Why AI-Generated Text Is Hard to Understand (New Study)

6 min read · Last updated: October 25, 2026

AI-generated language is hard to understand because models are rewarded for packing in information, not for being clear. An October 2026 arXiv study found that all nine frontier and near-frontier models it tested use more low-frequency phrases and more inanimate subjects with animate verbs than human writers do.

Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.

The paper

• Why is AI-generated language so hard to understand? Clark, Martin and Mahowald, arXiv:2610.29162, submitted October 21, 2026

Key takeaways

• Researchers tested 9 frontier and near-frontier models against human writing and against round-trip-translation paraphrases.
• All nine use more inanimate subjects with animate verbs and more low-frequency n-grams than humans.
• Most also use longer dependencies, more stacked nouns and more reduced object relatives in conversation.
• The authors' explanation: models are optimised for math, code and reasoning, where density beats clarity.
• A fine-tuning experiment with a length penalty pushed a model away from human-like language.
• Models differ a lot by family, so "AI writing" is not one thing.

My take is blunt. If AI text feels harder to read than it should, you are not imagining it, and it is not a prompt problem you caused. It is a training goal problem. Models are being optimised for something other than being understood by you.

Why is AI-generated language so hard to understand?

AI-generated language is hard to understand because the models use more of the features that slow human readers down. According to the arXiv paper submitted on October 21, 2026, those features include object relative clauses, long syntactic dependencies and low-frequency n-grams, all known to correlate with human processing difficulty.

The paper starts from a feeling many of us share. In the authors' words, there is a widespread sense that AI-generated text has recently become harder to understand. Instead of arguing about taste, they measured it.

They built a new corpus of AI-generated text across diverse domains and counted a set of psycholinguistic features. These are properties of sentences that researchers already link to how hard a sentence is for a person to process.

A low-frequency n-gram is a short run of words that rarely appears together in normal writing. Your brain has seen "thank you for your help" a thousand times. It has seen far fewer sentences built from unusual word combinations, so each one costs more effort.

What did the study actually find?

The study found that AI models use difficult language features more often than round-trip-translation paraphrases, and more often than human writers. All nine models tested used more inanimate subjects with animate verbs and more low-frequency n-grams than humans. Most also used longer dependencies and more stacked nouns.

Here is the result in one place.

Feature	What the paper reports
Difficult features vs round-trip translation	AI models use them more often
Inanimate subjects with animate verbs	More than humans, in all nine models
Low-frequency n-grams	More than humans, in all nine models
Longer dependencies	More than humans, in most models
Stacked nouns	More than humans, in most models
Reduced object relatives (in conversation)	More than humans, in most models
Differences between model families	Substantial variation

Two details matter here.

First, the comparison to round-trip translation. That is text translated into another language and back, a crude way to rephrase something without a human voice. AI text came out harder than that baseline, not just harder than people.

Second, the variation. The authors report substantial differences across model families. Some models are worse offenders than others, which means you can choose.

What does an inanimate subject with an animate verb look like?

It is a sentence where a thing, not a person, does something only a person can do. My own example: "The framework embraces a philosophy of composition." A frameworks cannot embrace anything. A human would write "You build it from small pieces."

The paper counts this pattern as a hard feature, and all nine models used it more than humans. It is the kind of sentence that sounds smart and tells you little.

Stacked nouns are the same story. "Request validation error handling logic" is five nouns in a row, and you have to work out how they relate before you can read the sentence.

Why would a model write this way?

The authors propose that human language is optimised for successful communication, while AI-generated text is increasingly optimised for other goals such as math, code and reasoning. In those areas the pressure is for information density, not for being understood by a reader.

That is a hypothesis, and the paper labels it as one. But it fits something developers already know.

Think about what we reward. A model that solves a maths problem in fewer tokens is cheaper to run. A model that writes tight code is easier to check. Neither goal asks, "would a tired person understand this on the first read?"

The authors tested the idea in a small way. They fine-tuned a language model under a length penalty, which means shorter outputs scored better. The result was a divergence from human-like linguistic behaviour. The paper calls this a proof-of-concept, and I would treat it the same way: a clue, not a verdict.

The paper's conclusion points at what it calls the axes of misalignment between the processing demands of humans and of language models. It also points toward possible remedies, though the abstract does not list them.

What is the strongest counter-argument?

The strongest counter-argument is that dense does not mean bad. If a model packs more meaning into fewer words, maybe the reader just needs to adjust. Experts read dense text all the time. Perhaps this is a feature, not a flaw.

I take that seriously, and it is partly right. Density is useful when you are the target reader and you want a compressed answer.

But the paper's own framing answers it. The features measured are ones that correlate with human processing difficulty. That means readers spend more effort per sentence. For a document you will skim, share with a teammate or paste into a README, that effort is a cost, and it lands on the person least able to fix it.

There is also a fair limit on what I can claim. I read the arXiv abstract and listing page, not the full paper. The abstract describes features known to correlate with difficulty. It does not describe reader comprehension tests, so I will not say it proved readers understood less. What it shows is that the text leans on harder features than human writing does.

What should you do about it?

Treat AI output as a first draft written for a machine's goals, then edit it for a human reader. Ask for plain sentences, check for stacked nouns and odd phrases, and never paste raw model text into documentation your team has to read.

Here is the routine I would use:

1. Ask for the reader, not the topic. Tell the model who will read it and what they already know.
2. Ask for short sentences and plain words. This pushes against the density pressure.
3. Hunt for stacked nouns. If three or more nouns sit in a row, rewrite the phrase with a verb.
4. Fix non-human subjects. If a thing is doing a human action, name the person or rewrite the action.
5. Read it aloud once. If you stumble, your reader will too.
6. Compare models. The paper found substantial variation across families, so test two or three on your own writing.

The same discipline applies to code. If you want a method for checking machine output, see how to review AI-generated code like a pro. If you use AI to study, how to use AI to learn programming 3x faster is worth a read, because hard-to-parse explanations slow learning down.

For the bigger picture on where these tools fall short, I wrote about why we're being sold an unfinished product. And if you worry about the job side, here is the honest answer on whether AI will replace developers.

FAQ

Is AI-generated text really harder to read than human writing?

According to the October 2026 arXiv study, yes on the features it measured. All nine models tested used more inanimate subjects with animate verbs and more low-frequency n-grams than humans. Most also used longer dependencies and more stacked nouns.

Why is AI writing so hard to understand?

The authors' hypothesis is that human language is optimised for communication, while AI text is increasingly optimised for math, code and reasoning. Those goals reward information density. A fine-tuning test with a length penalty moved a model away from human-like language.

Do all AI models write this way?

No. The paper reports substantial variation across model families. The trend holds across all nine models for some features and across most for others, but the size of the effect differs by model.

Can this be fixed?

The authors say their findings point toward possible remedies, but the abstract does not name them. In practice you can reduce the problem by telling the model who the reader is, asking for short, plain sentences, and editing the output yourself.

AI text is written to be dense. Your job is to make it readable. Try the free tools on Art of Code for your daily workflow, and read why coding fundamentals still matter in 2026 next.

Written by Adithya Guttha, Founder of Art of Code.