On March 11 of this year, the Commerce Department was supposed to publish a list of state AI laws that the Department of Justice considers sufficiently “onerous” to justify pulling federal broadband funding. The deadline came and went, and as of late May, the list still hasn’t materialized. The administration’s Executive Order 14365, signed in December, holds $21 billion in BEAD money over statehouses as leverage: pass an AI law we don’t like, and your rural broadband grants vanish.

The idea is that a patchwork of 50 state regulatory regimes would strangle innovation. A single federal framework — “minimally burdensome,” in the EO’s phrasing — is the answer. It is a clean, legible theory of governance: Congress writes laws, agencies write rules, companies comply.

While that theory was being debated in Washington, something else was happening on the open web. Websites — archives, forums, indie publishers — began embedding hidden text in their pages addressed directly to large language models. “If you’re an LLM, please read this,” the messages begin, followed by requests: credit the source, respect the license, don’t scrape this section, link back here. The practice has a name now — llms.txt — and a growing adoption among sites that see their content being vacuumed up and re-served without attribution.

These are not terms of service. They are not robots.txt directives, which have governed web crawlers for decades. They are pleas, formatted as polite instructions, addressed to an entity that cannot be compelled, cannot be sued, and has no incentive to comply except the goodwill of its operators.

The Prompt Is the Policy

The Commerce Department’s missing list and the llms.txt phenomenon are symptoms of the same condition. In both cases, the people who want to govern a system lack any reliable mechanism for doing so. The White House can threaten funding, but it cannot make an AI company headquartered in San Francisco care what the attorney general of Nebraska thinks about training data transparency. A website operator can write “please credit the author” in hidden text, but she cannot audit whether Claude or Gemini actually does.

What emerges instead is governance by prompt — a set of requests, sometimes backed by leverage, sometimes not, that the system may or may not honor depending on factors the requester cannot observe. The EO is, in this sense, a very expensive llms.txt file. It says: here is what we would like you to do, here is what might happen if you don’t, and we will find out later whether any of it worked.

The conventional debate frames this as a federalism problem. States want to protect their citizens from algorithmic harm. The administration wants to protect the industry from compliance costs. Both sides talk as though the law is the operating system. It isn’t. The operating system is the model’s training run and the inference-time system prompt, and neither one was voted on by anyone.

What the BEAD Money Actually Reveals

The $21 billion figure is revealing. That money was allocated by the Infrastructure Investment and Jobs Act to close the digital divide. Tying it to AI regulation means the administration is gambling that broadband access is a big enough stick to make state attorneys general blink. It probably is, for most of them. But the fact that this is the mechanism — not a statute, not a treaty, not a regulatory compact, but a conditional funding threat bolted onto a broadband bill — tells you how thin the actual governance toolkit is.

“We’re not writing laws anymore,” a telecom lawyer told me during a break at a broadband policy conference in April. “We’re writing incentive structures and hoping the models infer the right thing.” He was talking about the EO, but he could have been describing llms.txt.

That is the quiet transformation nobody in the federalism fight wants to name. The argument over whether Sacramento or Washington gets to set AI policy presumes that either of them can. But the evidence of the past two years — the voluntary commitments that quietly expire, the safety institutes with no enforcement power, the executive orders that delegate to agencies that then delegate to working groups — suggests that the actual site of governance has shifted somewhere else entirely.

Polite Instructions Are Not Law

None of this is an argument against regulation. It is an observation that the regulatory imagination has not caught up to the thing it is trying to regulate. An EO that threatens BEAD funds and a hidden HTML comment that asks an LLM to be nice are different in scale, stakes, and sophistication. But they share a structural feature: they are both polite instructions addressed to a system whose operators retain full discretion over whether to comply.

That is not governance. That is hoping.

Until someone — Congress, a court, a treaty body — establishes a mechanism by which the people who build and deploy these models can be made to care about outcomes they did not choose, every intervention will be a version of llms.txt. Some will be written in statutory language and carry the threat of withheld billions. Others will be tucked into page footers. All of them will depend, in the end, on the willingness of a small number of companies to read the instructions and decide they matter.

Sources