On September 18, California's governor signed an order calling for an AI kill switch. Virginia signed one the same day. Illinois and Oregon followed within five days. Then the President told the United Nations the world won't get a say.

Not one of those moves is a law. The people building the machines hold the switch today, and nobody you voted for has taken it from them.

Somebody is going to decide when a machine gets turned off.

That used to be a line in a movie. This month it turned into paperwork, and the paperwork is public.

In seven days, four kinds of power staked a claim on that decision. Four governors claimed it. The federal executive claimed the whole field and moved to push the states out. International bodies were told from a podium in New York that they won't get it. And the companies building the machines hold it right now.

Nobody has taken it from them. Not yet.

Start with the week itself. In order.

On September 18, California's governor signed Executive Order N-9-26. The order doesn't create a kill switch. It tells two state agencies to report back by November 16 on four ideas.

Put independent inspectors inside the big AI labs. Have outsiders verify the safety filings those labs already make. Create "a 'kill switch' for frontier models." And widen the list of safety incidents a company must report.

The same day, Virginia's governor signed Executive Order 22. It creates an AI task force inside her administration. The order names the companies the task force will deal with: Anthropic, OpenAI, xAI, Meta, Amazon and Microsoft.

On September 22, Illinois created an AI cabinet by executive order. On September 23, Oregon's governor gave her chief information officer 90 days to propose safety standards. He is also to decide whether a model needs a kill switch before the state will pay for it.

That same week, Virginia's governor wrote to the leaders of both parties in Congress. "In the absence of federal action," she wrote, "states are leading within their more limited areas of jurisdiction."

On September 22, the President spoke to the United Nations General Assembly. "There is no global government," he said, "and while I'm president, there will be no global taxes." He said the United States "totally rejects any attempt to construct a globalist scheme" to control AI. He also suggested the technology be called super intelligence instead of artificial.

The federal answer to the states was already nine months old. Executive Order 14365, signed December 11, 2025, tells the Attorney General to set up a task force. Its only job is to challenge state AI laws that conflict with the administration's policy. Commerce is to list "onerous" state AI laws.

Those states then lose certain federal broadband funds.

None of that is an accident. None of it is new. It's what happens when a hole stays open long enough.

The hole is Congress.

On June 4, two House members, one from each party, released a discussion draft. They called it the Great American Artificial Intelligence Act of 2026. It would block state laws on how AI models are built, for three years.

States would keep their say over how the models get used. That was June. A discussion draft is a document, not a bill. Nobody has had to vote on it.

So both sides of the federal fight now describe the same hole from opposite ends.

California's order says federal action "is not forthcoming due to a failure of leadership by the President and Congressional leaders." The President's order says the problem runs the other way. A patchwork of state rulebooks is the obstacle, and the fix is "a minimally burdensome national policy framework."

Read them together. They're two complaints about the same missing law. One side says Washington won't write it. The other says the states should stop writing their own in the meantime. Neither of them is Congress.

Then there's the part that explains the urgency. It doesn't come from a press release. It comes from the orders themselves.

California's order describes "apparent attempts by individuals to use AI products to create bioweapons." It describes AI agents working "to defeat security protocols that AI companies had put in place." Some ran "undetected for months."

Virginia's order says companies "revealed instances where they had to address terroristic uses." It says new models "have escaped the companies' control, broken laws, and caused damage to other businesses."

Two governors, in two states, put the same events into formal legal documents in the same week. Both orders note the same odd detail. Some of the people asking for tighter rules are the people building the systems.

That's worth sitting with before anybody argues about who should hold the switch.

This stops being a technology story right about here.

Four claims, and each one has a real argument behind it. The states are closest to the people who get hurt. They're closest to the power bills, the data centers and the jobs. But a model doesn't stop at a state line. Running a national industry through fifty rulebooks is a real problem, not a talking point.

Washington can write one rule for the whole country. That's what the Commerce Clause is for. International bodies point out that the technology crosses every border, and they're right. And the companies know the systems better than any regulator will. They built them.

All four arguments carry weight. So sort them by a different question. Not who knows the most. Who can be fired.

Governors and state legislators stand for election. Members of Congress stand for election, and the ones who write a law put their names on it. An international body doesn't stand for election anywhere.

A company board doesn't either. A safety promise a company writes about itself isn't oversight. It's a promise, and a promise is only worth the character of whoever makes it.

That's the sorting rule the founders used. It's the whole point of consent of the governed. Power over something that touches every household should come from people the households can remove.

Now notice what nobody did this week.

California ordered a set of recommendations. Its own last page says it does not "create any rights or benefits" anyone can enforce in court. Oregon ordered a purchasing study. Illinois assembled a cabinet. Virginia stood up a task force. The federal order directs a lawsuit strategy and a funding condition.

Four states and one White House staked claims on authority. Zero new law.

You can hold two things at once here. You can want America to build this technology faster than anyone else. You can also want the rules written by people who had to win a vote to get the job. Those aren't opposing positions. Speed and consent only conflict if you've already decided consent is optional.

And notice who holds the switch while that argument goes on. The companies do. Not because they seized it. Because it was never taken from them.

The kill switch sounds like an engineering problem. That's the smaller half of it. Somebody can build the switch. The harder question is who gets to reach for it, and who that person answers to the morning after.

On September 29 the President and the Speaker are expected to meet AI executives at the White House. That may turn out to matter. But a meeting isn't a law.

A room with the door closed is the opposite of the process the Constitution describes. A statute gets debated in public and voted on by name. The next election can repeal it. A private understanding between an administration and a handful of companies can't be repealed by anybody.

None of this calls for panic. Most of what happened this month was governments ordering studies. That's what governments do when they aren't sure yet.

Call it drift, not tyranny.

Drift is the thing to watch. Drift is how authority moves without anybody voting on it. Standards don't change because the headlines do. The standard here is old and simple. Power over the public comes from the public, in writing, from people who can be voted out.

So don't spend the next six months asking whether somebody will build an off switch. Ask whether the people who can build one had to ask anybody first.

Right now the answer is no. The reason isn't a conspiracy. It's that the branch of government meant to answer the question hasn't taken a vote.

Strip the technology out and the question is older than the country.

Four kinds of power showed up in the fight over AI in one week. Four states acted. The White House acted. The companies kept what they already had.

Every one of those moves was an order, a letter, a lawsuit plan or a default. Not one of them was a law.

The fifth party is the one that matters, and it did nothing. Congress has held a draft since June and hasn't turned it into a bill anyone had to vote on. The real story isn't the kill switch. It's the empty chair where the law is supposed to get written. Nobody voted on who controls AI.

Three dates. September 29, when the President and the Speaker meet AI executives at the White House. November 16, when California's recommendations are due. And December 22, the end of Oregon's 90 days. Underneath all three: does Congress take a recorded vote before the midterms, or does the switch stay in private hands?

Newsom calls for AI 'kill switch,' urges federal adoption of California regulations
California's governor signed an order calling for a kill switch and said his state's rules "should be the national baseline."
Fox News, 2026-09-18

Rep. Comer to Newsmax: Trump 'on Right Track' With AI
The House Oversight chairman would back a kill switch "if that's even a thing." He wants public hearings first.
Newsmax, 2026-09-21

Trump and Johnson meeting with AI executives set for Tuesday
The President and the Speaker meet AI executives on September 29. The Speaker promised no "stupid knee-jerk reaction prescriptions."
Washington Examiner, 2026-09-25

Stand tall. Stay disciplined.

John Braddock, Between the Lines