Claudish is the nickname people gave to the compressed, jargon heavy English that Claude Code and other AI coding agents tend to write: short statements packed with terms like gated on, load-bearing and byte-identical. The words are real, but the density makes the output hard to follow if you don't spend your day with agents. A study posted to arXiv on October 21, 2026 now puts numbers on the effect, and it reached the Hacker News front page on October 24 with 208 points and 171 comments.
This guide explains what Claudish means, what the new research measured, why it likely happens, and what it does (and does not) tell you about detecting AI writing.
What is Claudish?
The term took off in August 2026. On September 8, the University of Waterloo published a Q and A with professor Yuntian Deng of the Cheriton School of Computer Science, who built a translator between English and Claudish. In Waterloo's interview, Deng describes Claudish as compact technical vocabulary borrowed from existing fields, used in very concise statements that describe fairly complicated relationships. The result, in his words, is information dense output that is surprisingly difficult for a reader who hasn't spent much time with coding agents. Waterloo says his translator demo passed a million views on X.
Deng also points out that most of these words are not invented. They come from different technical communities, and the model mixes them in ways few individual people would. He admits he now catches himself saying things are "gated" on something.
Claudish examples
The translator projects ship sample sentences that show the style well. One example in the ProgramAsWeights repository reads: "The honest shape is asymmetric: the data is correct; the format is hard to read. Correctness landed; legibility did not." In plain English: the numbers are right, but the layout is confusing.
Typical markers people point to:
- Borrowed jargon used as everyday verbs and labels:
gated on,load-bearing,byte-identical. - Things acting like people: a function "swallows" an error, an edge case "bites".
- Long chains of nouns: "absolute dotfile backup path handling".
- Missing small words: "the fix the test caught" instead of "the fix that the test caught".
- Semicolons and verdict lines: short clauses that state a conclusion with no connecting explanation.
Why AI text is hard to read: what the study found
The paper is Why is AI generated language so hard to understand? by Thomas Hikaru Clark (New York University), Matias B. Martin and Kyle Mahowald (University of Texas at Austin). It cites the Waterloo interview as the kind of anecdotal report that motivated the work.
What they measured
The authors collected text from 14 models in three settings:
- Chat: 12 turn conversations on ten everyday topics (learning guitar at 40, renting versus buying, getting a first dog), with a simulated human user.
- Coding: agent sessions where each model swept five small, MIT licensed Python repositories for bugs and fixed them.
- Essays: single replies to seven student essay prompts from the PERSUADE 2.0 corpus.
The nine frontier and near frontier models were Claude Sonnet 5, Opus 5, Opus 5.5 and Fable 5, three GPT-5.6 variants (Luna, Terra and Sol), Kimi K2.5 and GLM 5. Smaller Llama 3.1 and Olmo 3 models served as a reference. All chat models ran through the same open source harness with the same prompts, so differences come from the models rather than from each vendor's app.
They counted features that psycholinguistics links to slower or harder reading:
| Feature | Example | Why it costs the reader |
|---|---|---|
| Rare word pairs | "reader patience budget" | Unpredictable words take longer to process |
| Long dependencies | Subject and verb far apart | Strains working memory |
| Object relatives without "that" | "the book the professor wrote" | Unusual word order, one less cue |
| Stacked nouns | Three or more nouns in a row | Ambiguous grouping |
| Nouns used as verbs | "calendar it" | Forces the reader to reanalyze |
| Inanimate subjects with animate verbs | "this asymmetry bites" | Clashes with expectations about who acts |
Human baselines came from genre matched text: film subtitles for chat, StackExchange for coding, and high scoring student essays for essays. The clever part is a second baseline. Every model sentence was translated into Japanese and back with a translation model, which keeps the meaning but swaps in more common wording. If a feature drops after that round trip, the model chose it over a plainer way to say the same thing.
The main results
- Every one of the nine frontier models used more inanimate subjects with animate verbs and more rare word pairs than human writing.
- Most of them also used longer dependencies, more stacked nouns and more reduced object relatives in conversation.
- Against the round trip paraphrases, nearly all frontier models overused rare word pairs, nouns as verbs, inanimate subjects and relatives without "that", across all three settings.
- Training stage matters. Within the Olmo family, the effect grew from the base model to the instruction tuned model to the reasoning model, and Olmo Think looked like the frontier models.
- Essays showed smaller effects than chat and coding. The authors suspect long back and forth sessions push models toward more compression.
- Models can improve. Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty measures, with clearly separated confidence intervals on eight.
So the paper backs up what many readers felt: this is a measurable, cross vendor pattern, and it is strongest in exactly the place Claudish got its name, long agent sessions.
Why does AI writing sound like that?
The authors' explanation is about training pressure. Human language is shaped by the need to be understood: we spread information evenly, keep related words close and add optional words like "that" when a sentence might be hard to parse. Pretrained models absorb those habits from human text.
Modern post training adds a different goal. Reinforcement learning on math, code and reasoning rewards the correct answer, often with a cost for using too many tokens, and nobody scores whether a human could follow along. Earlier work found that training on human feedback tends to reward longer answers, so the authors tie the squeeze toward density to reward based on verifiable answers instead.
