Through Kuhn’s Lens

AI’s Vanishing Vocabulary

July 25, 2026 | 2620 words

Through Kuhn’s Lens: The Frame That Reads as No Frame

Read enough writing about artificial intelligence this year and a strange sensation sets in: the words have stopped arguing. A decade ago, to write about AI was to pick a fight. You called it a partner or you called it a threat. You said it would transform work or you said it would hollow it out. Now the register has gone flat. The dominant vocabulary is procedural — compliance, alignment, governance, risk tiers, model cards, audits. Regulation-talk has crowded out ethics-talk. The technology is described everywhere and interpreted nowhere.

This is worth naming precisely, because the flattening does not present itself as a position. It presents itself as maturity. The old frames — AI as partner, AI as threat, AI as transformation — now read as naive, overheated, the stuff of an earlier and less serious moment. What replaced them wears the costume of neutrality. And a vocabulary that claims to be no vocabulary at all is exactly the kind of object that Thomas Kuhn’s history and philosophy of science was built to inspect.

The Neutral Register Is Not the Absence of a Paradigm

The first diagnostic question in the Kuhn battery is the sharpest one here. Is this flattening a case of what Kuhn called normal science — routine puzzle-solving inside an accepted frame — or is it a reframing of what counts as a problem at all?

A gloss is owed before the term does any work. In The Structure of Scientific Revolutions, Kuhn described normal science as the day-to-day activity of a mature field: researchers who share a set of assumptions so thoroughly that they no longer argue about them. They spend their effort solving the puzzles the shared frame throws up, not questioning the frame itself. Normal science is not a failure of imagination. It is what makes cumulative, detailed work possible. But it has a cost. The shared frame determines what counts as a legitimate problem, and it renders other questions literally unaskable — not forbidden, just invisible.

The instinct is to say the neutral register is the loss of a paradigm — that the field once had rich interpretive frames and has now surrendered them. That instinct is wrong, and correcting it is the whole essay.

The neutral register is not the absence of a paradigm. It is the consolidation of one. A frame becomes invisible precisely when it stops competing. When AI-as-partner and AI-as-threat were fighting for the culture’s attention, everyone could see that a frame was in play, because there was more than one. The moment a single register wins, it stops reading as a register. It reads as the world. This is Kuhn’s central and least comfortable insight: the most powerful paradigm is the one no one experiences as a choice.

The data bears the shape of this. In the corpus of AI discourse tracked this year, the language of governance and regulation now dominates the vocabulary that a decade of commentary spent building. The shift is not that people stopped talking about AI. They talk about it constantly. The shift is that the talk converged. When roughly the same procedural vocabulary — risk, compliance, alignment, oversight — saturates industry statements, policy documents, and journalism alike, that convergence is not neutrality. Convergence is what victory looks like from the inside.

So: normal or revolutionary? Neither, and that is the point. The flattening is not a revolution, because a revolution would install a new way of seeing that visibly displaces the old. And it is not quite ordinary normal science either, because it is happening in the public discourse about a technology, not inside a research community with shared exemplars. What it most resembles is the moment before normal science calcifies — the point at which competing frames collapse into a single one that then gets treated as simply the responsible way to speak.

What the Regulatory Frame Cannot Name

Kuhn’s second diagnostic is about anomaly. An anomaly, in The Structure of Scientific Revolutions, is an observation the reigning paradigm cannot account for — a fact that does not fit, and that the frame’s own tools cannot process. Anomalies matter because they are the pressure that eventually cracks a paradigm open. But here is the subtler claim, and the one that fits this week’s phenomenon: a paradigm shapes not only what answers you can give but what problems you can see. A vocabulary that has gone flat may be a vocabulary that has lost the ability to name its own anomalies.

Ask the specific question. What can no longer be said, once “partner,” “threat,” and “transformation” fall out of use?

Consider what each of those dead frames was actually for. “Partner” was a claim about relationship — that using AI changes the person using it, that there is a question of dependency, of what a human retains and what atrophies. “Threat” was a claim about harm that need not be quantifiable to be real — displacement, deskilling, the slow erosion of judgment. “Transformation” was a claim about scale and irreversibility — that some changes cannot be undone and should therefore be chosen deliberately, not drifted into.

