Why Is AI-Generated Language So Hard to Understand? What the New Paper Found

If AI-generated text has felt harder to follow lately, a new paper says you are not imagining it. In "Why is AI-generated language so hard to understand?" (arXiv 2610.29162, submitted October 21, 2026), Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald measure the effect and propose a cause. The discussion on Hacker News passed 70 points on Saturday, with readers trading examples of the same frustration.

This post explains what was measured, what was found, the authors' theory about why, what the experiment does and does not show, and what you can do about it today.

TL;DR: the questions people are asking

What the authors measured

The paper starts from a "widespread sense that AI-generated text has recently become harder to understand." Instead of relying on impressions, it counts features that psycholinguistics has linked to human processing difficulty:

• Object relative clauses, especially reduced ones without "that".
• Long syntactic dependencies, where the words that belong together are far apart.
• Low-frequency n-grams, meaning word sequences that rarely appear in ordinary text.
• Stacked nouns, such as a chain of nouns piled up before the head noun.
• Inanimate subjects with animate verbs, such as "the framework wants to".
• Denominal verbs, nouns turned into verbs.

The corpus covers three domains. Per the paper's HTML version, there are 120 chat conversations (1,428 replies, human baseline from OpenSubtitles), 45 coding sessions across five repositories (human baseline from StackExchange), and 980 essays from 14 models on seven assignments (baseline: 2,934 student essays scored 5 to 6). The frontier-model runs cost $211.78 in API fees.

The paraphrase control matters. A round-trip translation (English to Japanese to English with NLLB-200) rewrites text without an intent to be dense. If AI text is harder than the paraphrase, the difficulty is not just a side effect of generic rewriting. Robustness checks used a Russian pivot and a different translation model.

What they found

Frontier models use the difficult features more often than round-trip paraphrases. Against human writing, the abstract states that "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.

Details from the paper that are easy to miss:

• Domain matters. Effects are larger in chat and coding than in essays. The authors attribute this to essays being a single prompt rather than a multi-turn interaction.
• Models differ. The paper reports substantial variation across model families. Llama and Olmo base and Instruct models change least under round-trip paraphrase, while the reasoning variant Olmo Think looks more like the frontier models.
• Newer is not always worse. The authors report that Claude Opus 5.5 scores lower than Opus 5 on 19 of 21 comparisons, which they read as a sign that the problem can be reduced.
• Effect sizes live in figures. The text we read gives directions per feature but most exact sizes appear only in charts, so we do not quote numbers we cannot check.

If you track Anthropic's models, our Opus 5.5 community showcase covers what builders did with the model the paper singles out as improving.

Why? The information-density hypothesis

The central claim is a hypothesis, flagged as such. In the authors' words, "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 information density rather than successful communication."

Think about how reinforcement learning on verifiable tasks works. A reward for a correct answer, combined with any cost on length, favors compact output. Dropping "that" from a relative clause saves a token. Stacking nouns saves function words. Using an abstract noun as the subject packs a relation into fewer words. A human reader needs those small words as parsing aids; a model deriving the answer does not.

To test the idea, the authors fine-tuned Qwen2.5-1.5B-Instruct with GRPO on turning JSON event descriptions into sentences, adding a length penalty. Without a KL penalty, any positive length penalty pushed all seeds to more than 98% choice of the short variant for the trained "that" alternations. With a KL penalty the preference rose more gradually. The shift also transferred to untrained "that" alternations, and for the genitive alternation in the no-KL case.

What the experiment does not show

The authors are clear about the limits, and readers should be too:

1. Proof of concept, not a demonstration. A 1.5B model on a narrow task shows the mechanism is plausible. It does not show that frontier labs' training pipelines cause the measured features.
2. No causal access. Frontier models cannot be studied during post-training at scale, so attribution is indirect.
3. English only, subset of models. Chat covers 10 topics and coding five repositories.
4. Features are proxies. A higher count of hard constructions is not the same as a reader failing to understand. The paper measures processing difficulty correlates, not comprehension scores from human participants in the text we read.
5. Other explanations remain. RLHF-style rater preferences, training data mix and system prompts could all contribute.

What developers on Hacker News are saying

The thread has 54 comments at the time of writing, and the reactions split.

• User skeledrew pushed back, saying they have no trouble understanding AI text and asking why some people do.
• User paid_dot_expert reports that a follow-up "ELI5" prompt usually fixes output that is full of buzzwords, and calls it a failure mode worth snapping the model out of.
• User cmrdporcupine says models fail to intuit the audience and reward their own recent context, and that writing a messy outline yourself works best.
• User dividefuel, a technical writer, guesses that years of RLHF optimized for raters' first impressions: lists, headings and catchy phrases.
• User amelius notes that models also write for other models in feedback loops, so optimizing only for results may make them "speak AI".

These are commenters' impressions, not findings. They line up with the paper's hypothesis but do not test it.

How to get clearer output today

The paper's remedy is at training time: reward effective communication rather than correctness alone, or align outputs with human comprehension needs, citing work on preference optimization based on uniform information density. You cannot retrain a frontier model, but you can steer it:

1. Name the reader. "Explain to a first-year analyst who has not seen this codebase" changes the target of the optimization.
2. Ask for plain structure. Request short sentences, explicit "that", and one idea per sentence.
3. Use a second pass. Have the model rewrite its own answer for readability. The same trick helps with verification tax problems, where redoing AI output eats the time savings.
4. Constrain vocabulary. A controlled language can force simplicity; see our guide to ASD-STE100 Simplified Technical English with AI.
5. Edit before you ship. If readers will see the text, assume the first draft has the very features this paper measures.

For teams that publish AI-assisted content, the readability angle also connects to how audiences and platforms are reacting to low-effort output; see how Gamma 5 tackled the "AI smell" in presentations and why Steam's Next Fest filled up with AI slop.

Why this matters beyond style

If models are trained to be dense, and dense output is harder for people, then comprehension cost is a hidden externality. It shows up as longer review time for generated code explanations, more clarifying prompts, and readers giving up on AI-written documentation. It also complicates monitoring: if humans must read model reasoning to catch errors, text that is hard to process weakens oversight.

The finding also suggests an evaluation gap. Benchmarks reward correct answers; few reward answers a person can absorb quickly. A readability metric built from the paper's features could be added to evals cheaply, since the frontier runs cost about $212. Whether labs will optimize for it is the open question.

What to watch next

• Whether the authors release their corpus and feature code, so others can replicate.
• Whether the Opus 5.5 improvement holds up in independent tests and in the OpenAI models, where the HN thread argued differences were smaller.
• Human-subject studies measuring actual comprehension and reading time on these outputs.
• Whether any lab adds readability to its reward models.

Related reading

• Claude Opus 5.5 community showcase
• The AI verification tax and how to cut it
• Using ASD-STE100 Simplified Technical English with AI
• Gamma 5 and the AI smell in presentations
• Steam Next Fest and AI slop
• Spotify removing 50 million slop tracks

Details reflect the paper's first version (v1) and discussion available on October 24, 2026.