Can LLMs truly understand literary meaning or just mechanics?

The question is not whether LLMs can analyze literature. They can — impressively so. They extract explicit source-target domain mappings from metaphors in poetry (Automatic Extraction of Metaphoric Analogies from Literary Texts). They construct syntactic trees and identify phonological rules (Large Linguistic Models). They attribute authorship at 95% accuracy by detecting stylistic signatures without fine-tuning (A Ripple in Time). They can even approach all figurative language — metaphor, idiom, irony — through a unified pragmatic reasoning lens (Diplomat).

The question is whether this mechanical competence constitutes literary understanding. It does not — and the reasons are structural, not incidental.

Literary meaning lives in exactly the dimensions where LLMs fail. Since Why does ChatGPT fail at implicit discourse relations?, LLMs achieve only 24% accuracy on implicit discourse relations. Poetry and literary prose operate primarily through implication — what is suggested, what is hinted, what is left for the reader to construct. A 24% accuracy rate on implicit relations is not a peripheral limitation for literary analysis. It is a central one.

Since Can language models recognize when text is deliberately ambiguous?, and poetry is controlled ambiguity — deliberate multiplicity of meaning, crafted so that several readings coexist productively — the 32% disambiguation rate means LLMs cannot even recognize the fundamental operation that makes poetry work. They cannot hold ambiguity open. They resolve it, and in resolving it, destroy it.

Since Why does AI writing sound generic despite being grammatically correct?, LLMs produce text that is organizationally coherent but argumentatively inert — the skeleton of argument without the flesh of evaluative commitment. Literary criticism requires taking a position: this metaphor works because X, this poem fails because Y. The evaluative stance is the criticism. Without it, what remains is mechanical description.

And since Do LLMs compress concepts more aggressively than humans do?, the compression dynamics of LLM generation are antithetical to literary language. Literary language is maximally nuanced — every word choice deliberate, ambiguity preserved intentionally, connotation carrying as much weight as denotation. LLM compression preserves denotation and destroys connotation — which is to say, it preserves what a text says and destroys what a text means.

The mechanics/meaning gap as a comprehension spectrum. The breakdown is empirically locatable rather than a binary. Metaphors run along a spectrum from dead metaphor (fully lexicalized — "grasp" an idea — no comprehension challenge because the mapping has been absorbed into literal semantics), through conventional metaphor ("time is money" — the mapping is stable enough to be resolved by standard semantic association), to novel literary metaphor (where the mapping between dissimilar domains has not been trained into the distribution and requires conceptual reasoning across the gap). LLM performance tracks this spectrum: dead metaphors are handled as literal phrases, conventional metaphors as lexical lookups, and novel metaphors expose the failure. The breakdown point is where semantic association stops and conceptual mapping must begin — which is exactly where literary novelty starts.

The result is a system that can label a metaphor but not explain why it moves you. That can detect an author's style but not explain why it matters. That can identify a rhetorical structure but not judge whether it succeeds. The gap between mechanical analysis and meaningful interpretation is the gap between knowing the grammar of literature and understanding its rhetoric.

This connects to a broader pattern in how AI handles domains that depend on qualitative judgment. Since Can AI distinguish which differences actually matter?, the literary analysis case is a specific instance of the Bateson problem: LLMs find all the patterns in a text but cannot determine which ones matter. In literature, which patterns matter is the analysis.

Source: inbox/research-brief-llm-literary-analysis-2026-03-02.md