Can language models truly understand literary style?

GPT-2 + UMAP achieves approximately 95% accuracy attributing presidential State of the Union addresses to their authors, detecting both temporal patterns and individual stylistic signatures without any fine-tuning. Style is detectable even when "the Zeitgeist and language matter more than the actual politics" (A Ripple in Time: A Discontinuity in American History).

This is an impressive capability — and it reveals a boundary. LLMs can detect that an author has a distinctive style. They cannot explain why that style matters.

In literary prose, style is not decoration. It is content. Hemingway's short sentences are not a preference for brevity — they are a philosophy of communication: the unstated carries more weight than the stated, and every word must earn its place. Dickens's periodic sentences build to moral conclusions — the syntactic structure mirrors the argumentative structure. Faulkner's nested clauses perform the entanglement of memory, time, and consciousness that his novels are about. In each case, form and meaning are inseparable. Interpreting style as content is what literary criticism does.

Since Can imitating ChatGPT fool evaluators into thinking models improved?, we know that style is what LLMs (and human evaluators) detect most readily — coherence, fluency, apparent completeness. But since Why does AI writing sound generic despite being grammatically correct?, the evaluative dimension — judging whether a style choice succeeds, and why — remains structurally absent. Detection without evaluation is cataloguing without criticism.

Research on evaluation skill scaling confirms the mechanism: "readability and conciseness saturate early while logical reasoning improves with scale" (FLASK: Fine-grained Language Model Evaluation based on Alignment Skill Sets). Style detection saturates early because it operates on surface patterns. Style interpretation scales differently — or may not scale at all — because it requires the kind of evaluative commitment that alignment training actively suppresses.

The implication: LLMs can be excellent tools for stylometric analysis — detecting who wrote what, tracking style change over time, identifying signature patterns. But they cannot move from detection to interpretation. They cannot tell you that Lincoln's Gettysburg Address is extraordinary not because of what it says but because of how it says it — the way the syntax performs the democratic ideal it articulates. That judgment requires a reader who understands not just the pattern but its significance.