The Categorization Trap — Why Humanities Get Stuck Classifying and How AI Changes the Game

Here is a confession: TellDear maintains a taxonomy of 534 aspects of critical thinking, organized across 6 dimensions. We built it knowing — knowing — that it is imperfect, incomplete, and in some places arguably wrong. We shipped it anyway. This article explains why that decision was not reckless but necessary, why the humanities have spent centuries trapped in the pursuit of perfect categories, and why artificial intelligence may finally offer a way out of the trap — not by achieving perfection, but by making imperfection survivable.

I. The Ancient Dream of Perfect Classification

The urge to categorize is among the oldest intellectual impulses. Aristotle's Categories — written around 350 BCE — proposed ten fundamental types of predication: substance, quantity, quality, relation, place, time, position, state, action, and passion. It was, in effect, an ontology: a claim about what kinds of things exist and how they relate. It was also, almost immediately, controversial. His student Theophrastus questioned the boundaries. The Stoics proposed alternatives. The debate lasted two thousand years and never resolved.

This is not a historical curiosity. It is a pattern. Every ambitious classification system in the humanities follows the same arc: initial elegance → boundary problems → proliferating exceptions → contested revisions → eventual either abandonment or ossification. The arc is so predictable that it might itself deserve a name. Call it the categorization trap: the phenomenon whereby the pursuit of a perfect taxonomy consumes more intellectual energy than the taxonomy was meant to save.

Linnaeus and the Illusion of Natural Order

Carl Linnaeus published Systema Naturae in 1735, proposing a hierarchical classification of all living things. It was revolutionary, beautiful, and wrong in ways Linnaeus himself began to suspect. His system assumed fixed species, clear boundaries, and a divinely ordained natural order. Evolution — which he didn't know about — blew all three assumptions apart. Yet the Linnaean system persists, patched and extended, because replacing it entirely would be more disruptive than living with its flaws.

The lesson is subtle but important: a useful classification doesn't need to be correct. It needs to be functional. Linnaeus gave biology a shared vocabulary. That vocabulary enabled Darwin, who overthrew Linnaeus's theoretical foundations while keeping his filing system. The map was wrong, but people could navigate with it — and that was enough.

The DSM: Categorizing the Mind

Nowhere is the categorization trap more visible than in psychiatry. The Diagnostic and Statistical Manual of Mental Disorders (DSM) has gone through five major editions, each redefining what counts as a disorder. Homosexuality was a disorder until 1973. Asperger's syndrome appeared in DSM-IV (1994) and disappeared in DSM-5 (2013), absorbed into the autism spectrum. Grief was reclassifiable as depression starting in DSM-5, a change so controversial that the chair of DSM-IV publicly denounced it.

These are not minor taxonomic adjustments. They determine who gets treatment, who gets insurance coverage, who gets a diagnosis that shapes their identity. The categories are performative: they don't just describe reality, they construct it. As the philosopher Ian Hacking argued, psychiatric categories create "looping effects" — people classified as having a condition change their behavior in response to the classification, which in turn changes what the condition looks like, which forces the category to evolve.

The DSM is the categorization trap in its most consequential form: a system that must be definitive enough to guide clinical decisions, while simultaneously remaining open enough to absorb new evidence. It is, in practice, never both at once.

Wikipedia: The Infinite Edit War

For a contemporary example of the categorization trap, consider Wikipedia's category system. Wikipedia has over two million categories. Editors wage prolonged "edit wars" over whether articles belong in one category or another, whether categories should be split or merged, and whether category trees reflect "neutral" knowledge organization or embed cultural biases. The talk page for Wikipedia's categorization policy runs to hundreds of thousands of words — a quantity of text devoted to how to organize knowledge that rivals the knowledge itself.

The pattern repeats: the system must be both comprehensive and consistent, but comprehensiveness generates exceptions that undermine consistency. Every new edge case triggers a debate that consumes more effort than the case is worth. The categorization apparatus becomes an end in itself.

