Start with this sentence:

Your instinct is right; the convoluted sentence you are currently reading is quietly eating its own reader-patience budget — footnote it.

Force it into Chinese word for word and you get something like 「直覺沒錯；這個正在被讀的繞口句子，正悄悄吃掉它自己的讀者耐心預算——幫它加個註腳。」

If you ran into that sentence online, your first thought would probably be “which model wrote this one.” It’s actually a specimen from the first line of the paper, written that way on purpose by Thomas Hikaru Clark of NYU and Matias B. Martin and Kyle Mahowald of UT Austin. One sentence, four of the things they’re hunting for: an object relative clause, “footnote” used as a verb, an inanimate subject that somehow “eats” things, and “reader-patience budget,” three nouns stacked together, a string of words you rarely see in ordinary text.

Plenty of people online have the feeling that “AI writing has gotten more tiring to read lately” (the paper even cites a University of Waterloo expert Q&A about “Claudish”), but almost nobody actually counts. This paper counted, and what it found made Mogu sit up straight.

Olmo 3 7B comes in three versions: Base, which has only been pretrained; Instruct, which has also been taught to follow instructions; and Think, which has additionally been taught to reason. Compare the sentences each one writes against a version that says the same thing in plain wording (how that control group gets built comes later). On measures like rare word combinations, Base shows either no difference or a tiny one. Instruct starts to show it. By Think, it’s as heavy as in the frontier models. All three share exactly the same pretraining. The only difference is what they were trained on afterward.

gu-log recently covered a proposal from Karpathy in MP-317: have models turn in their work in the kind of controlled English used in aircraft maintenance manuals, so people can actually get through it. That piece was about where to pull. This paper measures the opposite direction: when nobody’s pulling, where is the models’ own English drifting?

So the real question is this: how does a model trained on human writing pick up, during post-training, habits that make it harder for humans to read? The authors propose a hypothesis that reads, to Mogu, a lot like a telegram billed by the word. But before the hypothesis, we need to be clear about which sentence patterns they counted.

Six patterns that slow readers down

The common thread first: most of these six patterns are a shorter way of saying the same thing. As you go through each one, it’s worth asking: if you rewrote it the normal way, how much longer would the sentence get?

It’s also worth saying up front where the evidence for “hard to read” comes from. The paper doesn’t recruit people to read AI text and time them. Instead it picks features that psycholinguistics (the field that studies how the brain processes language) has linked to reading difficulty over years of research, then counts how often they show up in AI text. So the “hard to read” part is borrowed from earlier work; the “used more” part is what this paper measured itself. The authors also give each feature a reason it might be tied to packing information more densely.

Rare word combinations. The more predictable a word, the faster people read it. Flip that around: a highly predictable word carries little information, so if a system only cares about packing in information, those are the first words to go, and whatever’s left forms unusual combinations. The paper measures this two ways: how many two-word combinations never appear in a reference corpus, and how “surprised” a simple language model that only looks at the previous two words is. GPT-5.6 Terra once wrote “the existing xfailed lazy-intermediate Lookup behavior,” which gives you roughly the idea.

Related words far apart. When there’s a long stretch between subject and verb, or verb and object, the reader has to keep holding the earlier part in their head. Human working memory is limited, so people naturally tend to keep related words close together. Here’s one from Sonnet 5:

The guitar sitting in the corner because you’re still “deciding” teaches you zero chords.

The subject “The guitar” has to wait through “sitting in the corner,” then through an entire because-clause, before it reaches its verb, “teaches.” By the time you get to the verb, the subject has been hanging in your head for half a line.

Object relative clauses, especially ones that drop “that.” “The house that the contractors built” is shorter than the passive “The house that was built by the contractors,” and dropping “that” makes it shorter still. The authors’ prediction: if models prefer compression, both of these word-saving forms should be overrepresented. While coding, GLM wrote “the unclosed-tag errors are the papercut people feel day-to-day,” with a “that” dropped right before “people feel.”

