Through Kuhn’s Lens

Governing What We Can’t Name

August 16, 2026 | 2552 words

Through Kuhn’s Lens: Governing What We Can’t Name

This week the discourse around artificial intelligence produced a strange asymmetry. Regulation dominated the conversation—hearings, frameworks, compliance regimes, liability debates. Meanwhile the words a society would need to say what it is regulating went almost silent. “Partner” barely surfaced. “Transformation” appeared as a slogan, not an analysis. Even “machine learning”—the technical name for the thing itself—hardly registered against the volume of governance talk. A society was busy drafting rules for an object it had declined to describe.

Consider the oddity directly. To govern something, you normally need a working account of what it is. Traffic law rests on a shared idea of a vehicle. Food safety rests on a shared idea of contamination. But this week’s AI governance discourse ran at full speed while the framing vocabulary—the words that would fix AI as a kind of thing—stayed absent. The regulation was loud. The naming was quiet. And nobody in the discourse seemed to notice the gap.

That gap is the phenomenon. Not the regulation itself, and not any particular bill. The phenomenon is the ratio: enormous governance energy directed at a conceptual blank. This essay uses the history-and-philosophy of science of Thomas Kuhn as an instrument to examine that blank—to ask what it lets a society avoid seeing, and who benefits from the avoidance.

The Instrument

Kuhn’s central claim is that what a community can see depends on the frame it reads through. His word for that frame was paradigm—the accepted set of assumptions, worked examples, and legitimate problems that a scientific community treats as settled and no longer argues about. A paradigm is not a theory you can state on a page. It is the background that makes some questions obvious and others invisible. In The Structure of Scientific Revolutions Kuhn argued that facts do not accumulate into discovery on their own. Discovery comes when the frame breaks.

A second concept will do the heavy lifting here: incommensurability. The term means that two communities can use the same words while meaning different things, so that they argue past each other without realizing it. They are not disagreeing about answers. They are running different questions through different frames, and the shared vocabulary hides the mismatch. Kuhn spent the last decades of his career sharpening this idea; the essays collected in The Last Writings — Incommensurability in Science treat it as the core mechanism of how communities fail to communicate across a conceptual divide.

These two tools—paradigm and incommensurability—are enough to work the case. The question for a phenomenon in the social aspects of AI is narrow and specific: what frame does a society read AI through, and are anomalies accumulating against that frame? This week the frame is not a description of AI at all. It is an absence dressed as neutrality.

Normal or Revolutionary?

Kuhn drew a sharp line between two kinds of scientific work. Normal science is puzzle-solving inside an accepted paradigm. The frame is settled; the community’s job is to extend it, tidy its edges, resolve its known problems. Revolutionary science is the rare event when the frame itself fails and must be replaced. The distinction matters because the two look nothing alike from inside. Normal science feels like progress. Revolution feels like crisis.

Run this week’s regulatory discourse through that distinction and something clarifies immediately. The regulation behaves exactly like normal science. It is puzzle-solving. It asks: who is liable when an AI system causes harm? What disclosure should be required? How do we audit a model? These are governance puzzles, and they are being worked with the confidence of a community that believes its frame is settled. The problems are bounded. The methods are familiar. The debates are about parameters, not foundations.

But here is the tension the case forces. Normal science solves puzzles inside a paradigm that has already named its object. The chemists who worked inside Lavoisier’s oxygen theory knew what oxygen was supposed to be. The governance discourse this week is running normal-science machinery over an object that was never paradigmatically fixed. It is puzzle-solving over an unnamed thing.

This is not a small awkwardness. It is the central feature of the phenomenon. The regulatory community has adopted the behavior of a settled field—the confident puzzle-solving, the bounded debates—without the settlement that behavior is supposed to rest on. The framing vocabulary that would have done the settling is the vocabulary that went missing. “Partner,” “transformation,” “machine learning”: these are not decorative words. Each carries a theory of what AI is. A partner is an agent with which one collaborates. A transformation is a process that changes the shape of work and life. Machine learning is a specific technical claim—that a system’s behavior is derived from data rather than written as rules. Drop all three, and you are left with an object that has no asserted nature at all.

