Before anyone closes a patient, the room counts.

Two people count sponges, needles, and instruments aloud against a written record. If the count comes up short, nobody closes. The surgeon does not close. The attending does not close. The case stops, the room goes quiet in a particular way, and everybody starts looking, because a sponge that is not on the table is somewhere worse.

I learned this as a Navy Hospital Corpsman, working in surgery. And the thing that struck me then, the thing I have carried into every job since, is who gets to call it.

The scrub tech and the circulating nurse, not the surgeon. The lowest-ranking people in the room hold the authority to stop the highest-ranking person in the room, and the reason is not fairness. The reason is line of sight. The surgeon is inside the patient. The scrub tech is looking at the table. Whoever can see the problem first gets the power to halt, or the power is useless.

That is the whole design, and it has nothing to do with courtesy. Risk shows up in one person's field of view before it shows up in anyone else's, so the system gets built around that person or it is theater.

I thought about the count all week.

The industry called it one.

On the weekend of September 12, Anthropic CEO Dario Amodei published an essay arguing that the companies building frontier AI should slow down together. He called it pacing the frontier. His proposals were specific: bring in independent evaluators who can look at safety, establish common safety standards across companies, and verify that anyone claiming to follow them actually does.

Sam Altman agreed publicly and said OpenAI would bring in independent evaluators too. OpenAI also delayed its expected public listing, which should get your attention, because delaying an IPO is not a press release. It costs money.

Elon Musk, Satya Nadella, and Demis Hassabis all broadly agreed over the same weekend, without committing to anything in particular.

And in the reporting around all of it, an Anthropic researcher named Evan Hubinger said he personally puts the odds that AI kills every human being on earth within the next decade at better than one in ten.

Read that sentence again and notice where it came from. That is a man describing the product he goes to work on.

The scrub tech called a count.

The following Saturday

On September 19, President Trump posted on Truth Social that he is creating an AI Force, modeled on the Space Force he stood up in his first term, and that he will name an AI czar shortly.

The post said the government "will not in any way hinder or stifle the Growth of this incredible Industry." He said bad behavior would be handled through, in his words, "our already existing Criminal and Civil Justice System." He called the worry that AI could destroy humanity a hoax. He said the only guardrail the situation requires is a strong and smart president.

He also said AI could become the next Industrial Revolution and could account for as much as 25% of the country's GDP.

Put a number on that, because nobody else did. The American economy runs around 29 trillion dollars. Twenty-five percent of it is roughly 7.25 trillion. Every dollar this country spends on health care (every hospital stay, every prescription, every premium) added up to 5.3 trillion in 2024, which was 18% of GDP and made it the largest thing we buy. Manufacturing came in at 9.4% at the start of this year. So the claim on the table is that a single technology ends up worth more than American health care. That could happen. It has never happened; he posted it on social media, and no one was in a position to ask him to show his work.

The count got called. The answer from the top of the building was to keep cutting.

Except that is not the whole government

Here is where it gets strange, and where I think most of the coverage is flattening something worth seeing.

While the White House was announcing a new force whose stated job is to protect the industry from interference, Congress was moving in the opposite direction. Dozens of AI bills have now been filed. Senator Bernie Sanders and Representative Greg Casar introduced legislation to ban artificial superintelligence outright and temporarily pause advanced AI development. States are also being actively pushed out of the room. An executive order signed in December 2025, titled Ensuring a National Policy Framework for Artificial Intelligence, stood up a Justice Department task force to challenge state AI laws in federal court, directed the Commerce Department to identify state rules it considers overly burdensome, and conditioned 42 billion dollars in broadband funding on states declining to enforce certain AI regulations. That is not a government sitting this one out. That is a government using real leverage to ensure oversight happens.

The industry leaders and researchers developing AI, referred to as "the builders," asked for brakes, and the government produced three answers at once. Ban it. Regulate it. Clear the field and handle the wreckage in court afterward.

Handling harm through the existing criminal and civil justice system means you find out afterward. It means the sponge is already inside the patient and now we are in litigation about whose sponge it was.