II. The Philosophical Impossibility of Perfect Categories

The categorization trap is not merely a practical problem. It reflects something philosophically deep about the relationship between language, thought, and reality. Several thinkers have argued that perfect categorization is not just difficult but impossible in principle.

Wittgenstein: Family Resemblance

In his Philosophical Investigations (1953), Ludwig Wittgenstein attacked the classical theory of categories — the idea, inherited from Aristotle, that categories are defined by necessary and sufficient conditions. His famous example was the word "game." What do board games, card games, ball games, and Olympic games have in common? Wittgenstein argued: nothing universal. There is no single feature shared by all games. Instead, there is a network of overlapping similarities — "family resemblances" — where some games share features with some other games, but no feature runs through all of them.

The implication is radical: many of our most important concepts — art, justice, religion, intelligence, argument from definition — cannot be captured by crisp, necessary-and-sufficient definitions. They are inherently fuzzy. Any taxonomy that forces them into sharp boxes will either exclude legitimate members or include illegitimate ones. The choice isn't between a good taxonomy and a bad one. It's between a useful approximation and a beautiful fiction.

Borges: The Chinese Encyclopedia

Jorge Luis Borges, in his 1942 essay "The Analytical Language of John Wilkins," describes a fictitious Chinese encyclopedia called the Celestial Emporium of Benevolent Knowledge, which divides animals into categories including: (a) those that belong to the Emperor, (b) embalmed ones, (c) those that are trained, (d) suckling pigs, (e) mermaids, (f) fabulous ones, (g) stray dogs, (h) those included in the present classification, (i) those that tremble as if they were mad, (j) innumerable ones, (k) those drawn with a very fine camelhair brush, (l) others, (m) those that have just broken a flower vase, and (n) those that from a long way off look like flies.

The passage is funny. It is also devastating. Borges's point is not that this particular taxonomy is absurd — it is that every taxonomy is, from a sufficiently external perspective, equally arbitrary. We don't notice the arbitrariness of our own categories because we inhabit them. They feel natural because they are ours. The Chinese encyclopedia feels alien because it is not. But the logic — grouping things by selected shared properties — is identical. The only difference is which properties were selected, and that selection is always cultural, always contingent, always debatable.

Lakoff: Women, Fire, and Dangerous Things

Cognitive linguist George Lakoff took Wittgenstein's insight empirical in his 1987 book Women, Fire, and Dangerous Things. The title comes from the Australian Aboriginal language Dyirbal, which has a grammatical category ("balan") that includes women, fire, and dangerous things. From an English perspective, this grouping is bizarre. From Dyirbal's internal logic, it is perfectly coherent.

Lakoff's broader argument is that categorization is not an objective mapping of pre-existing structure. It is a cognitive act shaped by embodied experience, cultural context, and metaphorical extension. Our categories are not mirrors of nature. They are tools of culture — and like all tools, they work better for some jobs than others.

This has direct implications for any attempt to build a universal taxonomy of anything — including critical thinking. The aspects that feel like "natural kinds" to a Western analytic philosopher might look arbitrary to a Confucian scholar. The false dichotomy between "universal" and "culturally relative" categories is itself a categorization error.

Foucault: The Order of Things

Michel Foucault opens The Order of Things (1966) with Borges's Chinese encyclopedia, using it to launch an investigation into the historical systems of thought — "epistemes" — that determine what counts as knowledge in any given era. Foucault's point is that categorization systems are not incremental improvements toward truth. They are epistemic ruptures: entirely different ways of organizing knowledge that cannot be ranked on a single scale of progress.

The Renaissance category system (based on resemblance and analogy) was not a primitive version of the Classical system (based on identity and difference), which was not a primitive version of the Modern system (based on function and history). They are incommensurable frameworks, each internally coherent, each invisible to itself. We cannot see our own episteme any more than a fish can see water.

