The Name That Governs

That is the origin of the phrase now embedded in the statute books of the European Union, the United States, China, and most of the OECD. It was not a description. It was a boundary marker in an academic turf dispute, chosen so that a young assistant professor would not have to attend meetings with a famous man he found difficult.

The obvious response is that this is a charming footnote. It is not. The name went on to do the single most consequential thing a word can do inside a legal system, which is to determine jurisdiction.

The Statutes That Cannot Use Their Own Word

Start with the tell.

The European Union's Artificial Intelligence Act — the most comprehensive AI statute in force anywhere — has to define its subject in Article 3. Here is what it says an AI system is: "a machine-based system designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments."

Now United States law. 15 U.S.C. § 9401(3), enacted through the National AI Initiative Act of 2020 and inherited from a 2019 defense authorization: artificial intelligence means "a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments."

Read both again. Neither uses the word intelligence. Neither uses think, understand, know, reason, or learn. Both describe a device that takes inputs and emits outputs which affect the world — which is, definitionally, a machine. The two most important legal definitions of artificial intelligence on earth are definitions of automation, wearing a different name on the door.

This is not sloppiness. It is the opposite: it is what happens when careful drafters, who must produce text that survives litigation, are handed a term of art that will not survive litigation. They keep the label because the label is what got the bill scheduled, and they write around it because the label cannot bear operational weight. The gap between the two is the whole subject.

And that cuts both ways, which the argument has to concede before it goes any further. If the name is what got the bill scheduled, then the counterfactual to the AI Act is not a labor-framed AI Act. It is no AI Act. Complex information processing — the term Newell and Simon preferred, and a considerably more accurate one — does not get a committee, a commissioner, or floor time; it gets a paragraph in a standards document. Nobody convenes a summit about complex information processing. Nobody runs for office on it.

So the claim here is not that a better name was sitting on the shelf and the field grabbed the wrong one. It is that a trade was made and never recorded as a trade: the word that made the legislation politically possible is the same word that fixed the register the legislation would be written in. Urgency was bought with a metaphor and paid for in jurisdiction.

Naming Is Jurisdiction

The reason the label matters more than the definition is that in a legislature, and in a bureaucracy, the name arrives first and everything else is downstream of it.

The name determines which committee gets referral. The committee determines which staff draft the text, which witnesses are called, which trade associations are in the room, and — decisively — which existing body of law the new statute is grafted onto. By the time anyone gets to the definitional section, the architecture has already been chosen. The definition is fitted to the architecture, not the reverse.

Watch this happen in Europe. The AI Act is not a standalone civil-rights instrument. It was legislated inside the New Legislative Framework — the EU's product-safety architecture, built from Decision 768/2008/EC and Regulation 765/2008, which governs some seventy-odd regulations covering non-food goods on the internal market. Under the AI Act, a high-risk system must undergo conformity assessment and bear a CE marking before it can be placed on the market. The same machinery, the same notified bodies, the same declaration of conformity used for personal protective equipment, in-vitro diagnostics, and elevators.

So Europe named it an intelligence and regulated it as an appliance. Annex III of the Act classifies AI used in employment — recruitment, task allocation, performance monitoring, termination — as high risk. That is not a labor-law provision. It is a product-safety provision about a tool used in workplaces, and it produces the obligations product-safety law produces: documentation, risk management systems, human oversight design, technical files. It does not produce the obligations labor law produces: bargaining, consultation, consent, a share of the gains.

The informed objection arrives here, and it is a good one: Brussels had no choice. The AI Act is built on Article 114 of the Treaty on the Functioning of the European Union — the internal-market harmonization power — because that is the competence the Union actually holds. In social policy its role is supporting and complementary, and Article 153(5) TFEU excludes pay, freedom of association, and the right to strike from the Union's reach outright. The bargaining-and-consent version of this statute was not available to the European legislator no matter what the technology was called. On that reading the CE marking descends from the Treaties, not from a 1955 turf dispute, and the strongest section of this essay is also its most attackable one.

The objection is correct, and it is not sufficient, for two reasons. First, competence explains the instrument Brussels reached for; it does not explain the reaching. The Union was not obliged to legislate at all on a matter where its available instrument produced product law and the member states hold the labor power. Something made this urgent enough to legislate with the tool at hand rather than leave it to twenty-seven parliaments — and the thing that made it urgent was the name.

Second, and more to the point: the objection does not travel. The United States has no such constraint. Congress writes product law and labor law with the same pen, and the National Labor Relations Board's general counsel had opened the algorithmic-management question by 2022. Given a free choice of register, the American statute went to the same place — § 9401(3), the definition that also declines to say intelligence, sits inside a research-and-initiative act, administered through the National Institute of Standards and Technology, producing standards frameworks and coordination offices rather than duties owed to workers. Two legal systems, one constrained and one not, arrived at the same shelf.

Competence is a real limit on Europe. It is not an explanation for the pattern.

The Questions a Name Admits

A name does not only route the bill. It sets the range of askable questions, because it tells everyone what kind of thing is under discussion. Call something an intelligence and you have asserted a subject. Subjects have capacities, and possibly interests, and arguably standing.

Follow that into copyright. Stephen Thaler applied to register a work generated by a system he called the Creativity Machine, listing the machine as author. The Copyright Office refused. He litigated it to the D.C. Circuit, which on March 18, 2025 held that human authorship is required as a matter of statutory law; the Supreme Court denied review on March 2, 2026.

One litigant is thin evidence, and Thaler is a particular litigant — the same man who filed DABUS patent applications across a half-dozen jurisdictions expressly to manufacture test cases. Taken alone, the case shows only that a determined plaintiff can compel courts to say something obvious at length.

