Why AI Text Is Exhausting to Read: Six Patterns From an NYU and UT Austin Study

If you read a lot of AI output, you have probably felt it: the answer is correct, but reading it takes effort. A study reported by The Decoder on October 26, 2026 puts numbers on that feeling. Researchers at New York University and the University of Texas at Austin say frontier language models overuse six linguistic patterns that are hard for humans to parse, and that smaller, less heavily post-trained models write more like people.

One caveat up front: we could not find the paper. The Decoder's figures credit "Clark, Martin & Mahowald (2026)", but our searches turned up no matching preprint. Everything below is The Decoder's account of the study, so treat the details as reported, not verified against the source paper.

TL;DR

The six patterns, with examples

The Decoder lists these constructions and the examples the researchers quote:

1. Unusual word combinations that rarely appear in human writing, such as "reader-patience budget."
2. Distant sentence parts, where related pieces sit far apart and tax working memory. Claude Opus 5 wrote "A cat who can always get up and away is a much calmer cat," separating subject and predicate with a long clause.
3. Shortened references with a dropped connector: "the book the professor wrote" instead of "the book that the professor wrote." It saves a word and creates a brief moment of confusion.
4. Stacked nouns, as in "horror movie soundtrack generator," which leave the reader to work out what modifies what.
5. Nouns turned into verbs: Claude Fable 5 wrote "calendar it" where a person would say "put it in the calendar."
6. Animate verbs on inanimate subjects: Claude Opus 5 produced "The whole conversation is eating its own tail."

None of these is an error. Each is understandable, which is why the study frames the problem as extra effort, a beat of processing per sentence, that adds up over a long answer.

Who comes out best?

Per the report, Llama and the Olmo base model sit closest to human text. Within the Olmo family, the deviation rises from the base model through the instruct variant to the reasoning model Olmo Think, which reaches frontier-level numbers. That ordering points at post-training, not model size or pretraining data, as the likely driver.

Claude Opus 5.5 beats Opus 5 in 19 of 21 comparisons, according to the study as relayed by The Decoder. That matters because it suggests the effect is not locked in: a lab can fine-tune against it. Anthropic's Opus 5.5 is covered across our blog, including its benchmark standing on the Epoch index and the head-to-head with Gemini 4 Argon, Grok 4.7 and GPT-6 Astra.

The explanation: from RLHF to RLVR

The researchers' argument, as summarized: human language balances ease of production against ease of understanding. Modern models are trained more and more on math, code and logic, where correctness is rewarded and clarity is not. The shift from RLHF, reinforcement learning from human feedback, to RLVR, reinforcement learning with verifiable rewards, takes humans out of the loop. The model then develops a kind of dialect that is good at compressing information and solving problems and poor for human readers.

As a proof of concept, they trained a smaller model with a penalty that rewarded shorter answers. Under certain training conditions, it favored the shorter phrasing almost without exception, regardless of context, and even generalized the preference to sentence patterns it had never seen in training. Humans, by contrast, pick short or long variants depending on the situation.

The Decoder adds a similar view from inside the industry: Anthropic's Jackson Kernion, who works on Claude fine-tuning, recently explained why Opus 4.6 was the last good "writing model" from Anthropic, saying optimization for math, code and technical explanations aimed at other AI models pushed Claude toward a dense style that reads unnatural to humans. According to The Decoder, Opus 5.5 was the first push back, addressing sentence clarity and overloaded information blocks. We have not independently verified Kernion's comments.

What this does not show

• It is not a verdict on accuracy. The study is about readability, not whether the content is right.
• Comparisons depend on the human baselines. Subtitles, StackExchange posts and student essays are very different genres from professional writing.
• A reported 19 of 21 is a count of comparisons, not a measure of how much better.
• The paper was not available to us. The details, including model versions, come through one outlet.
• Earlier readability work exists. For example, Flesch or Fumble? evaluated whether instruction-tuned models follow readability standards when simplifying text, a different question but a useful baseline for anyone building a readability check.

How to get more readable output today

These are our suggestions, not findings from the study, and you should test them on your own tasks.

1. Ask for the plain version explicitly. "Use short sentences. Keep subject and verb close. Do not stack more than two nouns."
2. Ban the specific habits. Add "no noun-to-verb coinages, no metaphors" for professional writing, since patterns three to six are easy to name.
3. Edit with a second pass. Have the model rewrite its answer for a reader with no background, then compare.
4. Try a smaller or older model for prose and a frontier model for the reasoning, then merge. The study's results suggest the writing quality gap runs the other way from the capability gap.
5. Measure it. Pick a readability score and sample your own outputs before and after a prompt change.

The cost of dense AI text is real. Our write-up of the AI verification tax shows how checking and redoing AI output eats into the time it saves, and clearer writing is one of the cheaper ways to cut it. The same goes for the wave of low-quality AI-generated reports that led Google to freeze part of its open-source bug bounty, where unreadable and unverified output has a measurable cost.

A worked rewrite (our illustration, not from the study)

Here is how the patterns look in a sentence we made up to show the mechanics. A dense version: "The agent harness failure triage dashboard calendar-syncs alerts, and a sentence like this is quietly eating its own tail." It stacks four nouns, turns "calendar" into a verb and gives a sentence a tail. A plain version: "The dashboard that triages agent failures adds each alert to your calendar." Same facts, one clause, no coinages. Reading both aloud makes the extra beat obvious, and it is the beat the researchers say accumulates across a long answer.

What people are asking

Is this just "AI slop" by another name? Not quite. Slop usually means low-effort, low-value content. This study is about well-formed, correct text whose sentence structure is costly to process. The two overlap in reader frustration but have different fixes: slop needs editorial judgment, while this needs prompts or training that favor clarity.

Should teams stop using frontier models for writing? The study does not say that. It says smaller and less post-trained models write closer to human text, and that one lab improved between versions. For customer-facing prose, run a blind comparison of two or three models on your own content and let readers choose.

Will this get fixed? If the cause is verifiable-reward training, fixes are available to labs: add human or readability signals after RLVR, or fine-tune for sentence clarity as Anthropic reportedly did with Opus 5.5. Whether competitors follow is an open question.

Why does it matter for builders? Every agent that writes to humans, whether support replies, status reports or documentation, inherits the dialect. If your product shows model prose to users, readability is a quality metric worth tracking next to accuracy and cost.

What to watch next

Two things would settle the questions. First, the paper itself: with the full methods, readers can see how the 21 Opus comparisons were defined and how human baselines were matched. Second, whether other labs treat readability as a training target the way Anthropic reportedly did. If post-training on verifiable rewards is the cause, labs that add human or readability feedback after RLVR should show the same improvement in an independent replication. We will update this post if the paper surfaces.

Related reading on explainx.ai:

• Claude Opus 5.5 tops the Epoch Capabilities Index
• Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
• The AI verification tax and how to cut it
• Google freezes OSS bug bounty over AI slop reports
• Spotify removes 50 million slop tracks
• Steam Next Fest and AI slop

Based on The Decoder's October 26, 2026 report; the underlying paper was not located and details may change when it is.