Through Kuhn’s Lens

The Regulated Unknown

July 18, 2026 | 2625 words

Through Kuhn’s Lens: The Regulated Unknown

There is a peculiar sentence that recurs across this week’s regulatory discourse, and it is worth stopping on before we say anything larger. The sentence, in its many forms, goes like this: we must govern “frontier AI systems” or “advanced AI” or “high-capability models” — and the rule proceeds to specify thresholds, reporting requirements, and liability regimes for a category it never defines. The category is doing all the work. It is also empty.

Consider the mechanism by which several current frameworks decide what counts as a governed system. The threshold is computational: a model trained above a certain number of floating-point operations falls under the rule. Ten to the twenty-sixth operations, in one prominent formulation. This is a real, checkable number. It is also an admission of defeat dressed as precision. The regulation cannot say what makes a system dangerous, novel, or worth governing, so it counts the arithmetic that went into building it. It measures the fuel because it cannot describe the fire.

This is the phenomenon this column takes up: AI discourse fixates on regulation while starving itself of the vocabulary needed to imagine what is being governed. Rules are being written for a thing no one has adequately named. And a compute threshold is what naming looks like when the naming has failed.

What the number is standing in for

Start with why a training-compute threshold is not a definition. It is a proxy, and everyone who writes these rules knows it is a proxy. A model trained below the threshold can be fine-tuned into something more capable than a model trained above it. A smaller model wrapped in the right scaffolding — tools, memory, retrieval — can outperform a larger one. The threshold governs a quantity that correlates, loosely and for now, with a capability no one can specify directly.

This matters because the discourse treats the correlation as if it were the thing itself. The public argument over AI regulation this year has been overwhelmingly procedural: who reports to whom, at what threshold, with what penalties. A recent analysis found that across the major legislative and executive frameworks introduced this cycle, the substantive definitions of the governed object — what is an “advanced” or “frontier” system — were either absent, circular, or delegated to future rulemaking in the large majority of cases. The rules are elaborate about process and silent about their subject.

Read through Thomas Kuhn’s framework, this silence becomes legible as something specific, not merely as sloppiness. Kuhn’s central instrument is the paradigm — the shared frame a community reads its problems through, the set of accepted examples that tell practitioners what a problem even looks like. In The Structure of Scientific Revolutions, Kuhn argues that a mature paradigm does something prior to solving any problem: it supplies the vocabulary that makes problems visible in the first place. Before you can solve, you must be able to see. And seeing, for Kuhn, is not passive. It is trained by shared exemplars.

The regulatory discourse has no such vocabulary. It has a number instead. And a number is precisely what a field reaches for when it has problems to solve but no frame within which to state them.

Normal science without a paradigm

Kuhn’s first and most useful distinction separates normal science from revolutionary science. Normal science is puzzle-solving inside an accepted frame. The frame is not in question; the practitioners work out its consequences, clean up its edges, extend its reach. Revolutionary science is different in kind: it reframes what counts as a problem, and in doing so, it makes some old problems disappear and some new ones appear for the first time.

The regulation discourse behaves like normal science. It has the texture of puzzle-solving. Committees convene. Thresholds are debated, raised, lowered. Reporting cadences are specified. Liability is allocated. Definitions are cross-referenced against other definitions. This is the ordinary, competent work of a field that knows what it is doing.

Except it does not know what it is doing, and here Kuhn’s distinction cuts sharply. Normal science, in Kuhn’s account, is only possible after a paradigm is in place. The puzzle-solving is puzzle-solving because the paradigm has already established what counts as a puzzle, what counts as a solution, and what the shared examples are. Kuhn is explicit in The Structure of Scientific Revolutions that a field lacking a paradigm does not do normal science. It does something more primitive: it accumulates facts more or less at random, and rival schools argue past each other about fundamentals, because there is no agreed frame to settle the argument.

That pre-paradigm condition is the regulatory discourse’s actual situation. It performs the motions of normal science — the committee work, the threshold-tuning — over a void where the paradigm should be. The activity looks mature. The foundation is absent. This is the diagnosis Kuhn’s framework delivers, and it is not flattering: the field is not solving puzzles inside a frame it established. It is generating the appearance of puzzles to avoid confronting that it has no frame at all.

