The Category Error Behind Both Political Arguments and Machine Intelligence

What if the most dangerous mistake in public debate is not believing the wrong fact, but asking the wrong category to do too much work?

A single sentence about Jewish indigeneity can turn a layered history into a binary choice: either Jews are indigenous to the Middle East or they are European. A single sentence about artificial intelligence can make a similarly compressed claim: either language models merely predict text or they can generate genuine discoveries. These subjects appear unrelated, but they expose the same intellectual problem.

Complex systems resist one dimensional labels. They become intelligible only when we examine how several kinds of identity, evidence, and inference interact. The deeper connection is not that politics resembles machine learning. It is that both debates ask us to distinguish between a category and the reality that category imperfectly describes.

The seduction of the single frame

Consider the claim that if Jews demand an indigenous homeland, they must return to Europe because Jews are European. Its rhetorical force comes from treating “European” and “indigenous” as mutually exclusive descriptions. But they answer different questions.

A person can have European citizenship, speak a European language, possess ancestors who lived in Europe for centuries, and still belong to a people whose historical origins, religious imagination, and collective memory are rooted elsewhere. Jewish communities have been shaped by many regions, including Europe, North Africa, the Middle East, and Central Asia. Their histories are not identical, and no genetic, cultural, or political description captures every Jewish person. Yet precisely because the history is complicated, one label cannot settle the entire question.

The word “indigenous” is itself not a synonym for “genetically pure,” “first in every period,” or “morally entitled to unlimited political power.” In serious usage, it can refer to a people’s historical connection to a place, continuity of collective identity, political subordination, cultural institutions, and relationship to ancestral land. Those dimensions may overlap, but they are not interchangeable.

This distinction matters because public arguments often convert a classification question into a permission question. “Is this group indigenous?” quietly becomes “Does that give the group an unlimited right to rule?” Or it becomes “If the classification is contested, does the group have no legitimate claim at all?” Both leaps are invalid. A historical relationship to a place may be real without resolving every contemporary dispute about borders, sovereignty, equality, displacement, or state conduct.

A label can illuminate one dimension of reality while hiding another. The mistake begins when we treat partial illumination as total explanation.

The same error appears in discussions of artificial intelligence. Saying that a language model predicts likely continuations does not by itself prove that it cannot reason, discover, or produce novelty. Human thought also operates through prediction, association, memory, compression, analogy, and revision. The mechanism of generation does not settle the status of the result.

At the same time, saying that a model produced a surprising answer does not prove that it made a scientific discovery in the richest human sense. Novel output can be accidental, derivative, or impossible to reproduce. Here too, the crucial task is to separate categories that are often collapsed: originality, usefulness, explanation, verification, and understanding.

Why breakthrough ideas require more than accumulation

There is a familiar picture of intelligence in which progress comes from collecting enough information. Give a system more books, more examples, and more computational power, and eventually it will find the answer. This picture is partly right. Knowledge supplies the materials from which insight is built. But many breakthroughs do not emerge by simply extending an existing line of thought.

They arrive through a shift in the question.

A scientist does not merely calculate faster when moving from one conceptual framework to another. The scientist notices that the current framework leaves a residue of unexplained observations, then constructs a new model that makes those observations coherent. This is often called abduction, or inference to the best explanation. Deduction asks what follows from premises. Induction generalizes from repeated observations. Abduction asks: what hidden structure would make this surprising pattern unsurprising?

Imagine finding wet streets in the morning. Deduction might establish what follows if it rained. Induction might show that wet streets often correlate with rain. Abduction proposes rain as the explanation, while remaining open to sprinklers, flooding, or a street cleaner. The creative act lies in generating a candidate explanation that reorganizes the evidence.

Breakthrough science often depends on this reorganization. The important move is not always adding another fact. It is placing familiar facts into a new relationship.

This is why scaling alone may not be enough for discovery. More text can improve fluency, recall, and pattern recognition, but a larger archive does not automatically provide a new coordinate system. A model can know thousands of descriptions of a problem and still fail to represent the problem in the way required for a conceptual leap.

The proposed alternative, shifting toward joint reasoning, points toward a broader idea: intelligence may depend on coordinating multiple representations at once. Instead of asking one sequence predictor to do everything, we might combine language with formal logic, simulation, visual models, experiments, memory, external tools, and competing hypotheses. The point is not merely to add components. It is to make them constrain and revise one another.

A weather simulation can expose a contradiction in a verbal explanation. A mathematical proof can reveal that an intuitive story is impossible. An experiment can eliminate an elegant theory. A historical archive can show that a political category conceals discontinuities. Intelligence grows when representations enter productive friction.

The shared architecture of contested identity and creative discovery

The political and technological questions converge around a useful framework: identity is relational, and explanation is multilevel.

A person or people can be described biologically, culturally, historically, legally, geographically, and politically. A model can be described as a statistical system, a software artifact, a reasoning partner, a generator of hypotheses, or an instrument embedded in a human workflow. None of these descriptions is automatically the final one. Each becomes useful for a particular question.

The problem is not categorization. We need categories to think. The problem is category capture, when one description takes control of the entire discussion and refuses to surrender it even when the question changes.

