What structural barriers prevent LLMs from making evaluative judgments about writing?

This explores why LLMs struggle to render genuine evaluative judgments about writing — distinguishing good from bad, novel from derivative, meaningful from mechanically correct. The corpus suggests the barrier isn't a knowledge gap you could patch with more training, but a set of structural features baked into how these models read and generate.

The clearest barrier is that LLMs can dissect the mechanics of a text without accessing what makes it worth evaluating. Models reliably extract explicit features — metaphoric mappings, stylistic signatures, discourse markers — but collapse on the implicit dimensions where literary meaning actually lives: they hit 24% accuracy on implicit relations, 32% on ambiguity recognition versus 90% for humans, and fail outright at evaluative stance-taking and preserving connotation Can LLMs truly understand literary meaning or just mechanics?. This maps onto a broader asymmetry: LLMs handle explicit, consistent structure well but fail wherever structure must be inferred Where exactly do LLMs break down with language structure?, and their grammatical competence itself degrades predictably as a sentence's structural complexity rises Does LLM grammatical performance decline with structural complexity?. Evaluating writing demands exactly the inferential, holding-it-all-at-once reading these models are weakest at.

A second, deeper barrier is what 'understanding' even means here. The potemkin failure mode shows models can explain a concept correctly, fail to apply it, and recognize their own failure — a pattern incompatible with human cognition, suggesting explanation and execution run on functionally disconnected pathways Can LLMs understand concepts they cannot apply?. So an LLM can produce a fluent paragraph about what makes prose strong while being unable to act on that standard when judging a real passage. The judgment and the criteria aren't wired together.

There's also a generative reason. Token prediction is trained to flow smoothly toward the training distribution, not to explore competing or counter positions — the process is smooth, so the claims come out smooth, multiplying without genuinely weighing alternatives Does LLM generation explore competing claims while producing text?. Real evaluation is adversarial: it requires holding a thesis against its weaknesses. When LLMs are pressed into the judge role, this shows in how easily they're swayed by surface — they fall for fake authority signals and rich formatting in semantics-agnostic, zero-shot ways Can LLM judges be fooled by fake credentials and formatting?, and their assessments shift with the emotional tone of the prompt even when the question is identical Does emotional tone in prompts change what information LLMs provide?. A judge that confuses formatting and tone with quality isn't evaluating; it's reacting to cues.

The interesting twist is that these barriers are partly bypassable — and how reveals their nature. When you stop asking for a holistic verdict and instead decompose evaluation into explicit stages, performance jumps: a three-stage novelty pipeline (extract claims, retrieve related work, compare) hit 86% reasoning alignment with human reviewers, beating holistic baselines Can structured pipelines make LLM novelty assessment reliable?, much as forcing models to check warrants and backing through structured critical questions catches reasoning failures plain chain-of-thought lets slide Can structured argument prompts make LLM reasoning more rigorous?. That these scaffolds help so much is itself the tell: the model can't assemble an evaluative judgment on its own: the structure has to be supplied from outside, because the architecture doesn't generate it.