An AI Force is a real thing to announce. A czar is a real position to fill. But announcing an oversight body whose declared purpose is to prevent oversight is a sentence I had to read four times.

We have run this play twice before

None of this is new. It is the third time an American industry has stood at this exact fork, and the first two went the other direction.

In 1968, Jack Valenti built the film rating system, with four categories: G, M, R, and X. The Motion Picture Association's own history says he created it "amid mounting calls for censorship and the specter of government intervention." Local censorship boards were cutting films city by city, and Washington was circling. Valenti's answer was to rate the movies before somebody else did it for him.

In December 1993, the Senate held hearings on violent video games, driven by Mortal Kombat, Night Trap, and Doom. Senators Joe Lieberman and Herb Kohl ran them, and Kohl put the choice in one sentence: "If you don't do something about [content ratings], we will."

The industry moved fast. It announced a ratings system hours before the December 9 hearing even opened. Seven companies formed a trade association by April. The Entertainment Software Rating Board went live on September 13, 1994, nine months after Kohl said that out loud.

Look at the shape of both. The government said Regulate yourselves or we will, and the industry, facing a credible threat, built a rating body in months.

Now reverse every arrow. In September 2026, the companies building the most powerful technology on earth asked, publicly and in writing, to be evaluated by outsiders and held to a common standard. The branch of government that would have to enforce any of it replied that it "will not in any way hinder or stifle" them.

Valenti had to be pushed. Amodei volunteered. Volunteering turns out to be worth less.

Both of those worked because somebody refused to carry the product

A rating is a piece of paper, and paper does not stop anything. What stopped things was the shelf.

Take Basic Instinct in 1992. The ratings board looked at what Paul Verhoeven turned in and gave it an NC-17. He went back eight times and cut roughly forty seconds to get it changed to an R, and those same seconds were restored for the director's cut that premiered at Cannes that May. Nobody was arrested, and no law was broken. An NC-17 in 1992 meant theater chains wouldn't book the picture and newspapers wouldn't run the ad, so three letters from a private board cost a studio its distribution. When the longer version reached home video, it came out as an unrated director's cut, which tells you the studio feared the letters, not the footage.

In 2005, someone found that Grand Theft Auto: San Andreas had shipped with a sexual minigame sitting in the code, disabled but present on the disc. The mod that surfaced it was called Hot Coffee. And the voluntary system that most people assumed was decorative did something nobody expected.

The ESRB revoked the game's M rating and re-rated it Adults Only. Retailers pulled it off the shelves. Take-Two suspended production and recalled it. The company shipped a patch, stripped the code, and reissued the game as an M. The FTC settled with Take-Two and Rockstar in June 2006 over advertising a rating the disc did not match.

That is the whole machine working. A private body re-rated a product, the product came off the shelf, and the company ate the cost. No legislation passed, and none had to.

Which is the condition AI does not have

The count in an operating room works for three reasons. Everyone counts the same items the same way. Somebody outside the room checks the record. And when the stop gets called, the case actually stops.

Amodei proposed the first two. Independent evaluators are the outside check, common safety standards are the shared count, and to his credit, he put both in writing.

The third one is the problem, and the two rating systems are exactly why.

Film and games each had a chokepoint. Between the studio and the audience stood a theater chain. Between the publisher and the kid stood Walmart, Target, and GameStop. When a rating changed, those companies refused to carry the product, and that refusal was the enforcement. The rating boards never needed police power because retail already had it.

A frontier model has no shelf. It ships from the company that built it, over an API, to anyone with a credit card, in every country at once. No exhibitor can decline the booking, and no buyer at Target can pull the pallet. Take the chokepoint out of the ESRB, and what you have left is a logo.

So a lab can slow down, a competitor can decline to, and nothing anywhere in the distribution path punishes the second lab. That is the structure, not a moral failure on anyone's part. Self-regulation without an enforced stop is a promise, and promises are worth exactly as much as the cost of breaking them.