Here is a practical test. Whenever someone makes a sweeping claim, ask four questions:

1. What dimension is being measured? Ancestry, residence, legal status, cultural continuity, predictive accuracy, novelty, or something else?
2. What is the unit of analysis? An individual, a population, a state, a text, a model, or a system involving humans and tools?
3. What time horizon matters? A lifetime, several centuries, the history of a people, a training run, or the duration of an experiment?
4. What conclusion is being smuggled in? Is a descriptive claim being used to justify a moral, legal, or political conclusion?

Apply this test to the assertion that Jews are European. It may describe part of the lived history of many Jews, especially those whose families lived in Europe for generations. It cannot, by itself, disprove older connections to the Levant, erase non European Jewish communities, or resolve the political rights of Palestinians and Israelis. A demographic description has been asked to carry the weight of a theory of sovereignty.

Apply the same test to the assertion that language models cannot make discoveries because they predict tokens. It may correctly describe an important mechanism. It does not by itself establish what the larger human machine can accomplish when a model generates hypotheses, a scientist tests them, and a formal system verifies them. A mechanistic description has been asked to carry the weight of a theory of intelligence.

In both cases, the hidden mistake is confusing where a process comes from with what the process can become within a larger system.

Intelligence is often distributed, not located

This suggests a more realistic model of creativity. Instead of treating discovery as a magical property inside an isolated mind, treat it as a loop:

Representation, surprise, hypothesis, challenge, revision, verification.

A system becomes capable of meaningful discovery when it can move through this loop reliably. The source of each step may vary. A human may supply the surprising question. A model may propose an analogy. A simulator may test consequences. A laboratory may produce the decisive observation. A mathematical checker may reject an attractive error.

Under this model, asking whether an AI “is creative” may be less useful than asking whether the human and machine arrangement supports creative inference. The relevant unit is not always the model. It is the joint cognitive system.

This has a parallel in political thinking. The relevant unit is not always an isolated individual carrying one identity label. It may be a people embedded in institutions, migrations, memories, laws, and relations with other peoples. A person can participate in several histories at once. A community can have a genuine ancestral connection to a land while also being responsible for present political actions that deserve criticism. Recognizing one layer does not require denying the others.

The joint system perspective also introduces accountability. Distributed intelligence is not distributed responsibility. If an AI assisted in producing a false scientific claim, humans still need to establish standards for evidence, attribution, and correction. If a state invokes historical identity to justify policy, historical connection does not exempt that policy from scrutiny. Complexity should prevent simplistic conclusions, not become an excuse to avoid judgment.

A useful principle follows:

The more layered the system, the more carefully we must separate explanation from justification.

History can explain why a group sees itself as connected to a place. It cannot by itself justify every action taken in that group’s name. A model’s architecture can explain how it generates an answer. It cannot by itself establish whether the answer is true or whether using it is wise.

A practice for making better leaps

If breakthrough thinking depends on moving between representations, individuals can cultivate it deliberately. The goal is not to produce random associations. It is to create disciplined opportunities for a better explanation to appear.

Start by writing the dominant frame in one sentence. For a political dispute, it might be: “This is only a question of who arrived first.” For an AI dispute, it might be: “This is only a question of next word prediction.” Then list what the frame explains and what it leaves unexplained.

Next, change the unit of analysis. Ask what becomes visible if the focus shifts from individuals to communities, from communities to institutions, from a model to a human tool system, or from a single output to a tested sequence of outputs.

Then force a second representation. Turn a verbal argument into a timeline. Turn a scientific intuition into an equation. Turn a moral claim into a list of affected parties. Turn a model’s answer into a hypothesis with a proposed test. The new representation should not merely decorate the first one. It should be capable of contradicting it.

Finally, preserve uncertainty where uncertainty is real. A mature conclusion may say: this historical connection is substantial, this legal claim remains contested, this output is novel but unverified, and this political action is judged independently. Such a conclusion sounds less decisive than a slogan because it is doing more intellectual work.

Key Takeaways

1. Separate description from justification. A fact about ancestry, residence, or historical memory does not automatically settle a question of political rights or moral responsibility.
2. Use multiple representations. Combine narrative, data, formal reasoning, simulation, and lived experience. Let each expose the blind spots of the others.
3. Treat abduction as a trainable habit. When evidence feels anomalous, ask what new structure would make it coherent instead of merely collecting more examples.
4. Judge systems, not isolated labels. Evaluate what humans, institutions, tools, and models accomplish together, while keeping responsibility visible.
5. Ask what a category cannot explain. The boundary of a concept is often more informative than its apparent certainty.

The deepest lesson is not that every perspective is equally valid, nor that truth dissolves into complexity. It is that truth often becomes visible only after we stop demanding that one frame answer every question.

A people can have multiple histories. A person can inhabit several identities. A machine can generate an idea without independently understanding its significance. A human can recognize the significance without having generated the first idea. In each case, what matters is the structure of relations among the parts.

The next great leap, whether in science or public reasoning, may therefore begin with a modest refusal: do not ask which single label wins. Ask which combination of perspectives explains more, predicts better, survives challenge, and leaves fewer important facts in the dark.