Bullet Trains without Tracks
On Fable 5 and regulating infrastructure
On Friday evening, the United States stopped a piece of software at the border. At 5:21 PM Eastern on June 12, the Department of Commerce issued an emergency directive ordering that Anthropic’s newest models, Claude Fable 5 and its unredacted foundation Mythos 5, be made unavailable to any foreign national, anywhere in the world. So, Anthropic pulled the models offline entirely. Software that had been writing code for people in Lagos and Lyon and São Paulo a few hours earlier simply went dark.
This event has refocused discussion around a larger question, one that has been building for years and that I think we are only now learning to ask correctly. The question is usually phrased as should we regulate AI? I think that phrasing hides the real issue. The better question is: what kind of thing are current AI systems, such that a government would reach out and recall it the way you might recall a faulty car? My thought is that current AI is part of an infrastructure of information. Not just a product, and not a tool in isolation. It is the newest part of a type of infrastructure we have been building, mostly without calling it as such, for a while now. And infrastructure, if history rhymes, is governed in its own peculiar and bloody way.
The thing we already built
When we say a system is infrastructure, we usually mean that it has stopped being optional. Electricity, once a marvel you paid to see at a fair, became the thing whose absence we call an emergency. Roads, water, the grid: these are not products you choose among so much as the floor you stand on. When they fail, it is not a bad purchase, it is a crisis.
Information technology has quietly become this kind of floor. The internet, the world’s data, the technologies we build on top of them, social media, web search, and now large language models, together form a single, continuous thing. I’d call it, simply, information infrastructure. It runs in an unbroken line from the telegraph, undersea cables, Starlink satellites, through the World Wide Web to the model that wrote you an email this morning. Most nations now depend on it the way they depend on the power grid, and it is welded to the power grid besides: data centers in the US already draw something like four percent of the country’s electricity, and that share is climbing. The trains, it turns out, need a lot of coal.
The US government did not treat Fable 5 as a consumer product that had disappointed its customers. It treated it as a strategic asset, a piece of the national substrate, a thing whose flow across borders was a matter of state. Whether or not that response was wise or measured, it was the response you give to infrastructure, not to an app.
The railways
To understand how a society absorbs a new kind of infrastructure, it helps to look at a recent time that we did it from scratch: trains.
I find the railway a convincing metaphor for what is happening now. Things map pretty precisely. The internet is the track, the fixed, physical bed over which everything else must run. Data is the fuel that’s fed into the engines. And the information technologies, the search engines, the social platforms, and now the AI models, are the trains themselves. We’re at the point of making new high-speed machines that carry us places in this information infrastructure, and that also occasionally jump the rails.
The early decades of rail were a glorious, dangerous mess, and they do correlate well with our present. In the nineteenth century there was no agreement on something as basic as how far apart to lay the two rails. Companies chose their own track gauges on purpose, so that a competitor’s train physically could not run on their line; at the boundary between two networks, every passenger and every crate had to be unloaded and hauled onto a different train. We are living through our own gauge wars right now. We just call them walled gardens, proprietary APIs, vendor lock-in, data silos. The startups that ported their whole business to Fable 5 and woke up Saturday to find the track torn out beneath their new train car learned the nineteenth-century lesson the nineteenth-century way.
And the trains exploded back then. Before anyone had invented the standardized signal, the automatic air brake, or the federal inspector, steam locomotives blew their boilers and ran headlong into one another with appalling regularity. When a model today hallucinates a fact into a legal brief, leaks the code it was supposed to protect, or helps tip a market into a flash crash, we are watching a boiler explosion. We have built the bullet train before we have finished laying the track, and we are trying to install the brakes while the thing is already moving at speed. This is not a new predicament. It is the normal predicament of new infrastructure.
What did we do about the exploding trains? This is the part everyone hoping for sensible AI policy needs to understand. The safety of the railways was not foreseen and legislated in advance by farsighted lawmakers. It was, in the grim and accurate phrase, written in blood. Synchronized time zones, mandatory air brakes, signal blocks, the public utility commissions that finally wrested control from the railroad barons: every one of them arrived after a disaster made it politically impossible to do nothing. Infrastructure regulation has historically been reactive. It is a scar, not a vaccine.
The story of trains does not end with rigid rulebooks, in the style of the EU AI Act. As rail networks grew too complex and too fast for any inspector to anticipate every failure, the most advanced regulators stopped trying to dictate the exact bolt. They moved instead to what’s sometimes called meta-regulation: the state sets broad, demanding safety goals and then requires each operator to build, run, and submit to audits of its own internal safety management system. The regulator stops writing the manual and starts demanding that you keep one, and prove it works. If AI governance is going anywhere good, it is going there.
Where the metaphor halts
Of course, AI models aren’t trains and data is not coal. Coal is gone once you burn it; that’s the whole economy of fuel. Data is non-rivalrous. The same dataset can train a model, feed a recommendation engine, and run an analytics pipeline all at once, and none of those uses leaves any less for the others. Any policy built on the intuition that information is scarce in the way coal is scarce will be wrong in ways that compound.
The AI trains keep rebuilding themselves, too. A locomotive, once it leaves the factory, does not quietly rewire its own engine on the basis of the scenery it passes. A deployed model does something uncomfortably close to that: its behavior shifts with updates pushed from upstream, sometimes overnight, sometimes invisibly. If a hospital or a defense network or a tax authority has woven such a model into its daily work, it cannot fully verify what the thing will do tomorrow, because tomorrow’s version may not be today’s. This is what the Fable recall exposed: developers found their pipelines silently rerouting mid-task from one model to a weaker one, billed at mismatched rates, the floor shifting under them without a sound.
