The short answer.
Open an article of a widely read text on the English site, for example Article 655 of the Penal Code, and you now read it in English under the official Arabic. The translation is by a machine, it is unofficial and it says so, with its date. It covers 20,679 articles of 306 texts. The Arabic pages are unchanged and never show it.
Why.
The English pages of these texts showed Arabic under an English menu. Only the titles had machine English. A reader who cannot read Arabic could find a law and not read it.
Which texts.
The texts that people read: those with 5 or more page views in the last 60 days, in any language. That was 307 texts on 2 November. Their English pages were already the only machine-English pages open to search engines.
How it was made.
Gemini translated the articles in batches of about 8 articles, run from a timer every two hours until all were done. A batch that failed was split in half and tried again, down to one article. The first 5,386 articles went live on 5 November and the rest on 6 November.
What keeps it honest.
Each translation is tied to a hash of the Arabic it was made from. If the Arabic text changes, its translation is no longer shown until it is made again. It is not shown either when the page displays a newer wording of the article. A check flags every number in the Arabic that is missing from the English. In the first batch it flagged 704 lines, and a sample of 20 were all false alarms, such as the article's own heading.
What is missing.
One article of the 20,680: the model stopped mid-sentence on most tries, so we left it out. The translation is a reading aid. For anything that matters, read the Arabic above it and the source it links.