The Leash They Cut to Fit
1,200 of the people who build AI are asking the state for a way to slow it down. The fight was never about the brake. It is about who gets to set the length of the leash.
On July 28, 2026, a letter titled "Pacing the Frontier" went out with 1,268 verified signatures from inside the frontier AI labs. It does not ask for a pause. It asks Washington to build the technical and governance infrastructure for a verifiable slowdown, to be triggered "if and when" AI moves faster than humans can oversee it. Within hours OpenAI and Anthropic signed off at the company level. Among the names: Dario Amodei, Jared Kaplan and Jack Clark, and for OpenAI Jakub Pachocki and Mark Chen. On paper it is the most responsible gesture of the year, the insiders asking to be held in check. Anyone who has walked a dog knows where the power in a leash sits, and it is not in the hand. It is with whoever sets the length. The hand can only pull once the dog has already reached the end; the length decides in advance how far the dog runs before anyone feels the tug. A two-metre leash and a twenty-metre leash hand the same person two different powers, and the hand does not choose which.
The letter asks the state to be the hand. Build the mechanism, hold the end of the leash, be the one who pulls. The length stays with the companies, because they define the two numbers that matter. The first is the threshold: what "faster than humans can oversee" measurably means. The second is the reading of the instrument that says whether the threshold was crossed. Whoever sets the threshold and reads the sensor decides what is being controlled. Washington keeps the final gesture, the most visible and least useful part: an off-switch it gets to pull once the speed has already been set by the people it is meant to check.
Take the other side, because it is real. The companies are the only ones who can measure the machine's speed today. They see new capability first, months ahead of anyone in Washington, and a pacing mechanism written by people who cannot read the technology would be theater. So the competence to calibrate the control does sit inside those buildings, and pretending otherwise is naive. It does not follow that the calibration belongs to them. It follows that the expertise should feed a threshold set by someone else and verifiable by someone else. The moment the measure, the threshold and the instrument all sit on the same side, the brake stops braking and turns into a certification, issued by the runners, that they are running at the right speed.
There is a second-order effect the safety framing never names. A verifiable control built now does not slow the future in the abstract: it freezes the present standings. Whoever already leads gets a global apparatus that makes every rival's sprint costly and suspect, and gets to shape that apparatus to its own tolerances. The leash is cut to fit the runner who holds it, and in the same motion stitched tighter onto whoever comes next. It is the oldest move in rulemaking, the one under every call for a global safety authority: the limit is drawn by the party it constrains least.
None of these names is acting in bad faith, and that is exactly the point: the mechanism works even when the intentions are good. You judge a control system by where the measure, the threshold and the reading end up, whatever sits behind it. In "Pacing the Frontier" all three end up in the same place, and that place is not the state.
This reaches you if you work with the technology and do not build it. The question you will be sold in the coming months is "are you for or against slowing AI down," and it is the wrong one, because almost all of us agree on it. The question that holds the power is who sets the length. Every time someone hands you a brake, ask who set the length of the leash before it reached your hand. The brake is the part you see. The length is the part that counts.