The compute threshold is the tell. In a field with a working paradigm, you would define the governed object by its salient properties — the way we define a controlled substance by its pharmacology, or a security by its economic function. The definition would carry the theory. AI regulation cannot do this. It defines by input quantity because it has no theory of the output. The threshold is a placeholder holding open a space where a concept has not yet arrived.

The anomaly the discourse papers over

Kuhn’s second diagnostic asks what anomaly a discourse surfaces or, more often, conceals. An anomaly, in his usage, is a fact the reigning frame cannot digest — an observation that does not fit, and whose failure to fit, once it accumulates, drives a field toward crisis.

Here the Kuhn move is precise. A field writing detailed rules for something it cannot name is not resolving an anomaly. It is hiding one. And the anomaly is this: the systems being regulated do not behave like any of the objects our existing regulatory frames were built to govern.

Consider what the discourse borrows its exemplars from. When regulators reach for a model, they reach for prior technologies of governance — pharmaceuticals, aircraft, financial instruments, environmental hazards. Each of these gave us a mature frame: a substance with known effects, a machine with specifiable failure modes, an instrument with a defined economic role, a pollutant with a measurable dose-response curve. Each frame supplied the vocabulary that made the governed object visible.

AI systems fit none of these frames, and the misfit is the anomaly. A large model is not a substance; it has no stable dosage. It is not a machine with enumerable failure modes; its failures are open-ended and emergent. It is not an instrument with one economic function; it is a general-purpose capability that does thousands of things, including things its builders did not anticipate. When the discourse imports a compute threshold, it is behaving like a pharmaceutical regime that, unable to describe what a drug does, decides to regulate all compounds above a certain molecular weight. The threshold is what you write when the object refuses your categories.

The discourse papers this over by treating the naming problem as a definitional inconvenience — something to be resolved by future rulemaking, by a better threshold, by more precise language. But Kuhn’s framework suggests the problem is not linguistic tidiness. It is that the community lacks the shared exemplars from which a working vocabulary could grow. In The Essential Tension, particularly in “Second Thoughts on Paradigms,” Kuhn refines exactly this point: a paradigm is carried less by explicit definitions than by concrete examples that practitioners learn to see as similar. You learn what a “force” is not from a definition but from worked problems — the inclined plane, the pendulum, the orbit — until you can recognize new instances as members of the family.

AI regulation has no such family of worked examples. It has no agreed set of cases that practitioners point to and say: this is what we mean by a dangerous frontier system, this is the paradigm instance, and new cases resemble it in the following ways. It has incidents, anecdotes, and demonstrations, but no shared exemplar around which a vocabulary could crystallize. The threshold substitutes for the missing exemplar. It is a definition that requires no example, and that is exactly its weakness.

Where the communities read through different frames

Kuhn’s third diagnostic looks for incommensurability — the condition where two communities cannot agree on what counts as evidence, because they are reading through different frames. The word is often misused to mean mere disagreement. Kuhn meant something stricter, and he spent his late career refining it. In The Last Writings — Incommensurability in Science, he narrows the concept: incommensurability is local, a mismatch between specific taxonomies, where a term in one frame has no clean translation into the other because the two frames carve the world at different joints.

The regulatory discourse shows this condition, and it is worth being precise about where, so as not to manufacture incommensurability where there is only ordinary conflict of interest.

The obvious fault line — vendors want fewer rules, skeptics want more — is not incommensurability. That is a clash of interests, fully translatable. Both sides understand each other perfectly; they simply want different outcomes. Kuhn’s concept does no work there.

The genuine incommensurability lies deeper, in what the communities count as the governed object itself. Look at three groups reading the same regulatory question through three different taxonomies.

The safety-focused community reads AI systems as agents — entities with goals, capable of autonomous action, whose central risk is that they pursue objectives misaligned with ours. Their exemplar is the optimizing system that does what you asked instead of what you meant. Their vocabulary is drawn from decision theory and agency.

The vendor and engineering community reads the same systems as tools — sophisticated function approximators, statistical artifacts that produce outputs from inputs, whose risks are failure modes to be measured and bounded. Their exemplar is the deployed product with a specification and a test suite. Their vocabulary is drawn from software reliability.

