On Tuesday, September 29, the President and six technology executives signed a one page pledge on artificial intelligence. A reporter asked whether it binds anybody. He said he thinks it is morally binding.

You will never sit at that lunch. But what a safety standard is worth, and who gets to grade it, reaches your work too.

On Tuesday the East Room filled up. Technology executives sat down to lunch with the President and the Speaker of the House. By the end of the afternoon six of them had signed a document.

It runs about 300 words. One page.

Then everybody walked out to the cameras, and a reporter asked the obvious question. Does this bind anyone?

"I think it's morally binding," the President said. He also called it "almost like a constitution in a way."

Skip the argument about the name that followed.

Read the page.

Start with who signed, because the list is the story.

Six executives signed with the President. Sundar Pichai of Google. Dario Amodei of Anthropic. Mark Zuckerberg of Meta.

Greg Brockman of OpenAI. Elon Musk of xAI. Jensen Huang of Nvidia. Their companies build the country's frontier models, and the chips those models run on.

The document carries a long title: the White House Accord on Super Intelligence. Under that sits a subtitle, Joint Commitment on Frontier Responsibilities. It asks each company to run four layers of checks.

First, internal controls that watch what a model can do. Those run during training and after it ships, in areas like cybersecurity, biosecurity and chemical threats. Second, an internal team to make sure the controls work.

Third, an outside auditor to assess whether they work. Fourth, a committee of the board to read the reports and see that problems get fixed. The page also says the companies will meet regularly to set standards.

Read that list again. It is a reasonable list.

Now read the verbs. The page says the signers "believe each company should implement" the four layers. Not one sentence on it says "must."

The page names no auditor. It names no standard. No deadline, no filing, no penalty. Its last paragraph says that "over time, it may make sense to codify these steps into laws or regulations."

The Speaker described it the same way. Voluntary commitments on behalf of the industry, he told reporters. The President said he is seeing "tremendous self-policing."

Hours later he signed an executive order on a different subject. Agencies are now to say "Super Intelligence" and "SI" wherever the law allows. The order tells the executive branch it "will not acknowledge" the old terms. It leaves the legal definition where Congress put it. It gives the science adviser 60 days to write proposed legal language.

One more detail about the accord, and it matters more than it looks. It is not in the fact sheet the White House published with the order. It reached the public as a picture the President posted to social media.

None of this is new. That is the part worth sitting with.

Go back to July 21, 2023. A different President gathered seven AI companies at the same building. They announced voluntary commitments. The seven: Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI. Eight promises, under three headings called safety, security and trust. None enforceable.

Those promises were specific. Outside security testing before release, partly by independent experts. A channel for third parties to report flaws. Watermarks on generated content. Public reporting on what the systems can and cannot do.

Three months later came an executive order with the same aims. Trump revoked that order by name on January 20, 2025. It appeared on a list titled "Initial Rescissions Of Harmful Executive Orders And Actions."

The pledges outlived it.

A year later reporters went back to check, and found the companies doing more testing and more watermarking. The promises were still voluntary. Two administrations with almost nothing else in common reached for the same instrument.

The industry did consider something firmer this year. In July, Google DeepMind's Demis Hassabis proposed a standards body modeled on the one that polices stockbrokers. It would test frontier models before release. Anthropic, Google and OpenAI talked it over for weeks.

Zuckerberg, Musk and Huang opposed it. They argued it would hand too much power to the three largest labs. The proposal went no further.

The lighter version is the one on the page.

And Congress? Hearings, and still no comprehensive AI law.

Both sides of this have a real case. You ought to be able to say each one out loud.

Start with the best case for the pledge. These models ship in months. A statute takes years. The people drafting it would be writing rules for a thing they cannot yet describe.

A promise that exists today beats a law that does not exist at all. The four layers are close to what the serious labs already do. The page's last line says these companies are committed to them whether or not anyone requires it.

