The paper argues that LLM-generated text suffers from unusually uniform per-sentence information density, heavy nominalization, and a hedging register that collectively raise cognitive load. Readers were slower and retained less from LLM passages even though they rated the writing as equally well-written on blind inspection, suggesting the deficit is a measurable processing problem rather than an aesthetic one.
The editorial frames the finding through surprisal theory: human writers spike information when introducing new ideas and relax into connective tissue so readers can consolidate. LLMs trained to minimize per-token surprisal produce suspiciously uniform text, which sounds easier but actually prevents the chunking and memory consolidation that comprehension depends on.
A paper making the rounds on Hacker News (173 points) takes a swing at a question practitioners have been muttering about for two years: why does so much AI-generated writing feel like wading through pudding? The authors run comprehension and processing-cost measurements on matched pairs of human- and LLM-written text and find that the gap isn't about grammar, factuality, or even coherence at the paragraph level. It's about information shape.
The paper's core claim: LLM prose exhibits unusually flat information density, heavy nominalization, and a hedging register that collectively raise cognitive load even when individual sentences are clean. Readers in the study were slower, retained less, and rated LLM passages as more effortful — despite rating them, on blind inspection, as equally "well-written." That mismatch between surface polish and actual comprehension is the finding worth paying attention to.
The sources it draws on are the usual suspects — Levy's surprisal theory, uniform-information-density work from Jaeger and Genzel, and more recent corpus studies on LLM stylometry. What's new is pinning the readability deficit to specific, measurable features rather than hand-waving about "GPT voice."
For years the critique of LLM writing has been aesthetic: it sounds like a LinkedIn post wrote a Wikipedia article. This paper reframes it as a measurable processing problem, and the mechanism it proposes is interesting enough to steal.
Human writers vary. Sentences spike in information when a new idea is introduced and relax into connective tissue when the reader needs a breather. Good prose is a wave. LLMs, trained to minimize per-token surprisal across huge corpora, produce text that is suspiciously uniform — every sentence carries roughly the same informational weight, which sounds like it should be easier but actually makes the text harder to chunk and remember. Your brain uses the dips to consolidate; without them, everything feels equally important and nothing sticks.
The second mechanism is nominalization. Compare "we deployed the fix" to "the deployment of the fix was completed." Both parse. One of them is how humans talk, the other is how a model trained on technical reports talks. The paper finds LLMs reach for the latter construction two to three times more often, which inflates word count and buries the verb — the one word that tells you what actually happened. If you've ever read an AI-generated changelog that says a lot while telling you nothing, this is the mechanism.
Third, hedging. The RLHF-era models were trained hard to avoid confident wrong answers, and the side effect is prose stuffed with "it's worth noting," "generally speaking," "in many cases." Each hedge is cheap on its own. Stack them and the reader's model of what the writer actually believes gets fuzzy, which is cognitively expensive. The paper's regressions suggest hedge density alone accounts for a meaningful chunk of the measured processing penalty.
There's a counterargument worth taking seriously: maybe the problem is sample selection. If humans mostly see LLM text in low-stakes contexts (slop blog posts, summarized emails, autogen docs), of course they'll report it as harder — they're reading content they didn't want in the first place. The authors try to control for this by matching topics and asking readers to score blind, but the control isn't airtight. The honest version of the finding is: LLM prose has a measurable readability tax, but how much of that tax is intrinsic to the generation process versus the result of a training-data register that happens to overlap with corporate blandspeak is still open.
If you ship LLM-generated text anywhere a human has to read it carefully — docs, PR descriptions, incident reports, release notes, marketing copy — the paper suggests three cheap interventions that are easier to implement than fine-tuning.
First, post-process for verbs. A regex-level pass that flags nominalizations ("the implementation of," "the completion of," "the evaluation of") and asks the model to rewrite with the underlying verb is maybe the highest-leverage edit you can make to LLM output, and it costs almost nothing. You don't need a smarter model; you need a dumber linter. Several teams have reported anecdotally that a "strip nominalizations" rewrite pass noticeably improves perceived quality of generated documentation, and the paper gives you the mechanism to argue for it.
Second, vary sentence length deliberately in your prompts. Models will happily produce prose where every sentence is 18 to 24 words. Ask explicitly for a mix — "use some short sentences, four to seven words, between the longer ones" — and the output picks up the rhythm humans actually read with. This maps directly to the information-density finding: short sentences are the dips your reader's brain needs.
Third, kill the hedges in your system prompt. Instructions like "state conclusions directly; do not hedge unless uncertainty is genuine" measurably reduce the "it's worth noting" problem. The models can do it; they just default to the trained-safe register.
For evaluation: if you have a dataset of human-written references, the paper's methodology — matched-pair comprehension scoring — is reproducible with any crowdsourcing platform and a few hundred dollars. If you care about whether your LLM-generated content actually lands with readers, this is a better benchmark than any automated readability score, which tends to just count syllables and miss the actual pathology.
The real target of this research isn't "AI writing bad" — that's a Twitter take, not a paper. It's the beginning of a measurement vocabulary for a problem that's about to matter a lot more: as LLM-generated text saturates the training corpus for next-generation models, any readability deficit that gets reinforced now compounds. If models learn from the uniform-density, nominalization-heavy register their predecessors produced, the gap widens. The honest practitioner takeaway is that the "AI voice" problem is a prose-engineering problem, not a model-capability problem, and the fixes available today are mostly mechanical. Fix the output, not the model.
One construction I'm surprised the paper doesn't mention (although, to be fair, I'm sure they had plenty of options to choose from... and you have to draw the line somewhere :P) is that LLMs love to use transitive verbs intransitively[0], in a way I almost never see in human writing.
LLMs are showing us that a lot of what we think of as cognition is actually part of the elaborate symbol system which is Language, as opposed to a property of "mind" (either biologically or artificially instantiated.)LLM-generated text is hard to understand because while it's perfectl
Consider too that AI training corpuses include vast amounts of commercial speech which is designed to numb the critical faculties and produce action; likewise they include numerous opinion articles, which are designed to persuade and often leverage rhetorical fallacies, whether cynically or as a ref
As usual most commenters reply to headlines, so I quote their conclusions.>We take these results as a proof of concept that a length penalty used during reinforcement learning with GRPO can directly affect grammatical features. We note that the model was not directly instructed to use compressed
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
As a technical writer, I believe default AI responses are optimized for what people _think_ is good writing rather than what actually is good writing.My guess is years of RLHF optimized for rater's first impressions of how well-written something was. It uses lists! It has headings! It uses fun