Kuhn would not call this a paradigm shift, and neither should the reader. There is no incommensurable successor frame replacing an older one. There is no crisis forcing a reframing. There is, instead, normal-science governance proceeding as if a paradigm existed when none has been articulated. The discourse has skipped the naming and gone straight to the regulating. That is not revolution. It is something Kuhn’s model does not quite have a name for: settled behavior without a settled object.

The Anomaly the Neutral Blank Suppresses

The most important move in this week’s discourse is the flattening of AI into a neutral blank—a tool, an object, a technology to be governed like any other. This flattening is what the absent vocabulary produces. When “partner” and “transformation” and “machine learning” fall out of the conversation, what remains is a thing with no asserted character. And a thing with no asserted character is easy to call neutral.

Kuhn’s framework is precise about what neutrality does. In The Structure of Scientific Revolutions he argued that a paradigm’s most powerful effect is what it renders invisible. The frame does not just organize what a community sees. It determines what counts as a fact worth noticing at all. Anomalies—observations the frame cannot accommodate—do not announce themselves. They are suppressed, ignored, or explained away, until they accumulate past the point of denial.

So the question is: what does the neutral-blank framing let a society not see?

Look at the missing words more carefully. Each one, if spoken, would force an admission the neutral framing avoids.

Say “partner,” and you admit that AI is being positioned as an agent—something that acts, decides, collaborates. That admission carries governance consequences the neutral frame dodges. You do not audit a hammer. You do not assign responsibility to a spreadsheet. The word “partner” would drag agency into the room, and agency demands a whole apparatus of accountability the neutral blank quietly excuses.

Say “transformation,” and you admit that AI is not an object to be regulated in place but a process reshaping the very institutions doing the regulating. Transformation is not something you govern from a stable outside. It moves the ground under the governor. The neutral frame keeps the regulator standing on solid earth, looking down at an inert thing. “Transformation” would tell the regulator the earth is moving too.

Say “machine learning,” and you admit the technical specificity of the thing—that its behavior is derived from training data, that it is opaque by construction, that its failures are statistical rather than logical. That admission would make most of the week’s governance puzzles far harder. You cannot demand a clean explanation from a system whose defining feature is that it produces outputs no one wrote rules for. The neutral blank lets regulators pretend they are governing a legible machine.

Here the essay’s editorial obligation comes due. Who benefits from the flattening? The answer is uncomfortable and worth stating plainly. A neutral blank is the most convenient object a powerful vendor could hope to have regulated. It carries no asserted agency, so no accountability attaches by default. It carries no admission of transformation, so incumbents keep their footing. It carries no technical specificity, so the hardest questions stay off the table. The neutrality that looks like caution—like refusing to hype the technology—functions instead as a shield. The absent vocabulary is not a neutral absence. It is a suppression, and the suppression has beneficiaries.

That is the anomaly the neutral blank papers over: the mounting evidence that these systems act, transform, and fail in ways no ordinary tool does. The evidence exists. The framing keeps it from counting as a fact worth governing.

Incommensurability: The Same Words, Different Worlds

Kuhn’s late work insisted that incommensurability is not vagueness or mere disagreement. It is a structured failure of communication in which two communities use identical terms anchored to different exemplars—the shared, obvious examples each community treats as the paradigm case. In The Last Writings — Incommensurability in Science he located the problem in the way communities learn their terms through examples rather than definitions. Change the examples, and the same word points to a different world.

This week’s discourse contains a clean instance, and it is worth showing rather than asserting.

Take the word “safety.” Both the AI-industry community and the AI-skeptic community use it constantly, and both use it inside the regulatory conversation as if it named a shared goal. It does not.

For the industry community, the exemplar of safety is the aligned model—a system that refuses harmful requests, avoids catastrophic outputs, and behaves within specified bounds. Safety, on this reading, is a property engineered into the artifact. It is a technical achievement, measurable, improvable, and demonstrable to a regulator. When the industry says a system is “safe,” it means the artifact has been built to behave.

For the skeptic community, the exemplar of safety is the protected person—the worker not displaced without recourse, the citizen not surveilled, the applicant not silently scored by a model no one can inspect. Safety, on this reading, is a property of the social arrangement, not the artifact. It is not something you engineer into a model. It is something you guarantee to a population.