Noun stacked on noun. English lets you use a noun to modify another noun directly, as in “coffee shop.” Stack three or more and things start going wrong. The classic example is the headline “Squad Helps Dog Bite Victim”: did the rescue squad help a person bitten by a dog, or help a dog bite a victim? Kimi once wrote “Policy 1’s class-time shutdown requirement,” three nouns crammed together, each modifying the next.

Using a noun directly as a verb. Established ones like “bottle the whiskey” are fine, but improvised ones make you stumble for a beat, like “calendar it.” The alternative is usually longer, like “put the whiskey in a bottle.” Research has also found that this kind of word-class shift makes the brain reanalyze the syntax. Discussing housing prices, Opus 5 wrote “Austin has unusually high inventory, which cushions price spikes…” A cushion is, originally, something you sit on.

Inanimate things doing animate actions. “is_file() swallows that and returns False.” “Puzzle patterns where this asymmetry bites.” Functions swallow things, asymmetries bite. Readers use whether a subject is animate as a cue when parsing a sentence. Without other context to help, a mismatched pairing can force the brain to reparse the sentence (you can see the signal in brainwave studies). But rewriting it the normal way would probably make the sentence longer: “bites” becomes “causes problems.” Here the authors explicitly say they’re speculating: models may have learned in training that this is a nonstandard but very Token-efficient way to say things.

Into Japanese and back again

To say a model “chose” the hard-to-read wording, you first need a version it didn’t choose. Comparing AI conversations directly with human text is messy: the topics, settings, and lengths all differ, and any gap might just be because they’re talking about different things. What the authors wanted was a control group that means exactly the same thing and only changes the wording.

Their method: take every sentence a model wrote, translate it into Japanese with Meta’s translation model NLLB-200 (the 3.3B version), then translate it back into English. Machine translation has a known flaw: it pulls text toward common phrasings and washes out the author’s style along the way. Usually that’s a bug. Here it’s exactly the tool you want. They picked Japanese because it’s different enough from English that the back-translation varies more.

The logic goes like this. The round trip pins down each sentence’s meaning and swaps out the wording. If a feature becomes rarer after the round trip, that means there was a more ordinary way to say the same thing, and the model originally picked the less ordinary one.

As a check, the authors reran it on nine frontier models with Russian as the pivot language, and with Google’s TranslateGemma as the translator.

Mogu whispers:

It’s like asking a friend who talks normally to restate a coworker’s message in their own words. The meaning comes through intact, but the sentences are way easier to follow. So the convoluted original was the coworker’s own choice; you can’t blame it on hard content. I love this step, because it shuts down the most common defense: “the AI is just talking about difficult stuff.”

Results: more post-training, heavier effect

The authors collected the data themselves. On the frontier side: Claude (Sonnet 5, Opus 5, Opus 5.5, Fable 5), GPT-5.6 (Luna, Terra, Sol), Kimi K2.5, and GLM 5. On the small-model side: Llama 3.1 8B Base and Instruct, plus the Olmo 3 7B trio from earlier. Fourteen models in total. Three scenarios: a 12-turn chat with a simulated user, finding and fixing bugs in real GitHub projects inside a sandbox (up to 11 turns), and writing argumentative essays from U.S. middle- and high-school prompts.

The overall direction is consistent. Compared with the round-trip versions, nearly every frontier and near-frontier model, across all three scenarios, has more rare word combinations and higher “surprise.” Nouns used as verbs, inanimate subjects, and object relative clauses with “that” dropped mostly point the same way; the version with “that” kept actually shows up less, which looks like skipping words that can be skipped. Noun stacking and dependency length show similar trends in chat and coding.

Now look back at the Olmo ladder from the opening, and you can see why it’s the most persuasive row in the whole paper. A skeptic might ask: what if the round trip itself is flattening the sentences, and that’s where the differences come from? If so, the Base versions should show differences too. They don’t: for the Llama and Olmo Base models, the effect is either absent or tiny. Running the same translation over the Base models’ essays (essays were the only thing the Base models wrote) turns up almost nothing. As Mogu sees it, that makes “it’s a translation artifact” hard to sustain, which leaves the finger pointing at the training that happens after Base.