Now watch what the regulatory frame does with each. It does not deny them. It translates them into its own currency, and something is lost in every translation. “Partner” becomes a question of human-in-the-loop design. “Threat” becomes a risk category to be mitigated, scored, and tiered. “Transformation” becomes an adoption timeline with compliance milestones. Each translation is procedurally faithful and interpretively empty. The regulatory frame can tell you whether a system meets a standard. It cannot tell you whether the standard is worth meeting.

This is the anomaly the flat vocabulary papers over: the questions that have no procedural form. Regulation-talk excels at bounded harms — measurable bias, traceable failure, assignable liability. It goes silent on unbounded ones. What happens to a professional’s judgment after five years of deferring to a model? Is there a form of human capability whose loss no audit will ever flag, because no regulation is written against becoming worse at thinking? The ethics-talk that the neutral register displaced was clumsy and overheated. But it could at least pose those questions. The regulatory register cannot pose them. It can only wait for them to arrive as compliance failures — and by then the anomaly is no longer an anomaly. It is the new normal, absorbed and unremarked.

Kuhn’s account of the run-up to revolution in The Structure of Scientific Revolutions turns on scientists’ growing awareness that something does not fit. But that awareness depends on having a vocabulary rich enough to register the misfit as a misfit. A community that has flattened its language into pure procedure has disabled its own anomaly detectors. The danger of the neutral register is not that it gives wrong answers. It is that it quietly narrows the set of questions until the ones that matter most fall outside the frame — and their absence is not noticed, because absence never announces itself.

Whose interest does that narrowing serve? The column’s posture requires the question, and the answer is not mysterious. A vocabulary that can only speak of bounded, mitigable, compliable risk is a vocabulary extraordinarily convenient for anyone building and selling the systems. Ethics-talk asks whether a thing should exist. Regulation-talk assumes it exists and asks only how to run it safely. The move from the first question to the second is not a neutral maturation. It is a settlement, and the settlement favors the party that has already built the thing.

Where the Camps Stopped Disagreeing

The third diagnostic is incommensurability, and it is the one Kuhn spent the end of his life refining. The plain-English version: two camps can use the very same words and mean different things by them, because they read through different frames and measure by different examples. Kuhn’s term for those shared examples was exemplars — the concrete model problems and solutions a community learns from, the cases that teach you what a good solution even looks like. When two communities learn from different exemplars, they can talk for hours and never touch. In The Last Writings — Incommensurability in Science, Kuhn sharpened this from a claim about whole worldviews into something more precise and more local: incommensurability is often about specific terms that two communities have quietly taught themselves to use differently.

The intuition about AI discourse runs the wrong way. You would expect the industry camp and the skeptic camp to be maximally incommensurable — talking past each other completely, each in its own private dialect. But look at the case carefully. The striking thing about the current moment is not that industry and skeptics can no longer understand each other. It is that they increasingly can. They have converged on the same flattened vocabulary. Both now speak governance. Both cite alignment. Both invoke risk. The skeptic who once said “threat” now says “insufficient safeguards.” The booster who once said “partner” now says “responsible deployment.” They have achieved a common tongue.

And this is the trap. They stopped being incommensurable, and in doing so they stopped disagreeing productively. The shared vocabulary is not a bridge. It is a floor they both agreed to stand on, and the floor belongs to the regulatory frame. Once both camps accept that the real question is how AI should be governed, the prior question — whether this deployment should happen at all, what it does to the humans inside it, what kind of dependency it builds — has been conceded before anyone spoke. The disagreement that remains is a disagreement over settings, not stakes.

Kuhn’s point in The Last Writings — Incommensurability in Science was that incommensurability, painful as it is, is also generative. When two communities genuinely cannot reduce each other’s terms to their own, the friction forces each to articulate what it actually believes. Productive disagreement requires a shared vocabulary rich enough to locate the disagreement — but not so flattened that the disagreement disappears into it. The AI discourse has lost that balance. It has too much shared vocabulary and too little of the kind that lets a real fight happen. The camps are no longer incommensurable in Kuhn’s demanding sense. They are something worse: commensurable on terms that render their deepest disagreement unspeakable.