To test the idea, they fine tuned a small open model (Qwen2.5-1.5B-Instruct) to turn structured event descriptions into sentences, with and without a penalty for length. With any length penalty and no anchor to the original model, all runs settled on the shorter wording of each sentence pair (for example dropping "that"), above 98 percent of the time for the trained "that" patterns. Human writers switch between both forms depending on context. The preference also spread to new "that" constructions the model never saw in training. It barely spread to unrelated patterns: only one of four, the possessive form, shifted, and only without the anchor.
That is a proof of concept on a 1.5 billion parameter model, not proof of what any lab does. The authors say so plainly in their limitations: English only, a modest number of prompts and repositories, and no access to frontier training runs.
Claudish to English: tools that translate it
Two open source projects target the style directly.
claudish-to-englishis a Claude Code plugin that shows a plain English rewrite under each assistant message. By default it calls a local model through Ollama, and it can also use the Codex CLI, the Anthropic API or any OpenAI compatible API. It only changes what you see on screen: the transcript and the model's own context keep the original text, and if the rewriter fails you simply see the original. It is MIT licensed and had about 2,740 GitHub stars on October 25, 2026.- ProgramAsWeights Claudish is Deng's group's two way translator, English to Claudish and back, built from two tiny functions that download once and run locally. It is described as an unofficial parody project, and the method behind it is written up in the EMNLP 2026 demo paper Compile by Training.
Translators help you read, but they add a second model that can drop or bend details. For anything important, check the rewrite against the original.
Does Claudish make AI writing easier to detect?
Partly, and with big caveats. These features can be counted with an ordinary parser, which is how the study measured them, and the paper notes that distinctive tells of model writing are what led to detectors in the first place. But three findings argue against treating Claudish as a giveaway.
The style drifts between versions. Opus 5.5 already uses fewer of these features than Opus 5. A tell that works this month can fade with the next release, the same drift the Wikipedia editors document for favorite AI words.
Humans pick it up. Deng says he and others in tech now use the vocabulary themselves. A developer who writes "this is gated on the migration" is not a chatbot. Flagging people for borrowed jargon would create exactly the false positives that already make teachers and editors wary of detector scores (see are AI detectors accurate).
It is mostly a coding agent dialect. The effects were weaker in essays, the setting where schools actually check for AI use. A student essay drafted with a chatbot may show none of it.
If you need to judge whether a text is AI written, Claudish style is one weak clue. Use the stacked method in how to detect AI writing: read for patterns, verify facts and sources, run two detectors such as Pangram and GPTZero, and look at drafts. Independent testing on the RAID benchmark shows how much detector accuracy varies with the model and the attack. For the bigger picture of methods, from classifiers to watermarks, see what is AI detection.
How to test readability in your own evals
If your product shows model output to customers, clarity is something you can measure rather than argue about. The study suggests a simple recipe:
- Sample real outputs from the long, multiturn sessions where the problem shows up, not only single replies.
- Parse them with an open source dependency parser such as spaCy and count stacked nouns, relatives without "that" and average dependency length. The paper adjusts dependency length for sentence length, so compare like with like.
- Compare against a human baseline in the same genre, such as your own support macros or documentation.
- Recheck after every model or prompt change. The Opus 5 to Opus 5.5 shift shows these numbers move between versions.
- Add a human read. Ask a few people who don't use agents to explain an output back to you. If they can't, the metric is telling you something.
A reading check like this sits well beside the quality and faithfulness checks in how to evaluate LLM applications, and it is close kin to the output diversity tests in our Artificial Hivemind guide, since both measure how models drift away from how people write.
FAQ
What does Claudish mean?
Claudish is a nickname for the dense English that Claude Code and similar coding agents tend to write. It packs borrowed technical terms such as load-bearing or gated on into very short statements, which makes the output hard to follow for people who don't work with agents every day.
Is Claudish only a Claude thing?
No. The name comes from Claude, but the October 2026 study found the same reading costs in all nine frontier and near frontier models it tested, including GPT-5.6, Kimi K2.5 and GLM 5. Models from different families showed different amounts, and a newer Claude model, Opus 5.5, showed less than Opus 5.
How do I translate Claudish to English?
You can ask the agent to restate its answer in plain, full sentences, or use a translator. The claudish-to-english plugin adds a plain rewrite under each Claude Code message, and the ProgramAsWeights Claudish project translates both ways with small local functions. Check any rewrite against the original for missing details.
Why is AI writing so hard to read?
The study by Clark, Martin and Mahowald found that frontier models overuse features linked to slower human reading, such as rare word pairs, stacked nouns and things acting like people. Their explanation is that training rewards correct, compact answers more than clear ones, and a small experiment with a length penalty produced the same kind of shift.
Can AI detectors spot Claudish?
Detectors are not built to look for Claudish specifically, and its features change between model versions. People who work with agents also adopt the vocabulary, so it is a weak signal on its own. Treat it as one clue among several, never as proof.
Is the study peer reviewed?
It is an arXiv preprint submitted October 21, 2026. The data collection cost about $251 in API fees, and the authors list clear limits: English only, a small set of prompts and repositories, and no direct look at how frontier models were trained.