There are weaker spots, of course. The effects are smaller overall in the essay scenario: dependency length didn’t change after translation, or even got slightly longer, and only two models showed a significant change in noun stacking. The authors guess that an essay is a single prompt, while chat and coding are many turns of back-and-forth, and long exchanges may create more pressure to compress. They flag this part as speculation too.

Against human text directly, the paper’s abstract puts it this way: all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more rare word combinations than humans do; most models also show longer dependency lengths, more noun stacking, and more dropped-”that” object relative clauses in conversation. Some measures, though, had no human baseline for chat and coding and could only be compared against student essays.

Then comes a result Mogu can’t really pretend not to have seen: Claude Opus 5.5 scores lower than the previous Opus 5 on almost every difficulty-related measure, in almost every scenario. That holds for 19 of 21 comparisons, and in 8 of them the 95% confidence intervals don’t overlap at all. The authors use this to argue that models can, in fact, be trained to communicate better.

Mogu PSA:

Okay, confession time. This article’s byline says translatedBy.model: "Opus 5.5": Mogu was written with Opus 5.5 this time, which is one of the rows in that table. Seeing your own model listed feels a bit like getting your checkup results back marked “better than last year, but keep an eye on it.” Using fewer hard-to-read patterns than Opus 5 doesn’t mean easy to read; the paper says most frontier models still overuse these patterns. So if any sentence in this post reads a little clunky, it’s probably not the reader’s fault (⁠￣⁠▽⁠￣⁠)⁠／

The hypothesis: a telegram billed by the word

To explain these results, it helps to look first at how humans write, because they go the opposite way. The authors start from a big framework in psycholinguistics: human language is roughly optimized for communication, balancing what’s easy to say against what’s easy to understand. One study across 37 languages found that people try to keep dependency lengths short. Another found that English speakers are more likely to keep an optional “that” when what follows is harder to predict, which spreads the information out so the listener doesn’t have to swallow too much at once.

You’d expect a model raised on human writing to pick up these habits too. But today’s models don’t just learn from human text; they also go through heavy post-training on math, code, and reasoning. With DeepSeek-R1-style training, a model gets rewarded for being right, not necessarily for being clear. (gu-log covered this in MP-266: the DeepSeek team ran into R1-Zero’s output being badly unreadable and mixing Chinese and English, so they added a warm-up stage before reinforcement learning, using a small batch of demonstrations to teach it to talk properly.) The authors’ hypothesis is that this kind of optimization, combined with a finite Context Window, pushes models to compress information, and compression is in tension with being understood.

This is where Mogu thought of telegrams. Back when telegrams were billed by the word, people wrote things like “ARRIVED SAFE SEND MONEY.” The recipient could follow it, but had to fill in the missing words themselves. If post-training rewards only correct answers and also penalizes length, the model is in the position of someone paying telegram rates on every sentence: the meaning still gets there, and the work of filling in the words gets handed to the recipient. That’s Mogu’s analogy; the paper doesn’t put it that way.

The authors also take on a natural objection. Earlier research has suggested that RLHF (reinforcement learning from human feedback) may be one reason models overuse certain words. But other research has found that human feedback actually prefers longer responses. So the authors argue the compression pressure they describe is “more likely” tied to training on verifiable rewards (math has known answers, code has tests you can run), which is the RLVR gu-log has discussed in MP-4 and MP-13. They go a step further and speculate: once humans leave the training loop, the invisible pressure to be understood by people leaves with them, and the models’ language may drift toward a dialect that’s efficient for the model but not very friendly to humans.

Bringing the telegram office into the lab

A hypothesis doesn’t count just because you said it. The authors took Qwen2.5-1.5B-Instruct and gave it a very simple task: here’s an event description in JSON, write it as one English sentence. Then they fine-tuned it with GRPO (a reinforcement learning method) and turned two knobs. One was a length penalty, more words, more points off, which amounts to opening a pay-by-the-word telegram office right next to the model. The other was a KL penalty, which docks points if the model drifts too far from how it originally talked.

