Why AI-generated text is getting harder to read

Written by

in

Table of contents

Thomas Hikaru Clark and his co-authors open their new paper with a staged sample of AI-generated text. The sentence describes itself as “eating its own reader-patience budget”, and I had to read it twice.

A sentence doesn’t eat anything, and there’s no such thing as a reader-patience budget. Still, anyone who spends the day with chatbots and coding agents has seen this kind of line before. The paper asks why AI-generated text has become harder to understand, and the answer it gives comes from training.

The sentence that eats its own budget

The team built a corpus of AI-generated text from chats and coding sessions of 14 language models. They also had the models write essays on school assignments, then counted constructions that psycholinguistics has linked to slower reading for decades.

98% of the time a model trained to save tokens picked the shorter, less human sentence.

The clearest one is a thing doing what only a living subject does. A function “swallows” an error, a puzzle pattern “bites”. Next to that the paper tracks five other markers of hard reading, from stacks of three nouns to rare word pairs.

All nine frontier and near-frontier models used more of these living-subject verbs than human writers did. They used more rare word pairs as well, and most of them stretched dependencies further in conversation.

The authors checked whether hard topics explain this by translating every sentence into Japanese and back. That keeps the meaning and swaps the wording. The difficult features dropped after the round trip, so the model had a simpler way to say the same thing and picked the harder one.

The gap was smaller in single-prompt essays and larger in long chats and coding sessions. The authors suspect that long back-and-forth conversations push models to compress harder.

Two problems in the same AI-generated text

The usual complaint about AI-generated text is bloat, and that complaint is fair. In Dossier 20 I covered a study run with the Internet Archive, which found that as AI content spreads online, texts get longer in word count while their semantic density drops.

Marcin Wilkowski walked through a slop taxonomy on his blog that was built from interviews with 19 experts. It lists information density as a feature of its own, described as long text that says little.

This paper finds the opposite problem in AI-generated text, with sentences too dense for the person reading them.

Both findings hold, and they come from different parts of training. The paper points to earlier work showing that human-feedback training rewards longer answers, while training on tasks with a checkable answer pushes toward compression. So one AI-generated text can carry padding around the edges and compressed knots in the middle.

That rules out the easy fix, because “make it shorter” is how the knots got there in the first place.

Why the model stopped writing for you

Human language has a built-in editor in the listener, because a sentence nobody understands fails and over generations forms like that get weeded out of the grammar.

Models now get a lot of training where nobody listens, since math and code tasks reward a correct answer whether or not a person can follow it. Long contexts cost money too, so every saved token counts. The paper’s hypothesis is that this pressure for information density leaks into the AI-generated text people read every day.

The authors tested it by training a small open model, Qwen2.5-1.5B-Instruct, with reinforcement learning to turn structured event data into English sentences under a penalty for length. The model found the shortcut on its own by dropping the optional “that” and picking the shorter of two grammatical options.

Without a term holding it close to the original model, it chose the short version over 98% of the time. People pick between the two forms depending on context. The habit also spread to “that” constructions the model never saw in training.

The authors call this a proof of concept, and they’re right to. A model with 1.5 billion parameters on a toy task says nothing direct about how GPT or Claude were trained, and the whole study covers English only.

In the Olmo family the effect grows with each stage of post-training, and the reasoning version ends up looking like the frontier models.

Why I care about it as a security person

The paper ends by naming user experience and monitorability as the two reasons to fix this, and the second one is my job.

Human in the loop only works if the human can read the loop. I build agents with Claude Code and I don’t know how to code at all, so what the agent tells me it changed is most of what I review. A line like “the handler swallows the error and the retry path bites on timeout” sounds confident, and I have to translate it before I can disagree with it. A tired reviewer approves it instead.

Picture AI-generated text in a SOC at 3 a.m., a triage summary written in noun stacks. It gets skimmed, and a skimmed summary is a rubber stamp on a decision nobody checked.

Claude Opus 5.5 scored lower than Opus 5 on nearly every difficulty measure in the paper, 19 of 21 comparisons. Someone rewarded clarity, and the numbers moved.

How to make AI-generated text readable again

Every post on this blog goes through a slop checklist and a detector before it goes up. The model rewrites the flagged sentences and I check them again until the count hits zero.

For AI-generated text you have to approve, one prompt does most of the work. “Rewrite this for a colleague at the end of a long day. Use people or named components as subjects and never stack three nouns.” The US government’s plain language guide gives the same advice about noun strings.

Then go through the verbs, and when a function or a config file “swallows” or “bites” something, ask what happens in plain words.

My rule for agent output is that if you can’t say in your own words what it did, it goes back to the model.

Source | https://arxiv.org/abs/2610.29162


Want More? Subscribe to The Dossier

Every week in your inbox:

📡 THE INTELLIGENCE FEED – 3-5 curated links: [Research] [Policy] [Tools] [Incidents]
💡 ONE ADVICE – One actionable AI/cybersecurity tip you can use today

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *