On top of the real concerns around otherwise selectable text in a writing system not supported by the user's keyboard, there's also the issue of whether or not they can even operate enough of a keyboard to transcribe whatever text they want to translate.
Just do whatever you want and then listen to your actual users' feedback.
I worked on an application that I had to make button text not selectable because the old people using it kept selecting text on the buttons by mistake instead of clicking/activating the button and getting stuck during a clinical trial.
Should I have left it selectable to pass the HN accessibility shamers purity tests, or listened to the users?
I empathize with translation, as I have to do it to pretty much every chipset firmware documentation I come across. So I just don't really understand where all of these issues are occurring with people not being able to translate stuff. Feels like a lot of people are maybe using a lot of websites that they aren't the target users for...
That's good advice. But there's an important caveat: telemetry is not user feedback.
This is where "data driven" approach often fails in practice: telemetry isn't feedback, it's evidence you gather to help you guess the user feedback in lieu of actually getting it. When that's not understood and given proper care (which is approximately always, because everyone has too little time and too many stakeholders breathing down their necks), it's very easy to just find proof for your own preconceptions in the data stream.