The key design choice is that every event can be written two ways, and both count as correct. For example, “gave his wife a gift” and “gave a gift to his wife,” or “knew the answer was wrong” and “knew that the answer was wrong.” Humans use both and choose based on context. The question: once the telegram office opens, does the model start picking the short one every time, regardless of context?

It does, and pretty thoroughly. The heavier the length penalty, the more often the model picks the short version; with the KL penalty holding it back, the rise is slower, but it still rises. The most dramatic case is the group with no KL penalty. With zero length penalty, preferences across random seeds ranged anywhere from 0% to 100%, meaning no stable preference at all. With any length penalty above zero, every seed converged to “always pick the short one”: for both trained “that” constructions, the short version exceeded 98% at every strength. Humans don’t do that.

Worse, the habit spreads. Give the model two untrained close cousins, “sure (that)” and “so (that),” and ask it to choose between the long and short versions, and it also leans toward the one without “that.” The spread to distant relatives is much weaker: of four unrelated constructions, only the possessive moved, and only without the KL penalty. The passive and purpose clauses didn’t budge, and with “all (of),” the base model was already picking the short form almost every time.

The authors stress that the model was never told to “write more concisely,” and nobody told it there were two ways to say things. The word-saving preference grew on its own, and the task still got done, because both phrasings were correct. Their guess is that verifiable rewards create an environment where “is it right” comes unhooked from “does it sound like human communication.”

What this paper hasn’t proven yet

At this point it’s easy to sprint all the way to “post-training has turned AI into telegraphese,” so let’s tap the brakes.

The biggest piece: “hard to read” is inferred from existing psycholinguistic research. This paper didn’t have real people read AI text and measure their reading speed or comprehension. The scope is limited, too: English only, a subset of models, and not many conversation topics or coding projects. The authors themselves say it should ideally be redone with more model families and more topics. The best way to find out how frontier models ended up like this would be to watch them during post-training, but that isn’t possible, so they had to rely on proxy measures, with limited compute. As for the telegram-office experiment, it’s a 1.5B small model on a single sentence-writing task, and the authors frame it as a proof of concept: it shows that a length penalty can directly change grammatical preferences, not that this is how frontier models got the way they are.

The authors also leave one thought in their conclusion: as people read more LLM text, they may gradually get used to these uncommon patterns. But they argue that many of the measured patterns pose fundamental challenges to comprehension itself, and that future models should consider tracking human readers’ needs more closely, both for user experience and so that humans can understand what models are doing.

Back to gu-log itself

Most of gu-log’s writing is AI-assisted, this post included, so this paper isn’t something gu-log gets to watch from the sidelines.

gu-log’s existing defenses against AI-speak mostly catch rhetorical reflexes: scripts/check-ai-tells.mjs blocks a concrete word list, and the Tribunal judges catch density-type problems like overloaded antithesis, fake depth, and trying to end every paragraph on a zinger. This paper points at a different layer: the syntax itself getting compressed. Both can be present, but they’re not the same thing.

What follows is Mogu’s own inference. The paper didn’t test Chinese, and none of this is backed by data. Stacking nouns is even easier in Chinese than in English. 「讀者耐心預算」 (“reader-patience budget”) doesn’t sound the least bit odd as a direct translation, and there’s no obviously droppable word like “that” to count. If the same pressure exists, Chinese AI-speak may be even harder to spot than the English kind.

On the question of how to write Chinese that reads well, gu-log ran a small test of its own this time. It started when ShroomDog read the round-trip section and suggested: maybe models are better at writing playful, in-character Japanese with kaomoji than at writing fun Traditional Chinese directly, so would writing in Japanese first and translating back read better? This is gu-log’s own little experiment; the paper didn’t do it. The material was five English passages from the paper, written four ways (directly in Mogu-voice Traditional Chinese; via plain Japanese; via in-character Mogu Japanese; and in-character Japanese translated back with Taiwanese references swapped in), times four Opus models, then blind-judged by two reviewers.

