teams, not a Manhattan-Project structure with a single director and a perimeter fence.

There is also the problem of which state. Nuclear weapons could be contained because three or four countries had them at the start and the rest could be policed by treaty. Frontier AI labs operate across multiple jurisdictions with branches in still more. The compute is portable. The weights, once released, are uncontainable. The closest historical analogy isn’t the bomb. It’s the printing press.

A nationalised AI programme would also lose what the private competition currently produces: the parallel exploration that has driven most of the safety advances of the last three years. RLHF, constitutional AI, mechanistic interpretability, model welfare — these came out of competing private labs publishing against each other. A monopoly state lab would have less of that, not more.

So the answer to “should the state take over” is probably: it can’t, and even if it could, the historical record on monopoly state research programmes is worse than people remember.

The two bad options

What’s left is the choice the field is actually making, which is between two unsatisfying futures.

Slow down, and risk falling behind. Every Western pause is a relative gain for whoever doesn’t pause. The most explicit version of this argument lives in the China comparison: if frontier development is slowed by Western governance, the global frontier shifts to a jurisdiction whose state has no commitment to the safety frameworks the West is building. This is not hypothetical. DeepSeek R1 in early 2025 demonstrated that the gap between frontier and follower labs is months, not years.

Charge ahead, and govern later. This is the current path. Build the systems, deploy them, hope the governance layer catches up before the harms compound. The optimistic case is that the technology is genuinely dual-use and the productive applications outpace the dangerous ones. The pessimistic case is the one Harari and others keep making: that we are deploying systems whose internals we don’t fully understand into institutions we depend on, and the cost of being wrong scales with capability.

Neither option is good. The honest framing is that we are running an unprecedented experiment in which the regulator is two cycles behind the regulated, in real time.

What this means for the rest of us

For anyone deploying AI in an enterprise context, the practical implication is that the governance layer you can rely on is the one you build yourself. The state’s frameworks will arrive late, will be partial, and will be written for systems that have already moved on by the time the rules ship. The EU AI Act is governing GPT-4-era systems. By the time enforcement begins in earnest, the models in production will be three generations newer.

This isn’t a counsel of despair. It’s a counsel of seriousness. The fact that the state did not build this technology does not mean the state will never shape its use. But it does mean that, for the period we are now in, the people closest to the technology — the labs, the deployers, the engineers, and the boards that fund them — are the governance layer. There isn’t another one. Not yet.

That is a heavy responsibility to place on a small number of private companies. It is also, factually, where we are.

References

  • US Atomic Energy Act (1946); Outer Space Treaty (1967); Asilomar Conference on Recombinant DNA (1975); NIH Guidelines for Research Involving Recombinant DNA (1976)
  • EU AI Act, Regulation (EU) 2024/1689 (March 2024)
  • Executive Order 14110 on Safe, Secure, and Trustworthy AI, White House (October 2023)
  • UK AI Safety Institute (renamed AI Security Institute, February 2025), founding mandate (November 2023)
  • Anthropic, “Responsible Scaling Policy” (2023, updated 2025)
  • DeepSeek R1 technical report (January 2025)
  • Yuval Noah Harari, Nexus (2024) — on dependence on non-deterministic systems
  • Open Philanthropy and Centre for the Governance of AI working papers, 2024–2025