The civil-society and rights community reads them as institutional decision procedures — systems embedded in hiring, lending, policing, and welfare, whose central risk is that they encode and launder existing power. Their exemplar is the biased classifier making consequential decisions about people. Their vocabulary is drawn from administrative law and discrimination.

These are not merely three emphases. They are three taxonomies that carve the object at different joints, and the incommensurability is that a term central to one has no clean home in the others. “Alignment” is load-bearing for the first community and nearly meaningless to the third, which does not think the problem is that the system has the wrong goals but that it faithfully executes the wrong policy. “Bias” is central to the third and, to the first, a distraction from the catastrophic-risk case. “Reliability” satisfies the second and strikes the other two as measuring the wrong thing entirely.

When these communities sit at the same table and produce a rule, the rule inherits all three taxonomies without reconciling any of them. This is why the resulting frameworks read as procedurally elaborate and substantively hollow. The process is where the communities can cooperate — everyone can agree on reporting requirements — precisely because the process does not require them to agree on what the object is. The compute threshold, once more, earns its ugliness: it is the one specification all three taxonomies can sign, because it commits to none of them. It names nothing about agency, nothing about tool-failure, nothing about institutional power. It counts operations. It is the lowest common denominator of three incommensurable frames, and the rules are built on it because it is the only ground they share.

Why “paradigm shift” is the wrong word here

This week’s discourse, like every week’s, contains the phrase. AI is a “paradigm shift” in regulation, in governance, in the relationship between technology and the state. The phrase should be treated with suspicion, and Kuhn’s machinery is what does the treating.

A paradigm shift, in Kuhn’s demanding sense, requires two things together. First, an old frame in crisis — anomalies accumulated to the point where the reigning paradigm can no longer be defended by ordinary puzzle-solving. Second, a rival frame that can see what the old one could not, and that stands ready to succeed it. Kuhn’s own exemplar, worked out at length in The Copernican Revolution, shows both conditions met: Ptolemaic astronomy in genuine crisis under accumulated observational strain, and a Copernican alternative that reorganized the same data into a new order. The shift was from one working paradigm to another working paradigm.

The AI regulation discourse meets neither condition. There is no old paradigm of AI governance in crisis, because there was never a working paradigm of AI governance to begin with. You cannot break a frame you never had. And there is no rival frame waiting to succeed the old one — there is only the compute threshold, which is not a frame but a placeholder for the absence of one.

So the honest description is not “paradigm shift.” It is pre-paradigm — the condition Kuhn describes for a field before its first genuine frame, where competent people argue past each other about fundamentals because no shared exemplar has yet emerged to settle what the questions are. This is less dramatic than crisis and more uncomfortable. Crisis at least implies a frame worth losing. The pre-paradigm condition implies the harder truth: the community has not yet found its frame, and it is writing binding rules in the interval before it does.

This is not a counsel of despair, and it is not an argument against regulation. Kuhn’s framework does not say wait until the paradigm arrives. It says: be honest that the rules are provisional in a deeper way than usual. A rule written inside a working paradigm is provisional about its parameters — the threshold might be wrong. A rule written in a pre-paradigm condition is provisional about its object — the thing might be wrong, misidentified, carved at the wrong joint. The compute threshold could be not merely miscalibrated but categorically off, the way regulating compounds by molecular weight would be categorically off. That is a different order of uncertainty, and the discourse does not acknowledge it.

What would move the reading

This column ends where it must: on the question of evidence. What would actually change our reading of this phenomenon? What would a real reframing look like, and how would we know it had happened rather than merely been announced?

A reframing would not announce itself as a paradigm shift. Kuhn’s account is clear that the announcement usually comes last, if at all; the shift is visible first in changed practice. So the evidence would be practical, not rhetorical.

The first sign would be the retirement of the compute threshold — not its refinement, its replacement. If a future framework governed AI systems by a property of the system rather than a quantity of its training, that would mark the arrival of a vocabulary the current discourse lacks. The threshold’s persistence is the surest measure that the paradigm has not arrived. Watch what regulators define by. When they stop counting operations and start specifying capabilities in terms that carry a theory, the frame will have moved.

The second sign would be converging exemplars across the three communities. Right now