Why is AI generated language so hard to understand?
arxiv.org ·
A study submitted to arXiv on October 21, 2026 finds that AI-generated text uses several linguistic features associated with greater human processing difficulty more often than comparison texts. The researchers analyzed a new corpus of AI writing from diverse domains and measured features including object relative clauses, long syntactic dependencies, and low-frequency word sequences. AI models used difficult features more often than paraphrases produced through round-trip translation.
Against human writing, all nine frontier and near-frontier models examined used more inanimate subjects paired with animate verbs and more low-frequency n-grams. Most models also produced longer dependencies, more stacked nouns, and more reduced object relatives in conversational text, although the results varied substantially among model families. The researchers propose that human language is shaped for successful communication, while AI output is increasingly shaped by objectives such as mathematics, code, and reasoning that reward information density. As a proof of concept, fine-tuning a language model with a length penalty caused its language behavior to diverge from human-like patterns. The findings identify potential mismatches between the linguistic demands of human readers and language-model output and suggest that model training objectives may contribute to harder-to-understand text.
What people are saying
Commenters largely agreed that current assistant prose often feels less like communication than like an optimization artifact: polished on the surface, padded with headings and rhetorical flourishes, yet thin on the facts a reader needs.
What readers liked
- Base models can reportedly match ordinary human style better than their instruction-tuned counterparts, suggesting the problem is a reversible…
- Models still produce strong code, giving some readers confidence that prose quality can be treated as a presentation failure rather than evidence of…
Concerns raised
- Assistant prose often wins a superficial writing contest with headings, lists, and confident flourishes while failing the real test: delivering the…
- Technical readers find the output unusually hard to skim because meaning is diluted across flowery paragraphs, so skipping supposedly nonessential…
Read full discussion analysis
Source: Hacker News · Based on 15 sampled comments · Analyzed
Discussion overview
Commenters largely agreed that current assistant prose often feels less like communication than like an optimization artifact: polished on the surface, padded with headings and rhetorical flourishes, yet thin on the facts a reader needs. Technical writers in particular described models that mistake textbook-style corrections or generic completeness for audience-aware clarity. The result is prose that is hard to skim, overexplains irrelevant context, and sometimes uses syntactically odd constructions that make precise technical documentation unreliable.
The thread split over the cause. Some blamed the web-scale corpus: advertising, clickbait, persuasive essays, and other language built to create an impression rather than transfer understanding. Others argued that base models can sound more human than instruction-tuned ones, making post-training, engagement incentives, and reinforcement learning the more plausible culprits. Several readers focused on the paper's specific mechanism: rewards for short, verifiably correct reasoning can select compressed grammatical forms even without an explicit instruction to write densely. That tradeoff worries commenters because models increasingly use language internally for reasoning and agents, while users still need prose shaped around a human reader's knowledge and goals. A few noted that models remain useful for code, critique, and cleanup when humans supply the structure and perform the final edit.
Points of disagreement
The sharp dispute is whether unreadable AI prose comes mainly from the training corpus or from post-training optimization. The corpus camp sees models reproducing a web saturated with salesmanship, clickbait, and empty persuasion; the opposing view points to more natural base-model writing and argues that instruction tuning, engagement targets, and rewards for concise verifiable answers distort it after pretraining. બંને sides agree that optimizing a model for success metrics other than reader comprehension produces language that feels alien.
What readers liked
- Base models can reportedly match ordinary human style better than their instruction-tuned counterparts, suggesting the problem is a reversible alignment choice rather than an unavoidable property of language models.
- Models still produce strong code, giving some readers confidence that prose quality can be treated as a presentation failure rather than evidence of universally weak reasoning.
- Using an LLM to expand and clean up a human-written outline, then editing manually, preserves its drafting speed while restoring audience-aware structure and relevance.
- A full-draft-then-revision workflow may need fewer corrective prompts than trying to force clarity sentence by sentence, making AI writing assistance more practical.
Concerns raised
- Assistant prose often wins a superficial writing contest with headings, lists, and confident flourishes while failing the real test: delivering the exact information its intended reader needs.
- Technical readers find the output unusually hard to skim because meaning is diluted across flowery paragraphs, so skipping supposedly nonessential sentences can erase the point.
- Instruction tuning and reinforcement rewards may push models toward compressed, internally useful syntax that sacrifices the explicit links and grammatical precision human readers rely on.
- Models routinely miss the audience's perspective, elevating details from the latest prompt context into permanent documentation even when they are irrelevant to future readers.
- Odd verb usage, ambiguous phrasing, and mechanically scattered pseudo-main-points make AI-generated technical documentation feel like legalese without the excuse of legal necessity.
Keep reading
Sources and details
arxiv.org · Published
Why is AI generated language so hard to understand?
Read report →
Community analysis available for 1 of 1 linked discussions.
Additional discussion searches are still incomplete.
Coverage details and dates
Latest available community analysis:
Summaries: 1 current · 0 awaiting update · 0 not verified
Latest summary update:
Hacker News · arxiv.org