Study links hard-to-read AI text to training that rewards information density over clarity

What the study measured

A paper posted to arXiv on October 21, which has since drawn attention on Hacker News, takes a quantitative look at a familiar complaint: prose written by large language models increasingly feels like work to get through. Rather than treating readability as a matter of taste, the authors scored machine-generated text against features that psycholinguistics has already linked to processing effort in human readers — constructions such as object relative clauses, long-distance dependencies between related words, and rare word combinations.

The team built a fresh corpus of AI-generated text spanning a range of domains and ran two comparisons. The first pitted model output against paraphrases produced through round-trip translation, a mechanical rewriting process. Across the board, the AI models reached for the harder constructions more often than those paraphrases did.

The second comparison, against human writing, produced the sharpest result. All nine frontier and near-frontier models examined wrote with more inanimate subjects attached to verbs that ordinarily expect an animate actor, and with more low-frequency n-grams, than people did. Most of the models also produced longer dependencies, denser stacks of nouns, and more reduced object relatives in conversational settings. The paper also reports meaningful differences between model families, suggesting the effect is not uniform across the industry.

Why models drift this way

The authors' proposed explanation starts from a difference in objectives. Human language, they argue, is shaped by the need to get a message across — writers and speakers constantly adjust to their audience. Modern LLM training, by contrast, increasingly optimizes for performance on math, code and reasoning tasks. In those domains the useful pressure is to pack as much information as possible into the output, not to keep it easy for a person to parse. Over time, that pressure shows up in the general-purpose prose the models produce.

Testing the hypothesis

To check the idea, the researchers fine-tuned a language model with a length penalty — a training signal that effectively rewards compression. According to the paper, that single change was enough to make the model's output drift away from human-like linguistic patterns. The experiment is described as a proof of concept rather than a definitive demonstration, but it offers a plausible mechanism: reward density, and density is what you get, with readability paying the price.

Why it matters

For anyone publishing with LLM assistance, the study reframes an aesthetic gripe as a measurable cost. If model text systematically leans on constructions that raise reader effort, then editing for readability is not cosmetic polish — it is compensation for a known bias in the tool. Documentation, tutorials and other material aimed at broad or time-poor audiences are the most exposed.

The variation across model families also suggests a practical lever: some systems appear to be heavier offenders than others, and the specific features the paper measures — sentence complexity, dependency length, rare word choices — are the kinds of things a publishing pipeline could audit automatically before content goes out.

The longer-term implication is the one the authors emphasize. They frame the gap as a form of misalignment between what humans find easy to process and what current training objectives reward, and they point toward remedies along those lines — training setups that value clear communication, not just correct or compressed answers. Until such fixes arrive, the burden of keeping AI-assisted text readable remains with the humans in the loop.