For taxonomy builders, Foucault's message is humbling: your categories reflect your era's way of thinking, not reality's actual structure. Future generations will look at your taxonomy the way you look at medieval bestiaries — not as wrong, exactly, but as from a different world.

III. TellDear's Position: The Courage to Categorize Imperfectly

Given all this — the historical pattern of taxonomic failure, the philosophical arguments against perfect categories, the demonstrated impossibility of crisp boundaries for fuzzy concepts — why would anyone build a 534-aspect taxonomy of critical thinking?

Because the alternative is worse.

The categorization trap has a mirror image: the categorization paralysis. If you wait for the perfect taxonomy before acting, you will never act. The humanities are littered with projects that never shipped because the framework wasn't quite right, the categories weren't quite settled, the edge cases weren't quite resolved. Decades of committee work producing a classification system that is theoretically unimpeachable and practically useless — because by the time it's finished, the world has moved on.

TellDear chose a different path: ship the imperfect taxonomy, then iterate. Our 534 aspects across 6 dimensions (D1: Argumentation, D2: Manipulation & Propaganda, D3: Cognitive Biases, D4: Statistical Deception, D5: Argumentation Schemes, D6: Digital Information Literacy) are not a claim that this is The Correct Way To Organize Critical Thinking. They are a working hypothesis — a scaffold that enables analysis now, while remaining open to restructuring.

Some aspects overlap. Argument from Verbal Classification and Deceptive Framing are arguably two sides of the same phenomenon. Confirmation Bias lives in D3 (Cognitive Biases) but also operates as a mechanism behind several D1 (Argumentation) fallacies. The boundary between False Dichotomy and False Dilemma is contested — are they the same fallacy or different? TellDear treats them as related but distinct. Other taxonomies merge them. Neither choice is objectively correct.

We know this. We shipped anyway. Here is why.

The Pragmatic Argument

A taxonomy that helps people identify Strawman Fallacies in political debates today is more valuable than a theoretically perfect taxonomy that won't be ready for a decade. Critical thinking education doesn't have the luxury of waiting. Disinformation operates on news cycles, not academic publication schedules. The Nirvana Fallacy — rejecting a practical solution because it isn't perfect — applies to taxonomy design itself.

This is where TellDear becomes a case study for its own content: we recognize the Nirvana Fallacy as an aspect in our system, and we apply it to the system itself. A taxonomy that commits the Nirvana Fallacy about its own completeness would be self-refuting. (See also: Zero-Cost Critique, which examines how easy it is to criticize systems without offering functional alternatives.)

The Map-Territory Distinction

Alfred Korzybski's famous dictum — "the map is not the territory" — is often invoked as a reason to distrust maps. But the correct inference is the opposite: precisely because the map is not the territory, we need maps. The territory is too complex to navigate without simplification. A map that captured every feature of the territory at 1:1 scale would be the territory itself — and therefore useless as a map.

TellDear's taxonomy is a map. It simplifies. It distorts. It omits. That is what makes it navigable. The question isn't whether it's a perfect representation — it's not, and cannot be. The question is whether it helps people navigate the territory of critical thinking better than they could without it. We believe it does. The evidence from our analysis tools — where users consistently identify manipulation patterns they would have missed without the taxonomy — supports this belief.

IV. AI as Game-Changer: Refactoring Humanities

Here is where the story takes an unexpected turn. The categorization trap has persisted for millennia because refactoring a taxonomy is extraordinarily expensive. When Linnaeus's system needed updating, it took generations of biologists decades of work. When the DSM gets revised, it consumes years of committee deliberation and millions of dollars. When Wikipedia's category tree needs restructuring, it requires thousands of editor-hours.

The expense is not primarily intellectual. It is mechanical. Reorganizing a taxonomy means:

• Re-examining every item against new criteria
• Identifying items that belong in different categories under the new scheme
• Updating all cross-references and dependencies
• Propagating changes through every system that uses the taxonomy
• Verifying consistency at scale

For a taxonomy of 534 items with rich cross-references, this is thousands of hours of human work. Or — and this is the shift that changes everything — it is a few hours of work with a Large Language Model.

