Very weird LLM behavior when asked to translate en_US source to en_GB

[unfulvio]

Hey, appreciate the work you did here.

I noticed today a strange quirk... which is probably for the LLM (OpenAI) to blame.

Basically I tried to translate a .pot source that is en_US into en_GB -- these have very minor differences but was hoping/testing to see how the AI would handle them. To my surprise the translations did not contain British (UK) English, but a variety of other languages, in the same file. All but English, of any kind.

I think the LLM hallucinated because the source is in English and did not understand the task? Very weird behavior but makes me think the same could happen in other language pairs within the same family. Say your strings are French from France but want to translate to Canadian French... Or I am not entirely sure how it would behave with non-latin writing systems in instances like the ones described.

I tried with both gpt-4.1-nano and gpt-4o-mini but didn't make a difference. I tried forcing the language with --lang but that also did not make a difference.

Since this may not be a bug with your tool you could close this, but I wanted to report as you might want to test yourself and maybe add some disclaimer. Or, if you have a solution around this, would be interested to know.

[thejamescollins]

I also noticed this today while translating from en-US to en-AU.
A lot of the translated strings appeared to be in Japanese.

Things worked well when translating to French (for example), but translating from one english variant to another did not produce good results at all.