It does not stand alone. On February 16, 2017, the European Parliament adopted a resolution on civil law rules for robotics that asked the Commission to explore, among other possible legal solutions, "creating a specific legal status for robots in the long run, so that at least the most sophisticated autonomous robots could be established as having the status of electronic persons responsible for making good any damage they may cause." That is not a crank filing. It is a chamber of the European legislature putting the legal personality of a device on the table for study. In April 2018, 156 researchers, roboticists, lawyers, and ethicists signed an open letter to the Commission arguing the proposal was incoherent — that no machine on any near horizon has the attributes personhood presupposes. The Commission did not pursue it.

That is the frame doing work: a legislature spending institutional attention on the moral status of an instrument, and a hundred and fifty-six specialists spending theirs to say no. Under the name the field nearly had, there is nothing there to propose. There is no paragraph about the legal personality of complex information processing. The question would have been the one copyright and liability law have handled without difficulty since the camera and the Jacquard loom: who is responsible for the output of an instrument a human operated. The intelligence frame generated a decade of moral-status argument and arrived, exhausted, at the loom answer.

The same frame drives the risk conversation. Because the systems are named as minds, the salient danger is what a mind might want. Executive Order 14110, signed October 30, 2023, built its reporting regime around training-compute thresholds — a proxy for capability trajectory, for how close the thing is getting to being the kind of thing the name promised. It was revoked in January 2025 by an administration with a different view of the same frame. Fifteen months, full swing, both directions, on a question about the systems' inner tendency.

Meanwhile the deployed harms — the scheduling system that fragments a shift, the scoring model that denies the claim, the monitoring stack that sets the pace of the warehouse — sat in an annex as a compliance line item. Not because anyone dismissed them. Because under an intelligence frame they are the boring case, and boring cases do not get the committee's floor time.

Where the Frame Broke

The argument would be too neat if the name were destiny, and it is not. There is a clean counter-case, and it is worth taking seriously because it shows exactly what escaping the frame costs.

In 2023 the Writers Guild of America and SAG-AFTRA struck, and both contracts came back with enforceable provisions on generative systems: limits on how studios may use machine-written material to affect writing credit and compensation, and consent-and-payment requirements for digital replicas of performers. They got binding terms in a single bargaining round — faster than any legislature has moved on anything comparable.

They got there because they never entered the debate the name invites. Nobody at that table argued about whether the model understands narrative. The question on the table was who gets paid when output is produced and who consents to the use of a likeness — a question about the distribution of gains from a machine, which unions have been answering since the power loom. They declined the intelligence frame and used the automation frame, and the automation frame had ready-made instruments.

But note the precondition. It took an organized workforce with an existing contract cycle, a credible strike threat, and a chokehold on production. That is the ability to route around the name — and it is rare. Most workers meeting these systems have no bargaining table to bring the other frame to. They get the frame the statute gives them, and the statute's frame was set in 1955.

What Neither Frame Asks

This essay has been arguing that the automation frame is the better of the two available registers, and it is. Before recommending it, though, it is worth naming what the two frames have in common, because the shared assumption is larger than the disagreement between them.

Both treat the system as a product with an owner.

Product-safety law asks whether the thing is safe to place on the market. Labor law asks who is paid for its use and who consents to it. Both questions presuppose a market, a placement, and someone who owns the thing well enough to place it there. Neither asks where the thing came from.

Look again at what the guilds won, because the boundary is visible in the settlement itself. SAG-AFTRA secured consent and payment for digital replicas — a performer's face, voice, and body, which the law already recognized as hers. The WGA secured limits on how machine-generated material may affect credit and compensation — writers' output, in a contract about writers' output. Both are the automation frame working exactly as advertised, on things at the table that were already property with a named owner.

Nobody at that table had standing to ask about the training corpus, and that is not a criticism of the negotiators. A bargaining table seats the parties to a contract. The corpus these systems are assembled from is the pooled written and recorded expression of an enormous number of people, nearly all of whom have no contract with anyone in the room, no union, no identifiable claim on any particular output, and no way to tell whether they are in it. The automation frame has no seat for them, because it is a frame about workers and employers, and being in the training data is not a job.

So the deepest question available here — whether a synthesis of everyone's expression can be privately enclosed at all — is not merely unanswered. It is unaskable in both registers, which is a worse condition than being answered badly. The intelligence frame cannot ask it because it is occupied with what the machine might want. The automation frame cannot ask it because it recognizes only parties to a contract. And there is no third frame with a committee.

What Is Actually Being Commemorated

A machine assembled from the pooled expression of a species — the letters, arguments, manuals, poems, and quarrels of everyone who ever wrote anything down — is a genuinely novel object, and a genuinely novel object deserves a name chosen for it. What this one got was a term selected to distinguish one lab's funding proposal from a rival's, at a moment when the technology could do essentially nothing.

Names harden. Once a word has a committee, a statutory definition, a trade association, an academic department, and a line in an appropriations bill, it stops being a description and becomes infrastructure. You cannot rename it without re-litigating every one of those attachments, which means the 1955 turf decision is now load-bearing in a way its author never intended and would probably have found funny.

The practical consequence is not that we are governing these systems badly. Both the EU and the US have produced serious, technically literate text. It is that we are governing them in the register the name selected — as products that might one day become minds — rather than in the register their actual deployment suggests, which is as instruments that redistribute work, attention, and money at scale, right now, in ways the law has old and effective tools for. And beneath both registers sits the question neither has a procedure for asking.

The seventieth anniversary being marked this month is not the birthday of a technology. The technology arrived much later and looked nothing like the plan. What has its seventieth birthday is a jurisdictional decision: four men, one summer grant, and a phrase chosen so that nobody would have to argue with Norbert Wiener.

The argument happened anyway. It is just happening in the wrong committee.