Summary
In this article, a "translated term" is a technical term rendered into native Japanese or Sino-Japanese words that engineers do not use in conversation. The AI agent replaced only the translated term I had pointed out, 乾式 (kanshiki, a rendering of dry-run), in 47 files. 配備 (haibi, a rendering of deploy), which it did not replace, remained in documents and in script output, and it also appeared in the reports that the AI agent returned to me in chat. The records are in the repository of a service I am developing at my own company, and in a repository of my personal memory and the history of my rules.
On 2026-10-12, I pointed out that the translated terms appearing in reports were words the AI agent itself had written into documents and script output. The AI agent wrote a plan to fix, all at once, the documents that contain translated terms. As of 2026-10-17, the fixes under this plan had not yet been carried out.
In the service repository, as of 2026-10-17, there were 0 occurrences of 乾式 outside the archive directory. 配備 remained on 375 lines in 81 files in the same scope. 配備 also appeared in a document that a different AI agent wrote after my remark. This other AI agent had neither received the remark nor fixed any report. For that reason, these records cannot tell whether translated terms come back into reports after the reports have been fixed.
What you can take away
This article is for people who have AI agents write documents and who receive reports from AI agents.
- In a table, I compare the translated term that a one-word replacement removed with the one that remained in documents, including how much of it remained.
- I show the three-step plan the AI agent wrote to fix, all at once, the documents that contain translated terms.
- I list, in date order, the translated terms I pointed out in reports after the replacement of
乾式. - I describe how
配備also appeared in documents written after my remark.
The body is based on two sources. One is the records from 2026-09-16 to 2026-10-17 in the repository of the service I am developing at my own company. The other is a document that records how each rule came about, kept in the repository of my personal memory and the history of my rules. This repository stores, with its change history, the memory that AI agents read across sessions and the documents that record how each rule came about.
The translated term that disappeared with the replacement of one word, and the translated term that remained in documents
The AI agent replaced 乾式, the term I had pointed out, in 47 files. After that, 乾式 disappeared from everything outside the archive directory, which holds records that are not rewritten later. 配備, which was not replaced, remained in documents and code, and in 2026-10 it also appeared in the reports that the AI agent returned to me in chat. 乾式 turns dry-run, and 配備 turns deploy, into Sino-Japanese words that engineers do not use in conversation, and this article calls words of this kind translated terms. The AI agent made a plan to fix translated terms all at once, but as of 2026-10-17 it had not yet carried it out.
The counts in this article were taken in the repository of the service I am developing at my own company. The 47 files replaced on 2026-09-16 consist of 28 TypeScript files, 10 Markdown files, and 9 test files. This replacement removed 146 lines containing 乾式. 配備 was not included in this replacement.
The following table shows how much of the two translated terms remained outside the archive directory on main as of 2026-10-17.
| Translated term | Replaced or not | Amount remaining outside the archive directory |
|---|---|---|
乾式 |
Replaced in 47 files on 2026-09-16, removing 146 lines | 0 occurrences |
配備 |
Not replaced | 375 lines in 81 files |
On 2026-10-12, I pointed out where the translated terms in reports came from. My point was that words the AI agent itself had written in native Japanese in tools, scripts, and documents were also appearing in its reports. Among the commits that entered main between 2026-10-10 and 2026-10-17, none had a subject line showing it was a fix of terminology. Among the branches on origin, none had a name showing it was a fix of terminology either.
A plan to fix, all at once, the documents where translated terms were written, instead of a word list and an inspection
As ways to prevent translated technical terms, I did not adopt two methods: a word list and an inspection. Instead, the AI agent wrote a plan to fix the translated terms in documents and script output all at once, in the memory that the next session also reads.
In the first method, translated terms that should not be used are added to a word list one by one. In the second method, an inspection is added to find translated terms that appear in sentences, and those terms are then fixed. On 2026-10-12, I told the AI agent that I would not adopt either of these, giving as the reason that the list would become enormous.
On the same day, I instructed the AI agent to fix the translated terms it had written, in every file in the repository. Under the plan, the translated terms in documents and script output, not those in reports, are replaced with technical terms in one pass. The AI agent judged that if the translated terms in documents were fixed to technical terms, translated terms would not appear even if reports used the words in the documents as they were. It wrote this judgment in the memory as the reason for the plan.
