AI-generated language is drifting away from plain human readability.
And nine models from the frontier and just under it all drift in the same direction.
They write text loaded with the exact features that make a human brain work harder to parse a sentence.
The study proving it does not hide its question.
The title is, without embarrassment, “Why is AI-generated language so hard to understand?” (arXiv).
The identifier alone, 2610.29162, dates the posting to October 2026, well over a year after the two companion papers it sits beside. The direct answer: yes, the drift is real and measurable. Across a corpus of AI-generated text spanning diverse domains, all nine models produced more inanimate subjects with animate verbs and more low-frequency n-grams than human writing did. Most also stretched dependencies longer, stacked nouns higher, and used more reduced object relatives in conversation. The proposed cause deserves a slow read, and it comes later in this piece.
Why AI-generated language is hard to understand
The researchers started from a hunch and turned it into counts. “There is a widespread sense that AI-generated text has recently become harder to understand,” they write. A sense is not a dataset. So they scored model output against psycholinguistic features already known to correlate with human processing difficulty: object relative clauses, long syntactic dependencies, and low-frequency n-grams.
The baseline is the quietly clever part. Model output was compared against paraphrases generated by round-trip translation, which pushes text through another language and drags it back with no intelligence choosing any of the new words. Judging against round-trip paraphrases is a bit like reheating a dish twice, then asking which plate tastes more homemade. Even measured against those clumsy rewrites, the models still reached for the difficult features more often.
Against human writing directly, the habits held in every one of the nine.
All nine produced more inanimate subjects with animate verbs, sentences where a thing performs an action only a person can perform, plus more low-frequency n-grams, the rare word pairings your eye snags on. Most of them lengthened dependencies, stacked nouns higher, and reached for more reduced object relatives in conversation. The drift is not uniform either, because the study found substantial variation across model families.
That detail matters if you are shopping for a house style.
One honest limit from my side: the paper reports family-level variation, not a ranking of which of the nine drifted worst. And I am not going to invent a ranking it does not contain.
The objective, not a bug
Here is the punchline of the whole paper, the mechanism the authors propose: “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.”
Then they tested the theory.
They fine-tuned a pre-trained language model under a length penalty and measured the linguistic profile of its outputs.
Squeezed for tokens, the model learned to succeed at its task using fewer tokens by exploiting syntactic alternations, the two routes grammar offers for saying one thing. It kept doing that even when the output turned less humanlike.
My read is the contrarian part: no model in that experiment misbehaved.
Every incentive in a modern stack points the same direction. Per-token pricing punishes length. Context windows punish length. Agent frameworks punish length. A model under that pressure behaves like a mover paid by the truckload instead of by the hour: everything gets packed tight. And whether a human can ever unpack it is somebody else’s problem. Punish length and the shortest grammatical option wins. The shortest option is the compressed, dense, less human one. Opacity: the objective working exactly as specified.
What AI-generated language means for a small operation
Run a lean shop that pastes model output straight into customer-facing surfaces? Then today’s risk is not hallucination.
It is text that is technically correct and quietly exhausting, a support reply that is accurate and still loses the reader inside a dependency chain.
Nobody files a bug report about prose.
Have you ever filed one?
Neither has anyone else.
They just stop reading.
Split your output into two lanes and hold the line between them:
– The internal lane holds dense spec summaries, code comments, and technical notes. Density is a feature here. Let it run.
– The customer lane holds emails, docs, and onboarding, anything with your name attached. Every word gets a human read before it ships. Treat that read as part of the pipeline, not polish smeared on afterward.
The pass itself is unglamorous, and it runs like this:
1. Read the draft out loud exactly once.
2. Stacked nouns announce themselves the moment you have to breathe through them.
3. Send the offenders back with one instruction: same meaning, human sentences.
4. If one model family fights you every single time, test a different family before deciding the problem is you. The study found substantial variation across families, so that is measured advice, not a hunch.
Sometime last spring I caught myself skimming a support reply our own pipeline had drafted.
And that is when the out-loud pass became a rule here rather than a suggestion.
Humans pull the language back
A separate line of research compared languages optimized for LLMs against languages optimized for humans and found the two subtly differ (arXiv, a March 2025 posting). The encouraging result came from what happened next: interactions between humans and LLMs alleviated those differences and produced vocabularies more human-like than LLM-like. The researchers frame this as maintaining referentially grounded, human-like vocabularies, the academic phrasing for the human keeps the words honest.
A third paper, from February 2025, is the warning shot. Emergent communication research watches AI agents develop their own protocols to achieve shared goals. And those protocols remain largely opaque and difficult to translate into human-readable forms (arXiv). One team tested unsupervised neural machine translation as a way to convert those emergent languages into readable text. And the inter-category setting scored better on both BLEU and METEOR. Read that plainly: we are already machine-translating machine languages back into ours, then grading ourselves on how well the round trip survives.
Line the three papers up and one trend shows.
Model prose drifts dense.
Agent-to-agent protocols drift opaque. The single force dragging language back toward human is a human being in the conversation. Collaboration is not a nicety bolted onto the pipeline, it is the alignment mechanism. And separate research groups keep confirming it from separate angles.
FAQ: AI-generated language, answered
Why is AI-generated text harder to read?
The features that make text hard are the ones models now use most: inanimate subjects performing animate actions, low-frequency n-grams, longer dependencies, stacked nouns. And reduced object relatives in conversation. All nine models in the study showed the first two. Most showed the rest.
Which models drift most?
The study found substantial variation across model families but crowned no single worst offender. Treat any specific ranking you read elsewhere as somebody’s guess, including mine.
How do I fix dense AI writing?
Read it out loud once, flag every sentence you breathe twice through. And send those back with one instruction: same meaning, human sentences.
A human pass before shipping has beaten every prompt tweak I have tried.
Is this the same problem as hallucination?
No. A reply can be factually correct and still lose the reader inside a dependency chain. The study measures processing difficulty, not truth.
Does the drift matter for internal docs?
Less. Spec summaries, code comments, and technical notes can stay dense, since density is a feature there. Customer-facing text is where the human read is non-negotiable.
Where do I start this week?
Take the next AI draft headed for a customer, read it aloud. And flag every sentence that makes you breathe twice. The next benchmark chart will not save you from a support email nobody finishes. So judge your models on task scores and judge shipped text on whether a tired person understands it in one pass. Those are other tests, and only the first one has a leaderboard.
The read-aloud pass remains the cheapest alignment tool you own.
