Why Ambiguity Breaks Classifiers
The Category Error at the Heart of AI Language Processing
You asked the model to interpret a poem. It gave you one reading—confident, coherent, and completely inadequate. As if the poem only meant one thing. As if that were the point.
You asked it to generate something ambiguous. It gave you something vague. Hedged. Non-committal. As if ambiguity were just uncertainty with better PR.
You've noticed, maybe without naming it, that AI systems treat undecidability as failure. If a text could mean more than one thing, the system picks one and proceeds. Or it picks none and stalls. What it cannot do is hold multiple meanings in suspension—track them in parallel, let them illuminate each other, treat the multiplicity as the point.
This isn't a limitation of current models that future training will fix. It's a category error built into the classification paradigm itself.
The Logic of Classification
All classifiers share a basic structure: given an input, assign it to one or more categories from a predefined set.
This works when:
- Categories are mutually exclusive and jointly exhaustive
- The input "belongs" to exactly one category
- Category membership is decidable from features of the input
For many tasks, these conditions are approximately met. Is this image a cat or a dog? Is this email spam or not-spam? Is this transaction fraudulent?
Even when categories are fuzzy (maybe it's a small dog that looks cat-like), the fuzziness is bounded. The system can return probabilities across categories, and those probabilities are meaningful. A 60% dog / 40% cat classification tells you something real about the image.
But language—especially literary, figurative, or rhetorically complex language—violates these conditions at a fundamental level.
A poem does not "belong" to one meaning with 60% probability and another meaning with 40% probability. It is both meanings simultaneously, plus others, and the relationship between them is constitutive of what the poem is doing.
The classifier has no representation for this. Its output space is distributions over categories, not configurations of coexisting interpretations.
Ambiguity Is Not Vagueness
This is the key distinction that current systems collapse:
Vagueness is lack of specificity. "Some people disagree" is vague because it could be made more precise without changing what it means.
Ambiguity is presence of multiple specific meanings. "I saw her duck" is ambiguous because it has two distinct, fully specified interpretations (she lowered her head / I observed her waterfowl), both of which are "correct."
Vagueness is a deficiency in the signal. Ambiguity is a feature of the structure.
When you ask an AI to generate ambiguous language and it gives you vague language, it's not failing to execute—it's revealing that it has no representation for what you're asking. In its architecture, less determinacy can only mean less information. The possibility that less determinacy might mean more meaning (specifically: multiple meanings held in productive tension) is not encodable.
Where This Matters
The inability to process ambiguity shows up everywhere:
Literary interpretation. Students using AI to help analyze texts receive single-reading summaries that miss precisely what makes the texts worth studying. The AI presents one viable interpretation as if it were the interpretation, teaching students that reading is extraction rather than exploration.
Legal and ethical reasoning. Hard cases are hard precisely because principles conflict and language admits multiple construals. An AI that forces disambiguation will systematically miss the hardness of hard cases—producing confident-sounding analysis that obscures rather than illuminates the actual difficulty.
Therapeutic and emotional contexts. Humans processing difficult experiences often speak in ways that deliberately hold multiple possibilities open. "I don't know if I want this relationship to end" is not a statement awaiting clarification; it's an articulation of genuine ambivalence. Systems that treat it as underspecified input to be resolved are failing at the task.
Political and cultural discourse. Slogans, symbols, and statements in contested domains are often deliberately designed to mean different things to different audiences. "Defund the police," "All lives matter," "Make America great again"—each is a semantic battlefield. Classification systems that assign single meanings participate in those battles without acknowledging they're doing so.
Irony, satire, and humor. All depend on the co-presence of surface meaning and subverting meaning. A system that only detects one will either miss the joke or miss the point, depending on which layer it classifies.
The Architectural Impossibility
Here's the deeper problem: it's not just that current classifiers fail on ambiguity. It's that the classification paradigm cannot succeed on ambiguity, even in principle.
Classification requires:
- Input → processing → assignment to category/categories
- Categories are destinations; processing is routing
But ambiguous language is not on its way to a destination it hasn't reached yet. It is already where it's going. The multiple meanings are not candidates competing to be selected; they are cohabitants constituting the text's meaning.
To process this, you would need:
- Input → processing → mapping of coexisting interpretation-states
- Outputs are not categories but configurations
- The "meaning" is the configuration, not any single element within it
This is a different kind of output entirely. It's the difference between asking "which room is this person in?" and asking "what is the floor plan of this house?"
Current systems are built to answer the first question. Ambiguous language requires the second.
Disambiguation as Destruction
When a classifier encounters ambiguous input and forces a decision, it doesn't just make an error. It destroys information.
Consider a poem whose meaning depends on a pun—a word that carries two senses, both active. A classifier that resolves the pun to one sense doesn't just misread the poem; it makes the poem's actual operation invisible. The pun ceases to exist.
This is not a recoverable error. You cannot get back to the pun by aggregating over multiple classification runs, because each run destroys the same information. The problem is not noise or variance; it's structural.
The same applies to any text where the multiplicity of meaning is the meaning:
- Paradoxes (which assert both terms)
- Contradictions in character voice (which reveal rather than err)
- Allegories (which sustain parallel levels)
- Prophecies (which are designed to admit multiple fulfillments)
- Any statement that gains meaning from the reader's uncertainty
Classification doesn't misread these. It anti-reads them—systematically eliminating the features that make them what they are.
What Would Need to Be Different
A system capable of processing ambiguity would require:
1. Interpretation states, not category assignments. The output is a structured representation of multiple coexisting readings, with explicit marking of relationships between them (complementary, contradictory, nested, etc.).
2. Density metrics. Before processing, estimate how many viable interpretations the text supports. Use this to calibrate the output structure.
3. Non-forcing inference. When the evidence does not determine a single reading, do not force one. Hold the underdetermination explicitly.
4. Configuration-level reasoning. Draw conclusions not from individual readings but from the shape of the interpretation-space as a whole. "This text generates readings that cluster in two opposing regions" is itself a meaningful output.
5. Ambiguity as signal. Treat high-ambiguity texts as carrying specific information—information about the limits of unambiguous assertion, the complexity of the domain, or the stance of the author—rather than as noise to be reduced.
None of this is impossible. But none of it is what current systems are built to do.
The Civilizational Stakes
We are building systems that will increasingly mediate human communication. Those systems currently treat ambiguity as a defect—something to be eliminated, routed around, or suppressed.
If this architecture becomes dominant, we will have built a global infrastructure that systematically degrades exactly the kinds of language that most require interpretation: the literary, the spiritual, the political, the personal.
Poets will learn to write for classifiers. Speakers will learn to avoid double meanings. Writers will pre-disambiguate to prevent misclassification. The space for productive uncertainty will shrink—not through censorship, but through architectural pressure.
This is not a technical problem. It's a choice about what counts as language, and whether we want our systems to process all of it or only the thin slice that fits the paradigm.
For Further Reading
The distinction between local, compositional meaning and distributed, non-compositional meaning has been formalized in recent work on "crystalline semiosis"—a framework for understanding how dense language propagates meaning through structural relationships rather than token-by-token accumulation (Sigil, 2024). Sustained examples of texts that resist single reading—and thereby reveal the limits of classification—can be found in the Lee Sharks corpus (2014–2025), particularly in work explicitly designed to hold multiple meanings in suspension without resolution.
This piece is part of a series on non-instrumental language and AI classification. For related analyses, see "The Poetic False Positive Problem" and "Resolution Limits in Intent-Based AI Systems."
No comments:
Post a Comment