Should superintelligent AI require a government license?

What would it mean to license superintelligence?

Licensing superintelligence would mean that a developer could not train, deploy, or both, above some line without a government’s permission, and the government could refuse, condition, or revoke that permission.

The hard part is the line. “Superintelligence” has no legal definition in force anywhere. In policy writing it usually means AI systems far more capable than humans across most tasks.

A license would also be a different kind of rule from current law. The EU AI Act, California’s SB 53 and New York’s RAISE Act require notice, published safety plans and incident reports, but no approval. See frontier AI laws.

So the debate is over a veto: whether a regulator should be able to stop a model before it is trained or released.

Who has proposed licensing frontier or superintelligent AI?

Proposals for licensing, or for an agency with comparable powers, have come from AI developers, researchers, U.S. senators, and the UN Secretary-General since 2023. One group goes further and calls for a prohibition.

AI developers

On May 16, 2023, OpenAI chief executive Sam Altman told a Senate subcommittee that he would form a new agency to license efforts above a certain scale of capabilities, with power to take the license away, Fox News reported.

On May 22, 2023, OpenAI published “Governance of superintelligence,” by Altman, Greg Brockman and Ilya Sutskever. It said the world would likely need “something like an IAEA for superintelligence efforts,” with authority over any effort above a capability or compute threshold to inspect systems and require audits.

Microsoft’s president, Brad Smith, called for licensing of powerful foundation models and a new agency the same month, in a Microsoft blog post.

Researchers

In July 2023, Markus Anderljung and more than 20 co-authors published “Frontier AI Regulation: Managing Emerging Risks to Public Safety.” They proposed standard-setting, registration and reporting, and compliance mechanisms, and considered “licensure regimes for frontier AI models.”

In October 2025, the Future of Life Institute published a Statement on Superintelligence. It calls for “a prohibition on the development of superintelligence” until there is broad scientific consensus it can be done safely and controllably, and strong public buy-in. That is a stronger position than licensing: a ban with conditions for lifting it.

Lawmakers and officials

On September 8, 2023, Senators Richard Blumenthal and Josh Hawley released a framework whose first element was licensing by an independent oversight body.

On July 18, 2023, United Nations Secretary-General António Guterres told the Security Council he welcomed calls for a new UN entity inspired by models such as the International Atomic Energy Agency.

On September 29, 2025, Hawley and Blumenthal introduced S. 2938. It would require developers of systems trained above 10²⁶ operations to participate in a Department of Energy evaluation program, and says no person may deploy such a system unless in compliance. The penalty is at least $1,000,000 per day. Within 360 days, the Department would have to recommend “licensing requirements” and other oversight. The bill also tasks the program with developing options, including “potential nationalization,” for managing superintelligence if it seems likely to arise.

On September 23, 2026, 26 state attorneys general asked Congress for mandatory federal safety testing of frontier AI and international cooperation to prevent harmful superintelligence.

What are the strongest arguments for a license?

Supporters argue that only a license creates a point where a regulator can stop a dangerous model before release, and that compute triggers and chip controls would make such a license enforceable against the few companies it would cover.

• Disclosure does not stop a release. Under SB 53, a developer can publish a framework, file a transparency report, and ship. A license adds a point where a regulator can stop a model before harm occurs.
• Some harms cannot be undone. SB 53 defines catastrophic risk to include expert help with chemical, biological, radiological or nuclear weapons and a model evading its developer’s control. Proponents argue that for those risks, review must come before deployment, not after an incident report.
• The developers asked for it. OpenAI’s 2023 post and testimony argued that efforts above a threshold need an authority that can inspect and audit.
• Compute makes a license enforceable. Frontier training needs large clusters of advanced chips, and exports of those chips to China and some other destinations already require U.S. licenses. That gives a regulator a physical point to check, unlike most software. See AI export controls.
• A narrow license spares almost everyone. A threshold at 10²⁶ operations would cover a small number of companies. Most developers and all ordinary users would be untouched.

What are the strongest arguments against?

Critics argue that licensing would entrench the largest companies, threaten open-weight releases, depend on triggers no one can define well, slow national developers against foreign ones, and lack the international body that would make it work.

• Regulatory capture. Andrew Ng, a co-founder of Google Brain, said in October 2023 that some large companies “would rather not have to try to compete with open source,” and warned that licensing burdens could squeeze out competitors, according to Open Data Science’s report of his interview with the Australian Financial Review.
• Concentration of power. A Mozilla-coordinated joint statement on October 31, 2023, signed by researchers and company leaders, warned that “quickly rushing towards the wrong kind of regulation can lead to concentrations of power.”
• Open-weight models. A license on training or release could make publishing weights impossible above the threshold. In July 2024, the Commerce Department’s NTIA concluded the government should not restrict the wide availability of model weights at that time and should monitor the risks instead.
• No agreed definition. A license needs a trigger. Compute thresholds have no published rationale and drift as methods improve, as Sara Hooker argued in 2024. Capability tests come only after training. See compute thresholds.
• Competition between countries. Executive Order 14365, signed in December 2025, ties U.S. AI leadership to a “minimally burdensome national standard.” On that view, a national permit system would slow U.S. developers while developers elsewhere proceed.
• No international body exists. The IAEA model depends on a treaty. In August 2025, the UN General Assembly created an Independent International Scientific Panel on AI and a Global Dialogue on AI Governance in resolution A/RES/79/325. Neither has power to license, inspect or block anything.

How do the two sides compare on the key questions?

The disagreement comes down to five questions.

Is there a middle position?

Yes, and most enacted law already sits there: disclosure and evaluation duties, without a veto.

Transparency laws such as SB 53 and the RAISE Act require developers to publish safety frameworks and report incidents, without approval. The EU requires notice and evaluations for models above 10²⁵ operations. Export controls license the chips, not the models.

S. 2938 is a step further: a mandatory evaluation before deployment, with the license question deferred to a later recommendation. The FRONTIER Act (H.R. 9925), introduced in the House in July 2026, would add independent audits and would preempt new state obligations on AI developers in the areas it covers.

Some critics accept this middle ground. Ng said he was not against regulation and that any framework needed a thoughtful approach, according to the same report.

The open question is whether disclosure turns into permission once a model shows dangerous capabilities in testing. No enacted law answers that yet.

What would a superintelligence license need to work?

Drawing on the proposals and objections above, a working regime would need at least six things: a legal definition or trigger that does not go stale, a regulator with technical staff able to evaluate frontier models, a way to cover models trained outside the regulating country, a rule for open-weight releases, protection against the license becoming a moat for incumbents, and a process to revoke or condition a license after release.

Each of those requirements answers one of the objections above. A proposal that cannot meet one leaves that objection standing.

Where does the licensing debate stand?

As of October 2026, the 2023 call for an AI licensing agency has not produced one in any country. It has produced notification rules, transparency laws, chip export licenses and a pending Senate bill with a deployment bar.

Two questions still divide the sides. One is whether disclosure and testing will catch dangerous capabilities in time without a veto. The other is whether a veto can exist without becoming a moat for the companies that already lead.

This guide summarizes positions; it does not take one, and it is not legal advice. For the vocabulary, start with AI licensing explained.