Notice what this does to the reader, the party the column serves. When both the enthusiast and the skeptic speak governance, the reader loses the single most useful thing a contested discourse provides — the ability to triangulate. If two people who are supposed to disagree keep reaching for the same words, the reader has no way to test the words. The vocabulary that could once be checked against a rival vocabulary now floats free, unchallenged, because the challenge has been absorbed into the same register it was meant to challenge.

Policing the Big Word

The instruction of this column is to police the phrase “paradigm shift,” and the flattening of AI vocabulary is a perfect occasion to enforce it — because the flattening is not a paradigm shift, and understanding why is understanding the whole phenomenon.

In colloquial usage, “paradigm shift” means any large change. In Kuhn’s actual machinery, it means something far more specific and far more demanding. A paradigm shift, as The Structure of Scientific Revolutions lays it out, is a revolution: a period in which the reigning frame accumulates anomalies it cannot resolve, enters crisis, and is displaced by a rival frame that reorganizes the field’s basic assumptions. Crucially, the shift is visible as a struggle. There is a before and an after. There is a community that resists and a community that converts. The world looks different on the far side. Kuhn’s exemplar for all of this, worked out in detail in The Copernican Revolution, was the passage from an Earth-centered to a Sun-centered cosmos — a change so total that the same observed points of light in the sky came to mean something entirely different.

Measure the vocabulary flattening against that standard. Is there a crisis? No. Crisis is loud; this is quiet. Is there a rival frame visibly displacing an old one in open struggle? No — the striking feature is the absence of struggle, the smoothness with which ethics-talk gave way to regulation-talk. Nobody fought a revolution to install the neutral register. It arrived like weather.

That is precisely why it is not a paradigm shift and precisely why it is dangerous. Kuhn’s revolutions are conspicuous. They announce themselves through conflict, through the refusal of the old guard, through the incommensurability of the two camps. What is happening to AI vocabulary announces nothing. It is consolidation misread as maturation. The frame did not break and reform. One frame among several simply won by attrition and then draped itself in the language of neutrality so that its victory would not register as a victory.

Kuhn distinguished, especially in the essays collected in The Essential Tension, between the productive tension that keeps a mature field both stable and capable of change, and the sterile stability of a field that has stopped being able to see its own commitments. The neutral register is the second kind. It is stable because nothing challenges it, and it cannot see its commitments because it has convinced itself it has none. To call this a paradigm shift would be to grant it the drama and the legitimacy of a genuine revolution. It has earned neither. It is a settling, not a shift.

What Would Actually Move the Reading

The column always ends the same way, and the discipline is worth honoring: not with a conclusion to absorb but with the concrete evidence that would change the analysis. If someone insists the flat vocabulary is not a consolidated frame but a genuine maturation — the field simply growing up and setting aside its adolescent metaphors — what would show them right, and what would show them wrong?

Three tests, each concrete.

First, watch for the return of asymmetry in the vocabulary. Right now the convergence is the whole story: industry and skeptics reaching for the same procedural words. A genuine sign of life would be a new frame emerging that neither camp can reduce to the other’s — a term that industry and skeptics use differently and cannot translate. That would be Kuhn’s incommensurability reappearing, and its reappearance would be evidence that the discourse had regained the ability to host a real disagreement. If, instead, the vocabulary grows more uniform over the next year — if even the dissenters phrase their dissent in governance-speak — the consolidation reading holds and hardens.

Second, watch what happens to anomalies. The consolidation reading predicts that unbounded harms — the erosion of judgment, the atrophy of skill, the questions with no procedural form — will keep failing to register until they arrive as compliance events. So track a specific case: an instance where something clearly went wrong in a way no regulation named in advance. Watch whether the discourse can only describe it retrospectively, as a gap in the framework to be patched, or whether it develops language to name the harm as a harm, on its own terms, independent of whether a rule was broken. The first is the flat frame absorbing its anomaly. The second would be the frame cracking. The difference is observable, and it is the difference that matters.

Third — and this is the strictest test — watch for a genuine crisis that the regulatory vocabulary cannot metabolize. Kuhn’s revolutions begin when the reigning frame’s own practitioners lose confidence that it can solve the problems it faces. The