Now watch what happens in a regulatory hearing where both communities use the word. The industry witness testifies that the system is safe—meaning the model has been aligned and tested. The skeptic witness responds that the system is not safe—meaning people remain unprotected. Each hears the other as either lying or ignorant. Neither is. They are running the same word through incommensurable exemplars. The industry witness cannot see the skeptic’s meaning as a fact about safety, because in the industry’s frame, safety lives in the artifact and the artifact tests clean. The skeptic cannot accept the industry’s meaning, because in the skeptic’s frame, a well-behaved model that still enables mass surveillance is not safe in any sense that matters.

The tragedy Kuhn identified is that they do not know they are talking past each other. The shared word hides the mismatch. The hearing produces the appearance of a debate about how safe the system is, when it is actually two communities disagreeing about what safety is. And crucially, the neutral-blank framing makes this worse. Because AI has not been named—because no shared account of what it is has been fixed—there is no common object to anchor the word. When the object is a blank, “safety” floats free, and each community fills it with its own exemplar. The missing naming vocabulary is precisely what would have forced the two meanings into contact. Its absence lets them coexist undetected, each community certain it is having the real conversation.

This is why the volume of regulatory talk is misleading. High volume looks like engagement. Under incommensurability, it can be its opposite—two communities generating enormous discourse while their central terms slide past each other, unexamined, because no shared frame exists to catch the slide.

What a Real Shift Would Require

The discipline of Kuhn’s model is that it makes “paradigm shift” expensive to claim. A shift is not a big change or a fast one. It is the replacement of one frame by an incommensurable successor, forced by anomalies that accumulate into a crisis the old frame cannot absorb. In The Copernican Revolution Kuhn traced how the geocentric frame did not fall because someone produced a decisive fact. It fell after its accumulated anomalies—the wandering planets, the epicycles piled on epicycles—made the whole apparatus unbearable, and only then did a rival frame become thinkable.

Apply that standard here, and the diagnosis is plain. There is no paradigm shift underway in AI governance. There is not even a stable paradigm to shift from. What exists is a neutral-blank framing that has never been articulated as a positive account of AI—it is an absence, not a theory—and normal-science governance proceeding on top of that absence as if it were settled ground.

The unsettling possibility is that the absence is hardening. A frame does not have to be true to become settled. It only has to become the unquestioned background through which a community reads its object. If “AI is a neutral tool to be governed like any other technology” becomes the tacit paradigm—never stated, never defended, simply assumed—then the naming vocabulary will not return, because the frame will have made the naming feel unnecessary. Why would you argue about whether AI is a partner or a transformation when everyone already knows it is just a tool? The neutral blank could become the paradigm precisely by never being spoken. That is the most Kuhnian danger in the case: a frame so successful it becomes invisible, governing a society’s vision without ever appearing in its vocabulary.

Kuhn’s refinements in The Essential Tension are useful here. He distinguished the paradigm as a community’s shared examples from the paradigm as its broader worldview, and he noted how the former quietly install the latter. The exemplars do the work of the worldview without anyone stating the worldview. In the AI case, the governance exemplars—liability, audit, disclosure—are all drawn from the regulation of ordinary tools. Adopt those exemplars, and you have adopted the neutral-blank worldview without ever asserting it. The frame arrives through the examples, unnamed, exactly as Kuhn described.

What Would Move the Diagnosis

The column’s obligation is not to summarize but to specify. What observation would tell us whether this reading is correct—whether the naming vocabulary is returning, or whether the neutral blank has hardened into the settled frame?

Watch the regulatory language for the return of the specific words, and watch which community uses them. If “partner” and “agent” re-enter governance discourse from the regulators’ side—not the vendors’ marketing—that would signal that agency is being forced back onto the table, and with it the accountability the neutral blank excuses. If regulators begin drafting rules that name AI as a process of transformation rather than an object in place—rules that acknowledge the ground moving under the institutions doing the governing—that would signal the neutral frame is cracking.

Watch the word “safety” for a forced confrontation of its exemplars. The diagnosis of incommensurability would be confirmed if industry and skeptic continue using the term at high volume while producing no convergence on cases—each certain, neither moved. It would be disconfirmed, and a genuine reframing signaled, if some event dragged the two ex