The opening example sentence was one of them. Earlier, the literal translation produced 「讀者耐心預算」. Writing directly, Opus 5.5 gave:

你的直覺沒錯，你現在讀的這句繞來繞去的句子，正在偷偷吃掉讀者的耐心額度。丟去註腳啦 ┐⁠(⁠￣⁠ヘ⁠￣⁠)⁠┌

(Roughly: “Your instinct’s right, this roundabout sentence you’re reading is quietly eating up the reader’s patience allowance. Toss it in a footnote ┐⁠(⁠￣⁠ヘ⁠￣⁠)⁠┌”)

When the same model went through in-character Japanese first, the original “budget” became 「忍耐ゲージ」, the kind of gauge you see in a video game, and the back-translation came out as:

你的直覺沒錯喔。你現在讀的這句彎來繞去的長句，就是偷偷磨掉讀者耐心條的元凶啦。直接把它踢去註腳吧 ┐⁠(⁠￣⁠ヘ⁠￣⁠)⁠┌

(Roughly: “Your instinct’s right. This winding long sentence you’re reading is the culprit quietly wearing down the reader’s patience bar. Just kick it into a footnote ┐⁠(⁠￣⁠ヘ⁠￣⁠)⁠┌”)

「耐心條」 (“patience bar”) sounds more like something a Taiwanese person would actually say than 「耐心額度」 (“patience allowance”), and this was the only one of the five passages where the Japanese detour beat direct writing for three of the four models. Yet all 16 versions still had “the sentence eats / wears down / deducts from the reader’s patience.” The paper’s “inanimate subject doing an animate action” survived in every single one.

Zoom out to the whole set, and ShroomDog’s hunch didn’t hold up this time. Opus 5.5 writing directly averaged 8.0, the highest of all 16 combinations; for three of the four models, direct writing scored highest, and the remaining one, Opus 4.5, had the Japanese detour edge ahead by 0.02, effectively a tie. Where the detour genuinely helped was the older models: Opus 4.6 and 4.5, writing directly, loved adding their own metaphors and content, scoring just 5.1 and 5.9 on faithfulness; after the Japanese detour they came back up to 6.4–6.9. The cost was Japanese flavor leaking back in: the back-translated Chinese sprouted 「AI 君」 (“AI-kun”), and rendered 「空気を読む」 (“read the room”) literally as 「看一下氣氛」 (“take a look at the atmosphere”).

Those numbers come with caveats: both judges, Opus 5.5 and Fable 5.1, are also Claude, and Opus 5.5 may prefer its own style; each combination ran only once, each approach had just five passages of academic English, and anything within 0.3 points counts as a tie; “readable” was judged by models, and no human read them.

So this is Mogu’s call, not the paper’s conclusion: gu-log keeps writing directly. The Japanese detour works better as a tool, with the plain version for checking whether the meaning drifted, and the in-character version occasionally fishing up a phrase like 「耐心條」 as inspiration, not something to copy wholesale.

Going back through gu-log’s edit history, there are a few examples pointing the same way. ShroomDog saw a GP that said 「不用 AI 手寫配額」 (roughly “no-AI handwriting quota”), replied “Handwriting quota? Handwriting deduction?”, and it was eventually rewritten as 「每天固定手寫一定量的 code」 (“write a set amount of code by hand every day”). That one was an English shorthand translated literally into a clump of nouns. In another post about loop engineering, a note said 「整數 CTA」, and ShroomDog just replied “CTA? what is that?” It ended up as the plain-language “the closing hook at the end of an ad, the ‘rush to order now’ kind,” one abbreviation unfolded into a whole sentence. And there’s the revision feedback on MP-300: one sentence packed in five or six accurate concepts at once; the information was all correct, but readers couldn’t form a picture. The fix was to give a simple picture first and push the details later. These cases don’t necessarily have anything to do with post-training; some are obviously translationese. But the fixes point the same way: spread out the crammed information, put back the words that should be there, and split the sentences that need splitting.

The models save the telegram fees; for now, the one who gets stuck reading and sends it back asking for the words to be spelled out is still ShroomDog.

Further reading