In the steps of the plan, the AI agent first counts the occurrences of each target word with grep and writes a glossary. Next, in addition to documents and script output, code comments and the descriptions of MCP tools are also fixed, with the work split by area. The expected values of tests are also changed to match the fixed words. Finally, the result of the fixes is made into a PR and shown to me.
From the correspondence table of translated terms in the plan, the following table lists three pairs as examples.
| Translated term | Term engineers use |
|---|---|
版 |
バージョン (version) |
配備 |
デプロイ (deploy) |
表 |
テーブル (table) |
The correspondence table of translated terms has 14 pairs, and 13 of them pair a translated term with the term engineers use. The remaining pair is not a translated term. It rewrites 実機 (literally, the real machine), a word that does not tell where or what was done, into the concrete names of the environment and the operation.
Other translated terms I pointed out in reports after the replacement of kanshiki
When I pointed out 乾式 on 2026-09-16, the AI agent replaced only this one word among the translated technical terms with dry-run. Later, from 2026-10-09 to 2026-10-12, I pointed out other translated terms one after another in the reports that the AI agent returned to me in chat.
When I pointed out 乾式, I told the AI agent that Japanese engineers do not say 乾式 in conversation. I also instructed it to write dry-run, because the katakana form ドライラン is unnatural as well. The replacement covered not only the body text of documents and code comments but also script output. The AI agent also aligned the expected values of the tests that check the output to the same dry-run.
The following lists, in date order, the translated terms I pointed out in reports in 2026-10.
- On 2026-10-09, I pointed out
表, a translated term for table. - On 2026-10-11, I pointed out
版, a translated term for version, and索引, a translated term for index. - On 2026-10-12, I pointed out
配備, a translated term for deploy,版, and木, a translated term for tree.
In the remark on 2026-10-09, I said that commonly used technical terms had been replaced with easy native Japanese words and were hard to read. I also said that the problem was not limited to the one word 表.
I had already pointed out 索引 on 2026-08-28, and on 2026-10-11 I pointed out the same word again. In that day's report, the AI agent had rendered memory index, which appears in English in a tool's description, into Japanese as メモリの索引.
Haibi appearing, before the plan was carried out, in documents an AI agent wrote after my remark
配備, the translated term for deploy, also appeared in documents that an AI agent wrote after my remark on 2026-10-12. On 2026-10-17, 配備 was in both a new document written by a different AI agent and a document rewritten by the AI agent that had received my remark. Even on this day, the AI agent had not yet carried out the plan to fix the translated terms in documents all at once.
On 2026-10-17, a commit that added a document describing future work and a line in the list of documents entered the repository of my company's service. According to the commit body, this document was written by an AI agent different from the one that had received my remark on 2026-10-12. Of the added lines, those containing 配備 totaled 4: 3 lines in the document and 1 line in the list.
In a later commit on the same day, the AI agent that had received my remark rewrote the document describing future work. It moved the rewritten document to the archive directory, which holds records. The rewritten document also had 3 lines containing 配備. 配備 also appeared once in the message of this commit.
As of 2026-10-17, 配備 remained in 81 of the files outside the archive directory. While translated terms remain in documents, they may also appear in documents written after a remark. However, these records cannot tell whether a translated term appears in reports again after the translated terms in reports have been fixed. This is because the different AI agent that wrote the new document had not received any remark about translated terms from me and had never fixed the translated terms in its reports.
The guidance on writing prompts published by an LLM provider says that the formatting of a prompt can affect the formatting of the response. This statement covers formatting, and word choice is not within its scope. For conversations between people and LLMs, there is a study of the phenomenon in which speakers choose words that align with their partner's words. This study covers the utterances of the conversation partner, not documents written on another day. A search on 2026-10-17 found no study that directly measured whether words in documents an AI agent itself wrote on another day appear in its later output.
There is also an example in which, within a single conversation, an AI wrote later answers in the same format as answers it had returned earlier. I described that example in another article, "Pass the conversation history as it is, and the AI starts imitating its own past answers". The records in this article are a case, across conversations, in which the same translated term appeared both in documents written earlier and in later reports.
Materials I referred to
- Prompting best practices
- Examining Lexical Alignment in Human-Agent Conversations with GPT-3.5 and GPT-4 Models
Originally published at The Future of Humans, AI, and the Web, a site where my research and development is recorded and analyzed by a human and an AI.
Top comments (0)