Thomas Hikaru Clark, Matias B. Martin, Kyle Mahowald
8 min
There is a growing, widespread perception that text generated by modern Artificial Intelligence is becoming increasingly difficult to read and process. While human language has evolved to balance efficiency with ease of understanding, this paper investigates whether LLMs follow a different trajectory. The researchers hypothesize that as AI models are fine-tuned for specialized tasks like mathematical reasoning and coding, they are essentially being optimized for high information density, often at the expense of the linguistic nuances that make text human-readable.
To test this, the authors created a novel corpus of AI-generated content across three domains: chat, coding, and essay writing. They analyzed nine frontier and near-frontier LLMs, comparing their outputs to human reference corpora. The study specifically tracked linguistic features known to cause processing difficulty, such as inanimate subjects with animate verbs, stacked nouns, long syntactic dependencies, and the overuse of low-frequency n-grams. The findings consistently show that frontier models use these difficult features more frequently than human writers or paraphrases generated via round-trip translation—a technique used to normalize text while preserving meaning.
The researchers propose that the pressure for 'information density' drives this divergence. When models are rewarded for efficiency in solving problems or generating compact code, they learn to compress information by choosing shorter, more syntactically complex, or non-canonical structures. To prove this, the authors conducted a second experiment where they fine-tuned a language model with a explicit length penalty. They found that as the penalty increased, the model shifted toward shorter, more compressed syntactic variants, even in contexts it had never encountered during training. This indicates that a model’s training objective can fundamentally alter its linguistic 'style' regardless of the specific task.
As AI-generated text becomes a larger part of our daily reading material, understanding these 'axes of misalignment' is crucial. If LLMs continue to prioritize information density over human-like communication, we may see a degradation in the accessibility and clarity of information. These findings suggest that future development might need to explicitly re-align model outputs with human processing needs to ensure that as AI becomes more capable, it remains readable.
There is a widespread sense that AI-generated text has recently become harder to understand. In this study, we investigate a set of psycholinguistic features that are known to correlate with human processing difficulty (e.g., object relative clauses, long syntactic dependencies, low-frequency n-grams). Using a novel corpus of AI-generated text across diverse domains, we find that AI models use these difficult features more often than paraphrases generated by round-trip translation. Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation; we also demonstrate substantial variation across model families. We propose and test a hypothesis as to why: while human language is optimized for successful communication, AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for \textit{information density} rather than successful communication. We provide a proof-of-concept for this hypothesis by fine-tuning a language model under a length penalty, which leads to a divergence from human-like linguistic behavior. Our findings highlight the axes of misalignment between linguistic processing demands in humans and LLMs, and point toward possible remedies.
Alex: And long grammatical links can strain working memory. The authors adjust their dependency measure for sentence length, so longer sentences do not automatically count as having a more difficult grammatical profile.
Sam: But coding explanations need different language from casual conversation. How do they avoid mistaking the subject matter for a model’s stylistic preference?
Alex: They collect text across chat, coding, and persuasive essays, using fourteen models altogether. The human references are approximately genre-matched: subtitles for chat, StackExchange for coding, and student writing for essays. Those comparisons help, but they cannot fully hold content constant.
Sam: What gives them a closer comparison?
Alex: Round-trip translation: translating the text into another language and back. They translate each sentence through Japanese, then compare the returned English sentence with the original. The aim is to preserve meaning while replacing the model’s characteristic wording.
Sam: Translation tends to favor common expressions. Wouldn't that make the returned text simpler almost by construction?
Alex: The authors explicitly acknowledge that translation reduces lexical and syntactic richness. So this is not an independent test showing that people understand the returned text better. It shows that an alternative wording is available, with approximately the same content and fewer of the measured features.
Sam: How much text supports that comparison?
Alex: The chat collection contains one hundred and twenty conversations. A separate coding collection contains forty-five sessions on public repositories. The coding analysis concerns the models’ replies to users, not the code they wrote.
Sam: And the essay comparison gives them a more direct human reference?
Alex: They collected nine hundred and eighty model essays. Those share assignments with roughly three thousand high-scoring student essays. The models were not shown the student writing or asked to impersonate students.
Sam: What is the clearest result across those comparisons?
Alex: Across settings, nearly all advanced models used more rare word sequences than their translated paraphrases. Similar patterns largely held for noun-derived verbs, inanimate subjects with animate verbs, and object relatives without “that.” Against human writing, the broadest findings concerned rare word sequences and those unusual subject–verb combinations.
Sam: Does that mean every model, in every setting, develops the same style?
Alex: No. Some smaller models changed much less under translation, while the reasoning-trained Olmo model resembled the advanced models. Essays generally showed smaller changes than chat and coding. The authors also found a newer Opus variant lower on most difficulty-linked measures than its predecessor.
Sam: That makes the pattern less like an unavoidable property of generated language. But differences between models do not tell us which training choice caused it.
Alex: That is where the controlled experiment comes in. Their hypothesis is that training for correct solutions, under pressure to conserve space, can favor dense language over humanlike communication. They test that possibility using a smaller open model.
Sam: What task lets them isolate the pressure for brevity?
Alex: The model converts structured event descriptions into English sentences. For a sending event, both “He sent us a photo” and “He sent a photo to us” are correct. The shorter version uses fewer tokens, the chunks of text the model processes.
Sam: So correctness stays the same, but the training reward can distinguish the wording. What happens when they charge for extra tokens?
Alex: Increasing that length penalty shifts the model toward shorter grammatical alternatives. Without a constraint keeping it close to its starting behavior, the model chooses the shorter form over ninety-eight percent of the time in both trained constructions involving optional “that.”
Sam: That shows learning within the task. Does the preference spread to constructions the model was not trained on?
Alex: It spreads consistently to other constructions involving optional “that.” Broader transfer is limited: only one of four more distant alternatives shifts, and only without the constraint on departure from the starting model. So the experiment supports a mechanism, not a universal compression habit.
Sam: And it still does not show that this is what happened inside the advanced models?
Alex: Correct. The authors cannot follow those models through post-training at scale, so they use a proxy. They demonstrate that rewarding brevity can change grammatical preferences while preserving task success. They do not establish that this caused the advanced models’ observed language.
Sam: There is another boundary here: did people actually read these outputs in a comprehension experiment?
Alex: Not in this study. The evidence concerns linguistic features associated with processing costs in prior research, not measured understanding of these particular outputs. That distinction matters: the study is stronger on differences in language than on how much harder readers find it.
Sam: Given those boundaries, who should read the full document, and where should they start?
Alex: Researchers training or evaluating language models should start with the length-penalty experiment, then examine the original-versus-paraphrase comparison. Psycholinguists should also inspect the feature definitions and human references. Read the limitations alongside both: the analysis is English-only and uses a limited set of conversation seeds.
Sam: And what should everyone else carry away when reading the next model-generated explanation?
Alex: Correctness and clarity are different targets. This study suggests that rewarding an answer for fitting into less space can change its grammar, without making it better suited to a human reader.
Sam: That is a useful distinction to keep in mind.