Not as of October 2026. No jurisdiction requires a general government license before a frontier AI model is trained. China’s pre-launch filing system comes closest. The EU, California, and New York rely on notification and transparency duties instead, and the United States controls frontier AI mainly through export licenses for chips.
“Licensing” an AI model can mean a government permit, a registration or filing, a notice to a regulator, an export license, or the private license that governs a model’s weights and training data. This site covers all of them and links its sources, using the primary text wherever it can be opened.
Apart from China’s pre-launch filing, they regulate disclosure more than permission. Pick a jurisdiction to see its trigger, core duty and maximum penalty.
In force; enforcement from Aug 2, 2026
In force Jan 1, 2026
Takes effect Jan 1, 2027
In force since Aug 15, 2023
In force Jan 22, 2026
No AI bill
No statute; EO 14365
Most frontier AI rules decide who is covered by counting the operations used to train a model. Enter a model’s size and training data, or its hardware and run time, to see where it lands. Open the full calculator.
This is a rough estimate. Estimates by outsiders can be off by a factor of two or more, and the laws put the duty on the developer, who knows its own numbers. California counts fine-tuning and later modifications toward the total. New York and California add a revenue test, so the heaviest duties fall on companies above $500 million.
Crossing a binding line triggers notice, documentation or safety duties. The 1023 line is non-binding guidance, and S. 2938 is a bill, not law. None of the enacted laws listed requires a government license before training a frontier model. Thresholds and statuses are stated as of October 2026. General information, not legal advice. Sources and method: compute thresholds.
The interactive calculator needs JavaScript. The method is simple: training compute ≈ 6 × parameters × training tokens. The EU’s systemic-risk line is 1025 floating-point operations; California, New York and South Korea use 1026. The full table is in compute thresholds.
7 guides to both meanings of the word: the rules governments apply to the most capable models, and the licenses that govern models and data. Each states its facts as of the date at the top.
The six things “licensing” an AI model can mean, and which ones actually exist in law today.
The EU, California, New York, China, South Korea, the UK and the U.S. federal government, side by side.
Why lawmakers count floating-point operations, how the count is estimated, and where the method breaks down.
The one place the U.S. already requires permission for frontier AI: export licenses for chips and, on paper, model weights.
Apache 2.0, MIT, the Llama license, Gemma terms and OpenRAIL, compared clause by clause.
The public content deals behind AI models and the court rulings that set their price.
Both sides of the licensing debate, with named sources for the main arguments.
Much of the current debate traces to May 2023, when OpenAI’s chief executive told a US Senate subcommittee that a new agency should license AI efforts above a capability threshold. Since then, lawmakers have mostly chosen lighter tools: compute thresholds that trigger reporting, safety frameworks that must be published, and incident reports. The policy debate guide sets out the arguments on both sides.
It is a free, sourced reference. It does not lobby for or against any rule, it is not written by a law firm, and it is not legal advice.