There Is No Arrival Date

Strip the vocabulary away and the argument is about five concrete things. How fast should these systems be built. Should anyone need permission to train or deploy one. Who pays when a system causes harm. May the weights be published. How much electricity may the whole enterprise consume.

Those are ordinary questions with separate answers. A person could favor strict liability and open weights at the same time without contradiction.

That is not how they are answered. They are answered as a bloc, and in the United States the bloc now has a party. Democrats say AI, and want a brake: licensing, audits, a slower road. Republicans say superintelligence, and want the road cleared. Tell me which word someone uses and I will tell you how they vote. That is not a description of a technology. It is a uniform.

Each word smuggles a theory of arrival

“AI” names a product. Something built, shipped, procured, bounded. A vendor on one side, a buyer on the other, a specification in between.

“Superintelligence” names a threshold. Something that does not exist, will exist later, and after which the world is different. A prophecy with the date left blank.

The words point in opposite directions and share one premise. Both assume the thing arrives. One puts the arrival in the future. The other puts it in the past. Neither describes a system that is still becoming something on the day you are trying to govern it.

Why nobody wins

Start with the side that wants to slow down. Their word asks for the wrong remedy. A product can be certified, and certification is the appeal of the word. It is also the trap. Certify a model and you have described it on one day, while the weights update on a cadence you do not see, the tools accrete, and the scaffold around it changes. They will win their licensing regime and receive a certificate that is stale the morning it issues. The harm will arrive through the update nobody inspected. Their word makes them fight for the wrong instrument and then call it a victory.

Now the side that wants to accelerate. Their word writes the brief against them. Nothing recruits a regulator like a prophecy. Announce that something more capable than its makers is coming, and the public will eventually ask the state to stop it, and the state will reach for the bluntest tool on the shelf. Every deck that promises superintelligence is a page in the other side’s filing. The threshold framing also demands a race, and a race is the one condition under which no firm can afford the slow internal controls that would make its systems reliable.

Then the loss they share. Both instruments aim at a fixed object. The object is not fixed. Whichever camp wins the vote, the winning instrument does not reach the thing it was written for, and the public is handed the appearance of a settled question with none of the substance.

And because the words are party property, nothing built survives. A rule with a party label has a half life of one election. Install a regime, lose a cycle, repeal it. The systems keep updating through every reversal, indifferent to all of it.

Worst of all is what goes unasked. Who is liable when an agent commits a payment. What an audit of a continuously updating system actually measures. Whether authority to act should be tested at release or at the moment of effect. Those questions sort nobody into a camp, so nobody’s standing is served by raising them. They are also the only questions whose answers will change how these systems behave.

Nothing settles

The reason both instruments miss is not political. It is physical.

Consider what moves in a deployed system. The weights update on a cadence the operator chooses. The tools accrete: a model that could only produce text in one quarter holds a payment credential in the next. The scaffold changes, and the same weights inside a different scaffold are a different actor with different reach. The retrieval corpus shifts underneath. The other parties it deals with are themselves changing.

There is no version whose behavior is a fixed fact. There is no state to certify.

This is not immaturity. A system that stopped changing would be a worse system, and the operator who froze one would lose to the operator who did not. Perpetual emergence is not a phase before maturity. It is an engineering commitment, renewed daily, under competitive pressure, by design.

This is the pacing problem, and it is not new. I argued more than a decade ago, beginning with Dynamic Regulation for Innovation, that stable rules cannot track moving objects, and that the remedy is not faster rulemaking but a different instrument. What is new is that the object no longer pretends to hold still between inspections. Earlier technologies outran the law. These systems change while the audit is being written.

The solution

Two parts, and the first is the name.

Call it emerging intelligence. The word describes what is actually true: these systems are perpetually emerging, by design, and there is no arrival to prepare for or to claim has already happened.

It needs one guard. “Emerging” ordinarily promises a destination. Emerging markets become markets. Emerging technologies mature. Strike that reading. These systems emerge perpetually and emerge toward nothing. The participle is the whole claim, and no noun waits at the end of it.

The name has a second property that matters as much as its accuracy. It cannot be worn. It does not tell you who is speaking. Anyone using it has given up the ability to signal a side and is left owing the listener something specific, which is the condition under which the unasked questions finally get asked.

The second part is the instrument, and it follows from the name. Stop certifying states and start supervising processes. We already know how to do this. We do it to banks, exchanges, and carriers, because we accepted long ago that a bank is a process and not an object. Nobody licenses a bank once and walks away. We watch what it does, continuously, and we test its authority at the moment it acts rather than at the moment it opened. Applied here that means one change above all: authority is tested at the point of effect, not at release, an argument I set out in Authority Must Hold Through Commit.

Continuous supervision has no party. It is not a brake and it is not an accelerator. It is the only shape of instrument that fits an object that never stops moving, which is why both camps can adopt it without conceding anything they actually care about.

The trajectory of these systems is an institutional choice and not a natural fact, an argument I have made in AI Evolution Is Mechanism Design. Naming was the first of those choices. We have now made it badly twice, once from each side, and both sides are losing to their own word.