The case against the people I just made sound noble

A company asking to be regulated is not automatically acting against its own interest. Independent evaluations, compliance verification, and common safety standards are expensive. Expensive rules are survivable for a lab with billions in backing and fatal for a startup with eleven people. Amodei may be entirely sincere and the proposal may still, as a side effect, pull the ladder up. Both are possible at once, and neither cancels the other.

There is also a straight-faced argument for the White House position that deserves better than a sneer. If American labs pace themselves and Chinese labs do not, the pacing did not reduce the risk; it relocated it. A unilateral slowdown in a race with more than one runner is a transfer, not a brake. You cannot wave that away, and the people who do are usually the ones who have never had to hold a position against a competitor who does not share it.

What I will not extend the same patience to is the word hoax. You can argue that the risk is overstated. You can argue the timelines are wrong. Calling it a hoax means claiming the people raising it are lying, and the people raising it this month are the CEOs of the companies building it, plus their own researchers, several of whom are putting numbers on the odds. Somebody in that picture is wrong. Nobody in that picture appears to be running a con.

What I would actually do with this

If you run something small, this week changed nothing about your Monday, and I am not going to pretend otherwise. But it changed what you should expect, and there are four things I would do.

Assume the rules arrive late and arrive unevenly. With the White House leaning one way, Congress leaning another, and the states fighting for the right to act at all, you won't get one clear rule you can build against. Write down now which AI systems touch your customers, what data goes into them, and who at your company signed off. You are building that record for a question somebody asks you in 2028.

Find out who in your building can stop the work. Whatever tool you have running, someone has line of sight on it that you do not. Whoever that person is, they need standing authority to halt it without asking permission, and they need to know they have it. A stop that requires escalation is not a stop.

Expect the rule that binds you to come from a company, not from Congress. That is the lesson of both rating systems. What disciplined the studios was a theater chain declining to book the picture. What disciplined Rockstar was Walmart clearing the shelf. If you build on somebody's AI, your real regulator is your vendor's terms of service, your app store's review policy, and your payment processor. Read those the way you would read a statute, because they will change faster than one.

Watch what companies spend, not what they say. OpenAI delayed a public offering. That is the most informative fact in this entire story, because everything else cost nothing to say. Apply that filter to every AI vendor pitching you this quarter. Statements are free.

Last thing.

The count is not really about sponges. Any competent surgeon knows roughly what went in. The count exists because the room agreed, in advance and in writing, on who gets to say stop, and it agreed while everyone was calm and nobody was bleeding.

That agreement is the product. The counting is just the part you can see.

Right now, the people closest to this technology are telling us the count is off. Congress wants to ban the procedure. The states are arguing about whether they are even allowed in the room. And the White House has announced that it is standing up a new force whose job, as described, is to make sure nobody stops the surgery.

We are going to settle who holds the authority to call a halt. The only question left is whether we settle it before or after somebody closes.

John M Pogue writes about how systems fail the people inside them. Curiosity is my superpower

References

The announcement

The industry's call for a slowdown

Congress and the states

The GDP figures

The rating systems

Further reading

For anyone who wants to follow a thread beyond what this piece covers.

On whether self-regulation forecloses regulation. Brown v. Entertainment Merchants Association, 564 U.S. 786 (2011). The Supreme Court struck down California's law restricting violent video game sales to minors on First Amendment grounds. A working voluntary rating system is part of the backdrop. It is the strongest argument that rating yourself well enough can permanently take the legislative option off the table.

On the chokepoint, and how the industry routed around it. In the pic-to-DVD shift, 'unrated' rates high, Variety, 2004. Studios learned to ship unrated home video rather than carry the NC-17 letters, and when both versions sat on a shelf together, the unrated one took 80 to 90 percent of sales. Worth reading alongside the claim that retail enforcement held.

On where state AI law actually stands. Where State AI Legislation Stands Halfway Into 2026, Tech Policy Press.

On what Congress is and is not doing. AI threats confront a Congress far from erecting guardrails, Roll Call.

On the money underneath all of it. The Questions You Should Be Asking About the AI Bubble, Harvard Business Review, September 2026.

Originally published on LinkedIn. Comments and the conversation live there.