What LLMs Actually Enable

Large Language Models don't solve the philosophical problem of categorization. Wittgenstein's critique remains valid. Borges's encyclopedia is still funny. Foucault's epistemes are still incommensurable. But LLMs solve the mechanical problem — the reason the categorization trap is a trap rather than merely a challenge.

With LLMs, you can:

• Propose alternative taxonomic structures and immediately test them against your full dataset ("Reclassify all 534 aspects into a 4-dimension scheme instead of 6. Show me what breaks.")
• Identify boundary cases at scale ("Which aspects have the weakest fit to their current dimension? Rank by ambiguity.")
• Generate cross-references that humans would miss ("For each aspect, identify the three most conceptually related aspects in other dimensions.")
• Simulate the impact of restructuring before committing ("If we merge False Dichotomy and False Dilemma, what downstream effects does this have on the analysis pipeline?")
• Translate restructured taxonomies across languages simultaneously, maintaining conceptual consistency

This doesn't make categorization easy. It makes categorization agile. The trap works because the cost of being wrong is too high — you invest years building a taxonomy, and changing it would cost years more. But if changing it costs hours instead of years, being wrong is no longer catastrophic. You can afford to be wrong. You can afford to ship. You can afford to iterate.

"Refactoring Humanities" — A New Paradigm

Software engineers have a term for this: refactoring — restructuring existing code without changing its external behavior. Refactoring is what allows software to evolve without collapsing under its own complexity. The reason software can be continuously improved while buildings cannot is not that software is simpler — it's that the cost of restructuring software is dramatically lower than the cost of restructuring a building.

AI makes taxonomy work more like software and less like architecture. "Refactoring humanities" — restructuring knowledge systems rapidly, safely, and iteratively — is now a realistic research program. The implications extend far beyond TellDear:

• Library science: Reclassifying collections as knowledge systems evolve, without decades of manual recataloging
• Medical taxonomy: Updating diagnostic criteria more frequently, with AI identifying inconsistencies and gaps
• Legal codification: Restructuring legal codes to reflect new precedents, with AI identifying affected cross-references
• Educational standards: Updating curricula frameworks without multi-year committee processes

The paradigm shift is not "AI creates better categories" (it doesn't — the philosophical problems remain). The paradigm shift is "AI makes categorization cheap enough to iterate." And iteration, not perfection, is how good systems emerge.

V. Lessons from Software Engineering

If humanities taxonomies are becoming more like software, what can the humanities learn from software engineering's hard-won lessons about managing complexity?

Agile Taxonomies

The software industry spent decades trying to design perfect systems upfront ("waterfall" methodology) before learning, painfully, that iterative development works better. The Agile Manifesto (2001) codified the lesson: "Working software over comprehensive documentation." "Responding to change over following a plan."

The taxonomic equivalent: a working classification system over a comprehensive theoretical framework. Responding to new evidence over defending existing categories.

TellDear practices agile taxonomy. Our aspect list has changed since launch. Aspects have been added, renamed, merged, and re-dimensioned. We don't treat this as failure — we treat it as the system working correctly. A taxonomy that never changes is either perfect (unlikely) or ossified (likely).

Semantic Versioning for Knowledge Systems

Software uses semantic versioning (e.g., v2.1.3) to communicate the nature of changes. Major version bumps signal breaking changes. Minor versions add features. Patches fix bugs. Everyone using the software knows what to expect.

Knowledge systems could benefit from similar conventions:

• Major version: Fundamental restructuring (merging or splitting dimensions, changing the ontological framework)
• Minor version: Adding new aspects, refining descriptions, adding cross-references
• Patch version: Correcting errors, updating examples, fixing inconsistencies

This would make taxonomic change legible. Instead of the DSM's traumatic every-15-years overhauls, knowledge systems could evolve continuously, with clear signals about what has changed and why. Users could "pin" to a version when stability matters and "upgrade" when they're ready for improvements.

