I’ll give you an example, using PT→EN because I don’t speak JP. Let’s say Alice tells Bob “ma’ tu é uma nota de três pila, né?” (literally: “bu[t] you’re a three bucks bill, isn’t it?”) . A human translator will immediately notice a few things:

• It’s an informal and regional register. If Alice typically uses this register, it’s part of her characterisation; else, it register shift is noteworthy. Either way, it’s meaningful.
• There’s an idiom there; “nota de três pila” (three bucks bill). It conveys some[thing/one] is blatantly false.
• There’s a rhetorical question, worded like an accusation. The scene dictates how it should be interpreted.

So depending on the context, the translator might translate this as “ain’t ya full of shit…”, or perhaps “wow, you’re as fake as Monopoly money, arentcha?”. Now, check how chatbots do it:

• GPT-4o mini: “But you’re a three-buck note, right?”
• Llama 4 Scout: “But you are a three-dollar bill, aren’t you?”; or “You’re a three-dollar bill, right?” (it offers both alternatives)

Both miss the mark. If you talk about three dollar bills in English, lots of people associate it with gay people, creating an association that simply does not exist in the original. The extremely informal and regional register is gone, as well as the accusatory tone.

With Claude shitting this pile of idiocy, that I had to screenshot because otherwise people wouldn’t believe me:

[This is wrong on so many levels I don’t… I don’t even…]

This is what you get for AI translations between two IE languages in the same Sprachbund, that’ll often do things in a similar way. It gets way worse for Japanese → English - because they’re languages from different families, different cultures, that didn’t historically interact that much. It’s like the dumb shit above, multiplied by ten.