Where in a conversation does the English actually show up?

Not only in the final report — it also appears in short remarks like "I'll check next", in question choices, and in explanations of operations. The check targets all text the user sees.

Why not just write more rules telling the AI to use Japanese?

The developer found that as a conversation grows longer the instruction fades, and English starts mixing into permission screens and similar text. He stopped adding rule text and instead inspected the content right before it is displayed.

What happens when the check gets it wrong?

To avoid a conversation that can never end, the same reply can be sent back for rewriting at most 3 times. The developer also learned to test hook changes on a copy before swapping them in, after a hook fix once stopped all operations.

Large Language ModelsZenn AI/ML4d agoOct 15, 2026, 10:00 JST

683 of 1,763 AI replies came out in English despite Japanese settings

Send on LINE

3 Key Points

What happened
A developer who told his AI to answer in Japanese counted 683 English replies out of 1,763 responses recorded before he added any checks. He then added hooks at three points: on receiving a user message, just before an action, and at conversation end.
Why it matters
Long conversations appear to weaken an instruction to write only in Japanese, so checking just the final answer is not enough — the check has to cover choices, plans and action explanations too.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The setup here was language settings plus a rule to write only in Japanese, which the developer assumed would be enough. In practice, longer conversations diluted the instruction, and English began appearing in permission-screen explanations and similar places. Rather than keep adding rule text, he moved the check to the moment just before display.

The hooks divide the work by timing. Code blocks, inline code, URLs and paths are excluded, and product names are wrapped in inline code so they are not mistaken for English sentences.

The developer is explicit about the limits: a remark between operations can still flash on screen once before being corrected, and the hook does not reach in-conversation thinking summaries or the AI tool's own interface text. He suggests starting small with the conversation-end check alone.

FAQ

Where in a conversation does the English actually show up?: Not only in the final report — it also appears in short remarks like "I'll check next", in question choices, and in explanations of operations. The check targets all text the user sees.
Why not just write more rules telling the AI to use Japanese?: The developer found that as a conversation grows longer the instruction fades, and English starts mixing into permission screens and similar text. He stopped adding rule text and instead inspected the content right before it is displayed.
What happens when the check gets it wrong?: To avoid a conversation that can never end, the same reply can be sent back for rewriting at most 3 times. The developer also learned to test hook changes on a copy before swapping them in, after a hook fix once stopped all operations.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.