"Ship Then Iterate"

The most important lesson from software engineering is also the simplest: shipping beats planning. Not because planning is bad — but because reality is the best test of any plan, and you can't encounter reality without shipping. Every day your taxonomy sits unreleased, it's being tested by zero users and improved by zero feedback loops.

Reid Hoffman's maxim — "If you're not embarrassed by the first version of your product, you've launched too late" — applies to knowledge systems too. TellDear's first taxonomy was rougher than the current one. The current one is rougher than the next one will be. That trajectory — launch, learn, improve — is not a compromise. It is the correct methodology for building knowledge systems in a world where perfect knowledge is impossible.

VI. Meta-Honesty: TellDear as a Case Study for Its Own Thesis

There is something pleasingly recursive about this article. TellDear is a platform that helps people analyze arguments, identify biases, and detect manipulation. This article argues that TellDear's own taxonomy is imperfect — and that imperfection is a feature, not a bug.

This is not false modesty. It is meta-honesty: applying your own analytical framework to yourself. If TellDear's tools can identify the Nirvana Fallacy in other people's arguments but can't acknowledge it in their own taxonomy design, the system is intellectually dishonest. If we teach users to detect False Equivalence but treat all our own categories as equally valid, we're not practicing what we preach.

So here is the honest accounting:

• Some of our aspects are better defined than others. The argumentation fallacies (D1) have centuries of philosophical refinement behind them. The digital literacy aspects (D6) are newer and less battle-tested.
• Some cross-dimensional relationships are underexplored. How exactly does Confirmation Bias (D3) interact with Framing Effect (D3) in the context of Agenda Setting (D2)? We have hypotheses, not proof.
• Our dimension boundaries are pragmatic, not principled. D5 (Argumentation Schemes) could arguably be a subset of D1 (Argumentation). We separated them for pedagogical clarity, not ontological necessity.
• 534 aspects is a snapshot. Some aspects will be merged. Others will be split. New ones will be added. The number will change. If it doesn't change, something has gone wrong.

This meta-honesty is itself a form of intellectual integrity — and it's what distinguishes a living knowledge system from a dead one. The DSM doesn't usually publish articles about its own categorization problems. Wikipedia's guidelines don't typically cite Borges. TellDear does. That's a choice, and we think it's the right one.

VII. Conclusion: Productive Imperfection

The categorization trap is real. The humanities have lost centuries to it. Perfect categories are philosophically impossible, and the attempt to achieve them is often more paralyzing than the problems they're meant to solve.

But the answer is not to abandon categorization. Categories are how we think. They are how we communicate. They are how we build tools that help others think and communicate. The answer is to categorize with open eyes — knowing that your system is imperfect, planning for revision, and using the best tools available to make revision cheap and fast.

TellDear's 534 aspects are not a monument. They are a working draft. A draft that helps people identify strawman arguments and false equivalences and manufactured consent right now, today, while philosophers continue debating whether our categories are optimally structured. We welcome that debate. We'll participate in it. And when it produces insights, we'll use AI to refactor our taxonomy in hours rather than years.

Aristotle started this conversation 2,400 years ago. Wittgenstein showed us it can never be finished. Borges showed us it's funnier than we think. And now, for the first time in intellectual history, we have tools that let us iterate on the conversation at the speed of thought rather than the speed of committee.

That's not the end of the categorization trap. It's the beginning of productive imperfection. Ship it. Name it. Use it. Fix it. Repeat.

This article explores themes from across TellDear's six dimensions. Relevant aspects include: Argument from Definition · Argument from Verbal Classification · Nirvana Fallacy · False Dichotomy · False Equivalence · Confirmation Bias · Deceptive Framing · Framing Effect · Strawman Fallacy · Manufacturing Consent · Agenda Setting. See also: Adaptive Shortcuts · Zero-Cost Critique · Hollow Rhetoric · Why Six Dimensions · The Architecture of Bad Choices.