Finally, the information tracks have no borders. A railway is welded to a place. You can stand at the frontier and watch the gauge change. Information infrastructure has no such geography. A model trained in California answers a student in Nairobi with no border to cross and, until last Friday, no passport to show.
On the good side, infrastructure that ignores borders can spread with incredible speed and generosity. An open-source model released on a Tuesday can be running in a hundred countries by Thursday, in the hands of people who could never have the resources to build it. This is the genuinely hopeful face of the thing, and it’s why I’ve argued that this kind of software tends, over time, toward being free and open. Infrastructure wants to be shared; that is how it becomes infrastructure.
On the bad side, infrastructure that ignores borders can be imposed. When the tracks are laid by a handful of very powerful actors, everyone else ends up living on someone else’s railway. In much of the world, Facebook and WhatsApp are not apps you might choose; they are, for practical purposes, the internet, the storefront, the marketplace, the post office, all owned by one company a continent away. Quiet adjustment to an algorithm can decide whether a local merchant eats. Google Maps is not a convenience; it is the layer that decides which neighborhoods are visible and which delivery routes exist. When you depend on infrastructure you don’t own, the owner can change it, price it, or switch it off, and you will only find out the moment they do.
This is why infrastructure is, in the end, a sovereignty question, and why governments are finally waking up to the sovereignty of information infrastructure. They have been circling it for years, mostly around data. Europeans will recognize the long saga of Schrems, cases brought by an Austrian lawyer that twice struck down the legal arrangements for shipping European data to American servers, killing Safe Harbor, then the Privacy Shield that replaced it. That fight was fundamentally a country asking who controls the tracks its citizens’ lives run on.
Countries, unsurprisingly, have very different means of answering such questions. Europe has reached for comprehensive, rights-based rulebooks. The GDPR, and now the AI Act with its pyramid of risk, ban some uses outright, hem in the dangerous ones with audits and human oversight, and demand that makers of big foundation models disclose what they were trained on. These laws are bureaucratic, slow, and this sometimes means Europeans get features last. But they are laid down in advance, on principle. On AI, China seems to be taking a state-centered path: models that face the public must pass state security review and register their algorithms before launch, their training data vetted for political alignment, while industrial AI is subsidized and sped along. And the United States, characteristically, has no single law. It is fast and unpredictable. It is the only nation that I know of with a power similar to the Federal Trade Commission’s power of algorithmic disgorgement, the authority to order a company to delete an entire model and its training data. A death penalty for software. The US has the fastest trains, the biggest train factories in the world, but maybe the most reactionary regulatory means. And at the federal level it is simultaneously degrading its own fuel by hollowing out public data, statistical agencies, and the research apparatus from which good models are built.
The border drawn through the train
Which brings us back to Friday, and the Fable recall.
It was the first time a government drew a hard border not around the data, the chips, or a data center, but through a live model, reaching into a running train and sorting passengers by passport. That it was done by an administration unusually fond of borders is not a coincidence; it is a trademark reflex to treat a strategic asset as something to be walled. The stated trigger was a security audit by Amazon researchers who coaxed Fable into finding and explaining software vulnerabilities. Anthropic’s own response was that the capability was nothing special (“widely available from other models”) and used every day by the defenders who keep systems safe. Reading the two accounts side by side, it is hard not to conclude that the recall was at least partly reactive: a show of force, with some existing friction between this administration and this company.
But I don’t want to let Anthropic off the hook either, because the company invited this action. Anthropic has marketed its models on the premise that they are too dangerous, that frontier AI is so perilous it requires government oversight, please and thank you. You cannot spend years insisting your locomotive might level the city and then act shocked when the city’s inspectors show up and impound it. And the specific shape of the Mythos and Fable launch sharpens this: Mythos was presented as so hazardous it had to be wrapped, redacted, fitted with real-time guardrails that silently reroute any dangerous question to a tamer older model, and then sold. To declare a thing an extraordinary danger, in vague and unfalsifiable terms, and then ship a guardrailed version of that very danger for money, is not the conduct of a careful engineer. It is the conduct of a marketer. It is hype manufacturing, and the recall is, in part, the bill coming due.
Laying better track
What heartens me in this story is that, today, millions of people take trains that don’t explode. We’ve arrived at this situation because of two things that the information infrastructure, and not just AI, needs as a whole.
The first is regulation over the whole infrastructure that treats it as such. History is rather clear about what that means: it will arrive late, it will arrive after harm, and the best version of it will not be a rigid manual but a demand that the builders prove they are managing their own risks. That is the meta-regulatory bargain rail eventually struck, and it is the realistic destination for AI as well. Sudden embargoes drawn along passports are not that. They are panic, a recall of exploding boilers, and not a new brake.
The second thing, the one that gets less attention because it isn’t anyone’s job to legislate, is responsible engineering. The railways were made safe not only by inspectors but by engineers who agreed on a gauge, published their standards, and stated their failure modes plainly so the next builder could design against them. Openness and clear, shared standards did as much to civilize the railway as any law. The equivalent for AI is not secrecy dressed up as caution. It is the honest, specific, checkable description of what a model can and cannot do, the kind of transparency that lets organizations, academics, and regulators study a system and hold its makers accountable, rather than guess at it through a marketing deck.
These are the rocky early years of a new kind of infrastructure, and we should expect them to be rocky; every infrastructure’s early years are. But we are not the first people to lay a new kind of track across the world, and we are not without a map. The trains came roaring out ahead of the rails once before, and we caught up by finally agreeing, in the open, on how to build the line. We can choose to remember that before the explosions. The bullet trains are already moving. The remaining question is whether we finish the track before they reach the end of the line.