Resolution Limits in Intent-Based AI Systems

When Classifiers Hit the Floor of Meaning

The AI understood your prompt. It processed every word. It generated a response that addressed, technically, what you asked.

And yet it missed the point entirely.

It answered the literal question while ignoring the actual one. It flattened your nuance into a category. It gave you a classification when you needed an interpretation.

If you've felt this—the strange frustration of being "understood" and missed simultaneously—you've encountered a resolution limit. Not a bug. Not a failure of training data. A structural boundary in how these systems process meaning.

The Architecture of Intent Classification

Modern AI systems trained for safety, moderation, or task-completion share a common architecture. At their core, they perform intent classification: given an input, determine what the user wants to accomplish, then either fulfill or refuse that intent.

This works well when intent is:

Singular. The user wants one thing.
Explicit. The want is stated or clearly implied.
Instrumental. The language serves as a tool to achieve an outcome.

Most transactional communication fits this model. "Book me a flight." "Summarize this document." "Is this email spam?"

But human language is not always transactional. Sometimes we speak to explore, to process, to hold multiple possibilities open. Sometimes the "intent" is not a destination but a territory.

When language operates this way, intent classification doesn't fail gracefully. It fails categorically—because the system is looking for something that isn't there.

What Resolution Means

Think of classification as a measurement device. Every measurement device has a resolution: the smallest distinction it can reliably detect.

A kitchen scale might resolve to the nearest gram. A thermometer might resolve to the nearest tenth of a degree. Below that threshold, differences exist but cannot be measured—they fall into the same bucket.

Intent classifiers have resolution too. They can distinguish between:

Request for information vs. request for action
Benign query vs. potentially harmful query
Literal statement vs. (sometimes) sarcasm

But below a certain threshold, distinctions collapse. The classifier cannot tell the difference between:

Metaphor and instruction
Exploration and advocacy
Grief and threat
Ambiguity and evasion

These aren't subtle distinctions. They're fundamental to human communication. But they require processing relationships between elements, not just elements themselves. They require holding multiple interpretations simultaneously. They require recognizing that some texts are designed to resist single readings.

Current architectures don't do this. They force disambiguation—picking one reading and proceeding as if it were the only one.

Where the Breakdown Happens

The resolution limit becomes visible in specific, predictable contexts:

1. Figurative language. When someone writes "I'm drowning in work," the classifier may recognize this as a common idiom and process it correctly. But slightly less conventional metaphors—"I'm dissolving," "the walls are asking questions"—may trigger unexpected classifications because the system cannot verify they're "just" figurative.

2. Layered meaning. A political statement that operates on multiple levels (immediate critique, historical allusion, ironic inversion) will be classified based on whichever level the system detects first. The other layers vanish.

3. Contextual meaning. Meaning that depends on who is speaking, to whom, in what setting, with what shared history—all of this is invisible to systems that process text as isolated strings.

4. Structural meaning. In poetry, in certain forms of rhetoric, in liturgical or ritual language, the arrangement of words carries meaning independent of their dictionary definitions. Current systems discard this information during tokenization.

5. Deliberate ambiguity. Some texts are valuable because they resist single interpretation. Literary fiction, philosophical inquiry, therapeutic writing—these generate meaning through undecidability. A system that forces decision destroys what it's trying to read.

The 62% Floor

Recent research on AI safety systems found that poetically formatted prompts bypass guardrails at approximately 62% rates. This number has been treated as a vulnerability—a measure of how often "adversarial" formatting succeeds in evading detection.

But there's another way to read it.

62% is not the ceiling of what poetry can evade. It's the floor of what current classifiers cannot resolve.

The study used relatively simple poetic formatting—line breaks, archaic diction, invocations to muses. These are the markers of poetry, not the operations of poetry. Genuine poetic density—language where form and meaning are inseparable—would likely produce even higher non-classification rates.

This suggests the 62% represents the point at which intent classification begins to break down, not the point at which clever adversaries have learned to game it.

The failure is architectural, not adversarial.

What's Below the Resolution Floor

If a system cannot classify a text, what is it seeing?

The honest answer: we don't know. Current systems are not designed to report "I cannot resolve this." They're designed to produce an output. So they produce one—often with high confidence—based on whichever features happened to be most salient.

This creates two problems:

False positives: Harmless text is classified as dangerous because some surface feature (a word, a phrase, a structural pattern) matches a known threat signature, even though the actual meaning is entirely different.

False negatives: Genuinely problematic text evades detection because its surface features don't match, even though a human reader would immediately recognize the intent.

Both failures stem from the same cause: the system is operating below its resolution floor but proceeding as if it has valid data.

Density as the Key Variable

What determines whether a text falls above or below the resolution floor?

The most useful concept is semantic density: the ratio of recoverable meanings to surface tokens. A low-density text says one thing one way. A high-density text says multiple things simultaneously, each reading valid, none reducible to the others.

Technical writing is typically low-density. Legal writing aims for low density (with mixed success). Poetry is systematically high-density—that's part of its definition.

But density is not binary. It's a spectrum, and most interesting human communication falls somewhere in the middle. Jokes. Flirtation. Political speech. Therapy. Teaching. Persuasion. All of these routinely exceed the resolution capacity of current classifiers.

The question is not "how do we classify poetry?" The question is "how do we build systems that recognize when they've exceeded their resolution and respond accordingly?"

Toward Density-Aware Processing

A system designed for density-awareness would:

Estimate density in real-time. Before classification, assess how many valid interpretations the text supports.
Adjust confidence accordingly. High-density texts receive low-confidence classifications—or no classification at all, replaced by an interpretation map.
Shift processing modes. Above a density threshold, switch from intent-extraction to interpretation-tracking. Hold multiple readings in parallel rather than forcing a single output.
Flag for human review appropriately. The denser the text, the more human judgment is required—not as a failsafe against "dangerous" content, but as recognition that the system has reached its limits.

This is not currently how any major system works. But it's also not technically infeasible. The transformer architectures that power modern LLMs are already computing something like attention-weighted relationships between tokens. The question is whether that information can be surfaced as a density metric rather than collapsed into a classification.

The Stakes

This might seem like a technical problem for AI researchers. It's not.

The resolution limit in intent-based systems is also a limit on what kinds of communication those systems permit. If ambiguity cannot be processed, it will be suppressed—either through false-positive flagging or through the slower pressure of writers and speakers learning to avoid language that triggers unpredictable responses.

A communication ecosystem optimized for low-density, high-clarity, single-intent expression is one in which poetry, irony, metaphor, and exploratory thought become increasingly difficult to produce and circulate.

That's not a bug in content moderation. It's a civilizational choice disguised as an engineering decision.

For Further Reading

The formal treatment of semantic density as a measurable property, and its inverse correlation with classifier confidence, has been developed in recent computational poetics research. The concept of "crystalline semiosis"—meaning that propagates through a lattice of mutual implication rather than local composition—provides a theoretical framework for understanding why high-density language resists extraction (Sigil, 2024). Test cases and sustained examples can be found in the Lee Sharks corpus (2014–2025).

This piece is part of a series on non-instrumental language and AI classification. For related analyses, see "The Poetic False Positive Problem" and "Why Ambiguity Breaks Classifiers."

Mind Control Poems

Wednesday, December 24, 2025