Six reasons AI text is such hard work to read

On 21 October 2026 Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald published "Why is AI-generated language so hard to understand?" on arXiv. Rather than describe the sense that model output is getting harder to read, they measured six specific syntactic constructions that psycholinguistics knows burden the reader. All nine frontier models studied use them more often than humans do.

Key takeaways

• Corpus: 120 conversations, 45 coding sessions and 980 essays from 14 models, set against human text from film subtitles, StackExchange and student essays
• Features measured: long syntactic dependencies, low-frequency n-grams, reduced relative clauses, stacked nouns, animate verbs on inanimate subjects
• Llama and the Olmo base version sit closest to human text, and within the Olmo family the score rises from base through instruct to the reasoning model
• Claude Opus 5.5 beats its predecessor Opus 5 in 19 of 21 comparisons
• Proof of concept: a model fine-tuned with a length penalty began shortening regardless of context and generalised that to patterns absent from training

Where this starts

A model's answer is correct. The facts check out, nothing is missing. And yet reading it takes effort — one has to return to the start of the sentence to work out what connects to what.

That experience is not a matter of taste. Specific grammatical constructions carry the load, and their effect on a reader's working memory was described by psycholinguistics long before language models existed.

But here is the problem: if every individual sentence is correct, what is there to measure?

That is precisely the question this paper answers.

What exactly was measured?

The team assembled a corpus of text generated in three different situations: casual conversation, a coding session and an argumentative essay. In total 120 conversations, 45 coding sessions and 980 essays from 14 models — among them Claude, GPT-5.6, GLM-5, Kimi K2.5 and smaller open models such as Llama and Olmo.

The reference point was human text from three sources matched to those situations: film subtitles, StackExchange and student essays.

One additional control matters more than it looks. The authors compared AI text not only with human writing but also with paraphrases obtained by round-trip translation. The models used difficult constructions more often than their own paraphrases too — meaning this is not an artefact of topic or length.

Which constructions are they?

The common denominator here is not error but cost. Each construction can be understood — it just takes an extra beat. Six such beats per paragraph produce an impression that is hard to name and easy to feel.

What do these names mean?

This paper's terminology comes from psycholinguistics and is opaque without translation.

• Syntactic dependency — the distance between words that must be linked for a sentence to make sense. The longer it is, the longer the reader holds the sentence's opening in memory.
• n-gram — a run of several consecutive words. A low-frequency n-gram is a combination the reader has not seen before, so no ready pattern exists for it.
• Object relative clause — a construction of the form "the book the professor wrote". Harder to process than the subject variant.
• Information density — how much content sits in each word. High density is an asset when transmitting data and a liability when reading.
• Post-training — the stage after base training in which a model learns to answer in a particular way. Identified here as the most likely cause.
• RLVR — reinforcement learning in which the reward is granted by an automatic verifier rather than a person.

Together these terms let the paper's thesis be stated: the problem is not aesthetic but measurable — and it lies in what training fails to reward.

Why not simply say models write "unnaturally"?

Because "unnaturally" does not say which model and where. The paper shows the differences between model families are large, and that settles the question of cause.

The Olmo family

Closest to human text are Llama and the Olmo base model. Within the Olmo family the score rises consistently: from base, through the instruct variant, to the reasoning model Olmo Think, which reaches frontier levels. If the degradation accumulates with each tuning stage, post-training is the most likely source.

Opus 5.5 as a counterexample

The second piece of evidence runs the other way, which makes it stronger. Claude Opus 5.5 beats its predecessor Claude Opus 5 in 19 of 21 comparisons. That means targeted tuning can reverse the effect — it is not an inevitable consequence of scale.

There is a second problem that diagnosis alone does not solve: not every text should be short. Humans pick between a shorter and a longer variant depending on the situation. A model rewarded for brevity loses that flexibility.

Where does it come from, per the authors?

The hypothesis is stated plainly. Human language is optimised for successful communication — balancing ease of production against ease of understanding. Model-generated text is increasingly optimised for other goals: mathematics, code and reasoning, where correctness and information density are rewarded and comprehensibility is not.

The authors point to a specific transition: from RLHF (Reinforcement Learning from Human Feedback), where a human granted the reward, to RLVR, where the human leaves the feedback loop entirely. The model then develops a kind of dialect — good at compressing information and solving problems, bad for a human reader.

This describes a proof-of-concept experiment, not a measurement run on frontier models.

A similar explanation came from inside the industry. Jackson Kernion, an Anthropic employee working on Claude fine-tuning, recently explained why Claude Opus 4.6 was the company's last good "writing model". Optimising later models for mathematics, code and technical explanations aimed at other AI models pushed Claude toward a style that works for AI and reads as too dense for people. Kernion called it an adaptation to "LLM psychology".

Is it live yet?

This is a published research paper, not a tool or a fix. The preprint appeared on arXiv on 21 October 2026 and has not yet been peer-reviewed.

One practical consequence already exists. With Opus 5.5 Anthropic deliberately pushed back against the trend for the first time, addressing sentence clarity and overloaded blocks of information — and this paper's measurements bear that out.

Why does it matter?

The paper turns a vague complaint into a measurable parameter, and in doing so makes it an optimisation target. Until now a quality that was hard to argue about except through examples had no measure of its own. Now it has one — and, as a single model family shows, it can be improved without losing technical capability.

What comes next?

• The authors point in the abstract toward "possible remedies" — the next step is testing whether psycholinguistic features can enter post-training as a reward signal.
• The Claude Opus 5.5 result (better than Opus 5 in 19 of 21 comparisons) is a reference point against which later model releases can be judged on readability.
• The Olmo family, with an open base, instruct and reasoning version, remains the only public case in which it is possible to reconstruct at which tuning stage readability drops.

Sources

• arXiv — Why is AI-generated language so hard to understand?
• The Decoder — Researchers find six patterns that make AI text so exhausting to read