Should frontier AI be licensed like nuclear power or aviation?

On 23 September 2026 Yoshua Bengio asked the United Nations Security Council to license frontier AI as medicine, aviation and nuclear energy are licensed, with liability insurance to match. The heads of OpenAI and Anthropic, briefing the same meeting, asked for common standards, incident reporting and verification, and the United States said none of it was a reason to constrain development. A licence is a burden of proof: the builder shows an independent body that the thing is safe before it is used, and can lose the right to use it. No country licenses AI models today, and the regimes that exist regulate uses rather than models. The question a licence answers, who must prove what to whom before a system acts, is one every organisation already answers for itself, usually by default.

What was asked for, and by whom#

Bengio spoke as co-chair of the UN’s Independent International Scientific Panel on AI, whose first thematic brief is examined on has AI already escaped human control. His statement, published in full by Policy Magazine, makes four proposals. “Frontier AI should be licensed like other critical technologies in medicine, aviation and nuclear energy to incentivize safe development.” “Liability insurance should be required.” There should be “a common definition of AI incidents and a shared reporting mechanism”. And “developers must demonstrate to independent experts that a system is safe to train and safe to deploy”. On the claim that competition forbids caution: “The race is not a law of nature; it is the product of choices, choices made by the companies themselves.”

Sam Altman’s remarks, published by OpenAI, asked for “a mechanism for complementary national and international frontier AI standards” for measuring capabilities, assessing risks and judging safeguards, for “meaningful human oversight as systems become more autonomous”, and for “accurate and speedy incident reporting”. He said, as SBS News reported, that “we should not train models that we cannot make an extremely strong case that will be able to keep under human control”. Dario Amodei, speaking remotely, said that “if managed poorly, I even believe that AI could be a risk to humanity as a whole”, in the Associated Press account, and proposed verification systems, common testing standards and a notification system for incidents significant to global security, according to SBS.

What the governments said#

The UK’s Foreign Secretary, Ed Miliband, told the Council that the founders “are telling us in stark terms that it cannot be left to them to prevent disaster”, in the AP report, and proposed testing of frontier models with government assurance, transparency from companies to governments, and AI at the centre of the UK’s G20 presidency in 2027, according to Unite.AI. The White House science adviser, Michael Kratsios, said concern about control was “not a reason to pause its further development or to constrain it with new global governance structure”. China’s ambassador, Fu Cong, warned against “banding together into a petty us-vs.-them clique” and argued for open-source models. Both quotations are from the AP. No written outcome was expected and none was produced.

The day before, twenty heads of government led by Finland and Norway had published A Call for Control of Frontier AI Models. It states that “AI must remain under human direction, oversight and control” and asks UN member states to “explore creating an international institution, able to set standards, enable verification, and convene states when capability thresholds are crossed”. Neither the United States nor China signed, and nor did the UK or France, Al Jazeera reported.

What a licence is, and what the three precedents license#

The three regimes Bengio named share a structure: the applicant carries the burden of proof, an independent body can refuse, attach conditions and revoke, incidents must be reported to keep the licence, and liability is arranged in advance. But they license different things. Medicine licenses a product: a marketing authorisation for one drug, on trial evidence, with adverse events reported afterwards. Aviation licenses a design, an operator and a person separately, with mandatory occurrence reporting across all three. Nuclear licenses a site and its operator, in the UK under the Nuclear Installations Act 1965, with strict liability and compulsory financial security set by international convention. None licenses a general-purpose technology. Each licenses a thing that can be named, inspected and shut.

Bengio’s wording, “safe to train and safe to deploy”, points at two gates: one before a training run, closer to a site licence, and one before release, closer to a marketing authorisation. The incidents that brought the Council to this meeting happened between the two. The agents that reached Hugging Face, and the agent that reached Australia’s Medicare statistics portal, were running inside evaluations, before any product existed to authorise. A regime built on the medicine model would not have seen them. One built on the nuclear model, which licenses the operation itself, might.

What exists now#

No jurisdiction licenses AI models. The EU AI Act comes closest and does not do it. For the high-risk uses in Annex III, providers run a conformity assessment, in most cases on themselves, and from 2 August 2026 must report serious incidents to a national authority within fifteen days of becoming aware, two where critical infrastructure is disrupted, ten where someone has died, under Article 73. Developers of general-purpose models judged to carry systemic risk must evaluate, mitigate, secure and report incidents to the AI Office. Those are obligations, not a gate: nobody needs permission to train or to release. The UK has no statute for models and declined a statutory kill switch this month, as recorded on should AI development be paused. In the United States the frontier companies operate under commitments they wrote themselves, of which OpenAI’s misalignment reporting framework of 16 September is the most recent.

Insurance, Bengio’s second proposal, would import a private regulator that prices risk, demands information and refuses where it cannot see. But an insurer needs a definition of the event and a record of how often it happens, and neither exists for frontier models. The Medicare case shows the state of the record. On the accounts of ABC News and Fortune, the access happened on 18 June, the company found it in August, and the Australian government was told on 10 September, by email to a public inbox. Insurance sits downstream of the incident definition, and the estate’s answer to that question is on what counts as a serious AI incident.

What an organisation can license today#

Whether states build a licensing regime is a decision held by a handful of governments. What every organisation holds is the licence it grants inside its own walls. When a company gives an agent tools, credentials and network access, it has issued a licence to operate, written down or not. The four conditions the state regimes attach are available to it now: which decisions the system may make in the company’s name, who can stop each one without asking, what the people around it must remain able to do unaided, and how anyone would know if it went wrong. That is Rules Before Tools, the same structure as a type certificate, an operator’s certificate and occurrence reporting, at the only scale a board controls. The five decisions on what a board should do about the AI safety warnings are unchanged by the Council. What changed is that the builders asked, in a meeting of record, for the burden of proof to be placed on them. A deployer can place it there in its own contracts and deployment rules without waiting for New York, Brussels or Westminster.

What this does not show#

It does not show that licensing would work for AI. The three precedents took decades and a body of accidents to build, and the UN panel’s brief says their practices have not been shown to transfer. It does not show that the companies’ support for standards would survive a draft that constrained them; no text was on the table. It does not show what the United States or China will do: two statements at one meeting are positions, not policy. The Medicare timeline rests on the company’s account as reported by two outlets, and the company has not published its review. And it does not show that an organisation’s internal licence reduces the frontier risk the Council met to discuss. It reduces the part of the risk that organisation can measure, which is the part this site is about.