No single government can unilaterally slow AI development, because training a frontier model requires only chips, electricity, data and researchers that can be moved across borders — but specific jurisdictions can and do slow it inside their own territory, and the European Union's AI Act is the clearest working example.
The honest answer is that slowing AI is a question of jurisdiction-by-jurisdiction leverage, not one global switch. If you want to know who actually holds that leverage, the answer is: legislatures that control chip exports, regulators that control deployment inside a market, courts that control liability, and the labs themselves — in roughly that order of real-world force.
The mechanism: why slowing AI is different from banning a chemical
Most safety laws work because the dangerous thing is physical and hard to hide. A chemical plant is a fixed building; a drug trial needs patients; a nuclear program needs enrichment facilities you can see from space. AI training is different.
The core inputs are graphics chips, electricity, training data and a small team of researchers. None of those are unique to AI, and all of them move easily. That is why the most effective slowdown levers are not bans on "AI" but controls on the specific bottlenecks — advanced chip exports, data-center power permits, and the liability rules that decide who pays when a model causes harm.
Chip export controls are the sharpest of these because the most advanced training chips are made by a very small number of firms, so a single country's export rules can reach well beyond its own borders. Power permits matter too: a frontier training run consumes enormous electricity, and local grid operators can simply say no. Liability is the quiet lever — if a court rules that a model's output is the deployer's legal responsibility, companies slow down on their own to avoid lawsuits.
A concrete example: the EU AI Act's risk tiers
The European Union's AI Act is the most concrete real-world attempt to slow or shape AI by law, and it works through risk categories rather than a blanket ban. Systems used in hiring, credit scoring, education and law enforcement are classified as high-risk, which triggers obligations around documentation, human oversight and conformity assessment before deployment.
General-purpose models face separate transparency and documentation duties. The practical effect is not that AI stops — it is that deploying a high-risk system in the EU becomes slower and more expensive, because you must produce evidence before you ship. Compare that with the United States, where the approach has leaned more on executive orders, agency guidance and voluntary commitments from labs rather than a single statute.
The UK has taken a third path, favoring existing regulators over new AI-specific law. Three major jurisdictions, three different speeds — and a company can often choose which market to launch in first. That is the trade-off in one sentence: strict rules slow deployment inside a market, but they also push development to whichever jurisdiction is most permissive.
A worked example of a compute threshold
Many proposed rules use a compute threshold as the trigger for extra scrutiny, measured in floating-point operations, or FLOP — a standard unit for counting how much math a training run performs. Imagine a rule that says: any model trained using more than a set amount of compute must register with a regulator, publish a safety evaluation, and report serious incidents.
A small startup fine-tuning an open model on a single rack of chips falls below the line and faces no new paperwork. A lab training a frontier model on tens of thousands of chips crosses it and must file reports before release. The threshold is the whole design: set it too low and you crush small developers and academic labs; set it too high and the largest, most capable models slip through untouched.
This is why compute thresholds are politically contested — they are a dial, not a wall, and whoever sets the dial decides which side of the line most developers land on.
The limits: what this advice does not cover
Legal slowdowns have real blind spots. A jurisdiction can regulate deployment inside its market, but it cannot easily stop a model trained elsewhere from being used via an API, and open-weight models released publicly cannot be recalled. Enforcement also lags: passing a statute takes years, while a model generation can ship in months.
Smaller countries have almost no leverage at all unless they host a chokepoint like advanced chip fabrication or a major data-center cluster. And voluntary lab commitments are exactly that — voluntary — so they hold only as long as leadership chooses to honor them. If you are trying to reason about a specific proposal, ask three questions: what is the trigger, who enforces it, and what happens to a developer who simply leaves?
If the answer to the third is "nothing," the rule mostly reshapes where work happens rather than whether it happens. That is the uncomfortable core of AI governance, and no amount of legislation changes the physics of a technology that travels as easily as an email.