Add the part nobody on that side should have to whisper. Beijing will not slow down because Washington passed something. And ordinary law has not gone quiet. Fraud is still fraud.

A day after the signing, the Federal Trade Commission confirmed an industry wide probe into Anthropic and OpenAI. It plans formal demands for information and testimony from executives. Its chairman says Washington should use the laws it already has before writing new ones. The Wall Street Journal's editorial page argued the same case this week, and argued it well.

Now the other case, which needs nothing but the page itself.

The verb is "should." The company picks its own outside auditor. Nothing on the page says what makes an auditor independent. Nothing requires anybody to publish the findings. Nothing happens if a company stops.

At a Senate hearing the next day, Senator Richard Blumenthal called the arrangement "completely voluntary and totally secret."

Sort the two the way you would sort any argument. Ask what each one leaves standing.

The pledge leaves the promise, and the one who made it. The same company writes the control and hires the grader. It reads the report. It decides whether you ever hear a word of it.

That can work. It works the way a handshake works, which is to say it works until it doesn't.

A rule leaves something different behind. A standard somebody wrote down. A record. A door a stranger can walk through.

One of those depends on who is holding the pen. The other does not.

A week like this one is what principles over personalities actually tests.

You will never eat lunch in that room. This still lands on your table.

Think about the standards that your own work already lives under, the ones you never got a vote on. A license. An inspection. An audit, a certification, a code.

Every one of them has the same four parts. Somebody wrote the standard down. Somebody from outside checks it. The result goes in a file. You can read the file.

Layer three of the accord asks for an outside evaluator. One already exists. A research group has run exactly those reviews for two of the companies that signed. Its people testified at the same Senate hearing, and they described the terms. The companies had no obligation to bring them in, and none to report publicly what they found.

The accord has the first two of those four parts.

It is missing the last two, and that is not a small gap. A standard nobody can read is a standard nobody has to keep. "Morally binding" puts the whole arrangement inside the person who signed.

A conscience is a fine thing to have in a room. It is not a record. It leaves the building when the person does.

One of these systems will do something nobody intended, and nobody at that table will have planned for it. When that day comes, the question will not be whether anyone felt responsible. It will be narrower. What can you point at?

None of this calls for alarm. These companies may run all four layers and run them well. The ones most likely to do it well are the ones who showed up to sign.

The point survives either way. You would have no way to know, and no step to take.

So watch one word in the months ahead. Voluntary. The page allows that codifying these steps "may make sense" someday. The order already tells a staffer to draft the bill. Watch whether that word survives, and watch who asks for it to stay.

A promise is worth exactly what the person who made it is worth. That is the definition of a promise, and no insult to anybody who signed one this week. A law is what you reach for when that person is gone.

Take the names out and this week was about the distance between a promise and a rule. Six companies promised to check their own work. They may do it well. Nothing on the page makes them, nothing makes them show you, and nothing happens if they stop.

A pledge is held by the one who made it. A rule is held by all of us, including the next administration and the next set of executives. That is the only kind of safeguard that survives a change of heart.

Washington picked a new name for the technology this week. Not a new rule.

Mark November 28. That is the order's deadline for the White House science adviser to deliver proposed language defining the technology in law. The trade commission's probe runs alongside it. The question worth holding until then: does the word voluntary survive the drafting?

Trump Has the Right Instincts on AI
The case for the light touch: a voluntary accord beats a federal takeover of the industry.
Wall Street Journal, 2026-10-01

FTC Probes AI Giants as Issue Hits Midterms
A day after the signing, the trade commission confirmed an industry wide probe into the safety of AI labs.
Newsmax, 2026-09-30

Trump, AI titans sign 'morally binding' agreement to 'self-police'
A safety researcher told the Senate that companies now rely mostly on AI to monitor AI.
Fox News, 2026-09-30

We want your feedback. Tap the answer you agree with most, and your vote is counted.

John Braddock
Between the Lines
Stand tall. Stay disciplined.