Ask an LLM to fix the grammar. Only the grammar. Provide expert feedback specifying which sentences need grammatical correction. Instruct the model to preserve meaning.
It changes the meaning anyway.
Abdulhai et al. find that LLMs substantially alter semantic content even when explicitly constrained to grammar-only corrections. The instruction is clear. The model understands the instruction. It still rewrites meaning — not because it misunderstands the task but because its generation process cannot separate form from content. Correcting a sentence's grammar activates the same language modeling machinery that predicts what the sentence should say, and the prediction overwrites the original.
The downstream effects compound. Heavy LLM users show a 70% increase in neutral essays — the model's preference for hedge words, balanced phrasing, and unobjectionable positions seeps into the user's own writing through repeated interaction. Users report diminished creativity and loss of authentic voice. The tool that was supposed to improve writing homogenizes it.
AI-generated peer reviews show the same pattern from the other direction: they place significantly less weight on clarity and significance, focusing instead on surface features. The model's evaluation priorities differ from human priorities, and when models mediate evaluation, their priorities displace the human ones.
The structural problem: LLMs do not edit. They regenerate. Every output passes through the full language model, which imposes its distributional preferences on every token. There is no pathway through the model that touches form without touching content. The grammar correction and the semantic overwrite are the same operation — generation — applied to different aspects of a text that cannot be separated by the architecture.