AI-like Japanese, and yomiyasu, which fixes it without changing the meaning

Contents
  1. How writing changed after generative AI
  2. Why AI-like Japanese appears
  3. Examples from medical documents
  4. Steps for checking
  5. What yomiyasu is
  6. What yomiyasu does well
  7. How to use it
  8. References

Japanese written by generative AI can be grammatically correct and still leave you unsure who does what. This article summarizes the features of this AI-like Japanese and steps for checking it, and introduces what the public tool yomiyasu does well.

How writing changed after generative AI

One study counted 70,524 articles on Qiita, a Japanese site for engineers, from 2019 to 2026. The average sentence length went from 35.4 to 39.2 characters, within the natural increase seen before AI. What changed a lot was formatting: bold text increased about 3-fold, bullet lists about 1.8-fold, and dashes about 10-fold. Things look more organized, but there is not more content1.

Vocabulary changed too. The share of articles using 効く (“to work, to be effective”) rose from 2.0% to 22.3%, 壊れる (“to break”) from 0.8% to 13.1%, and 実測 (“measured”) from 0.1% to 9.4%.

Why AI-like Japanese appears

The pretraining data of publicly available large language models is heavily skewed toward English2. Studies report that multilingual models show English syntax when they write in other languages3. Text that a model writes directly, without translation, is also often judged as translationese4. The English phrase “silently break” becoming 静かに壊れる (literally “breaks quietly”) is an example of this literal translation.

Examples from medical documents

Here are four example sentences, written for this article, imagining a patient information leaflet, an in-hospital manual, and a referral letter.

  1. Inanimate subject with a metaphorical verb

    AI-like: 丁寧な説明が、患者さんの不安に寄り添います。 “The careful explanation stays close to the patient’s worries.”

    Revised: 担当医は、患者さんが不安に思っていることを聞き、質問に答えます。 “The attending physician asks what the patient is worried about and answers their questions.”

    Write who does what to whom.

  2. Stacked nouns

    AI-like: 説明資料の共通化により、説明内容の標準化を図る。 “Standardization of explanation content is pursued through commonization of explanation materials.”

    Revised: 担当医は、共通の資料に沿って必要な項目を説明する。 “The attending physician explains the necessary items following the shared materials.”

    Turn the nouns back into verbs.

  3. Contrast that denies a misunderstanding nobody stated

    AI-like: 不眠は、単なる睡眠の問題ではなく、こころからのサインです。 “Insomnia is not just a sleep problem but a sign from the mind.”

    Revised: 眠れない日が続くときは、診察で睡眠や生活の状況、気分の落ち込みについて伺います。 “If you keep having nights without sleep, at the consultation we will ask about your sleep, daily routine, and low mood.”

    Write the needed information as a plain statement.

  4. Vague pointing words and roundabout openers

    AI-like: 重要なのは、この点について事前に情報を共有していただくことです。 “What is important is that you share information on this point in advance.” (Premise: the visit is on October 9, and recent blood test results are to be sent by the day before.)

    Revised: 10月8日までに、直近の血液検査結果を当科へお送りください。 “Please send the most recent blood test results to our department by October 8.”

    Write out what the pointing word refers to.

In the fourth example, the date and what to send were already decided as the premise. The revision only writes that information out; it does not add anything that was not there.

Steps for checking

  1. Write who did what.
  2. Replace metaphorical verbs with everyday words.
  3. Turn chains of nouns back into verbs.
  4. Remove unnecessary negations and openers.
  5. When revising, do not add information that was not in the original.

What yomiyasu is

yomiyasu is an Agent Skill that fixes unnatural metaphors, unclear subject–predicate relations, and unnecessary decoration in AI-generated Japanese, and makes it easier to read5. nanaism published it on GitHub in September 2026 under the MIT License; the README says it was built at ALGO ARTIS, Inc. It is aimed mainly at practical writing such as technical articles, design documents and specifications, pull request descriptions, and internal reports. You load it into AI coding environments such as Claude Code, Codex, and Cursor. As of November 14, 2026, the latest version is 3.1.0 and the repository has 3,225 stars on GitHub.

What yomiyasu does well

Before rewriting, yomiyasu checks four things in the original sentence: the claim, the relative weight, the strength of assertion, and the function of the sentence. Keeping these four the same after rewriting is its top priority. It even specifies how to handle strength, for example not turning “may happen” into a definite statement and not adding “always”. In patient leaflets and referral letters, turning a guess into a definite statement makes the content wrong. Putting meaning before readability suits medical documents.

It is also explicit about not adding information. It does not fill in, by guesswork, the actor, cause, condition, or number that the original does not make clear. Where the meaning cannot be settled, it does not put notes in the text; it asks the writer in a section after the text. The revision of the fourth example above follows the same idea.

The way it edits is also defined. It does not fix text by bulk find-and-replace; it reads the surrounding sentences and fixes one place at a time. Each proposed edit is compared with the original on six points: what is being talked about, what is affirmed and what is denied, who does what to whom, scope and conditions, strength of assertion, and numbers, proper nouns, and dates. If any one of them changes, the original is kept. Finally, the bundled scripts compare the text before and after, and check that the text being submitted is the same as the text that was checked.

It also changes how it edits depending on the type of document and the reader. There are 13 document types, including pull requests, articles, incident reports, minutes, email, and papers, and three reader levels from experts to general readers. Each type has rules that are not applied; in essays and fiction, for instance, metaphors are kept when it is clear what is being compared to what. The README also says that parts that already read naturally are left alone.

How to use it

The recommended way to install it is this command:

npx skills add nanaism/yomiyasu

After installing, ask your AI agent something like “draft.md を yomiyasu で読みやすくして” (make draft.md easier to read with yomiyasu). For environments where the full skill cannot be installed, the smaller yomiyasu-mini and yomiyasu-pocket, made from the full version, are also available.

The Japanese version of this article was also checked with yomiyasu.

References

  1. Sakasegawa-chan. How did engineers’ writing change before and after generative AI? Counting 70,000 Qiita articles (in Japanese). September 2026. https://nyosegawa.com/posts/qiita-writing-before-after-ai/
  2. Takase S, Honda U. Toward moving away from English-centric large language model development (in Japanese). Proceedings of the 32nd Annual Meeting of the Association for Natural Language Processing (NLP2026), C7-23. https://arxiv.org/abs/2605.15613
  3. Guo Y, et al. Do Large Language Models have an English Accent? ACL 2025:3823-38. https://aclanthology.org/2025.acl-long.193/
  4. Valentini M, et al. An Investigation of Translationese in the Generations of Multilingual LLMs. COLM 2026. https://arxiv.org/abs/2608.17399
  5. nanaism. yomiyasu. GitHub. https://github.com/nanaism/yomiyasu
English translation of the Japanese article.