Interestingly, it looks like some of the "bad examples" are precisely the kind of things ML would produce. Others are what non-native speakers would produce (starting with "あなたは" to translate a sentence starting by "you").

I'm sure some translators were using ML before it was integrated, and those guidelines are here in particular to tell them about those problems.

Also, ML is now really good to translate between European languages, but Japanese is very different in its structure so ML from English to Japanese is not as good. I'm sure some people who only know English/French/Spanish/German saw that ML is pretty good, and don't realize that for some other language it just doesn't work.

> Also, ML is now really good to translate between European languages

As somebody who has to regularly bear "German" machine-translated UIs and manuals that originate in English, I can only say: No, it's not. It's atrocious.

Best one was when gedit had the option to syntax highlight for a language named “Los.”

Not a bad name, to be honest!

lmao. This is a "research team and five years" task with current state of LLM.

Yes, but now that's the answer to questions like "how do we deploy AI without pissing off our communities?"

I have your answer: you provide a tool for your translators, you don't unleash a bot that makes changes left and right and creates new pages.

Typically, when a new page is written in English, don't automatically generate a version in all languages. When a translator starts creating the page in their language, provide a button to pre-fill with ML translation if they want to.

And for users, you can display the English version with a message, "this is not translated in your language yet but you can read an ML version if you want".

This sounds about right, and should be incorporated everywhere.

For reference: https://xkcd.com/1425/

I edited my comment to clarify I hope. Imagining what it could have done wrong and knowing what it did wrong are different.

Do you expect someone who has just watched a bot replace 20 years of their work, with no prior consultation or review, to now write a detailed post about how translations by the bot are not specifically wrong?

The core issue here is the way the bot was deployed. The fact that they had the poor taste to make it auto-replace articles written by their own volunteers is idiotic and disrespectful in the extreme. A new bot should work entirely in the back end, sending proposals for translations to the volunteers, who can choose to accept them or ignore them. Once the rate of acceptance is very high, for a specific individual language, then you might consider automating further.

And yes, this effort needs to be done for each language separately. Just because the bot works well in Italian doesn't in any way guarantee that it will work well in Japanese. Machine translation quality varies wildly by language, this is a well known and obvious fact.

This is Mozilla as usual, arrogant and tone deaf.