We used this a couple times at Sandstorm back in the day. At the end of the interview the candidate would play Factorio cooperatively with the team for a while.
I think it is remarkably effective at identifying the kinds of skills and personality traits that a software engineer actually needs to have in day-to-day work. You can find out if someone is self-directed, how fast they work, whether they produce clean designs or spaghetti code, whether they are good at cooperating or tend to go off on their own, etc. Some people will just sit and watch and do nothing unless instructed... that's bad. Some people will build stuff, but with obvious efficiency flaws and "bugs"... also bad. Some people identify what needs doing and get it done effectively but without trying to be perfect... that's good. Factorio essentially compresses real-world work patterns into a shorter time period, giving you the chance to see how someone works in the space of a couple hours. I don't know anything else that does that.
But, obviously, it's problematic. A person who has played the game before will have a big advantage. A person who doesn't play video games at all will have a big disadvantage. For these reasons, I don't think I could seriously recommend this approach be adopted more widely.
But then, I don't know of any approach to interviewing that I think is fair. Everything has problems:
* Regular interviews bias towards people that are charismatic, not effective.
* Coding interviews bias towards people who can code under pressure with someone looking over their shoulder, which is almost nothing like real-world coding.
* Puzzle-y algorithms questions identify people good at clever solutions but that's hardly what most people need to do day-to-day when coding.
* Take-home assignments bias towards people who have lots of free time, and still don't really tell you how that person works with others.
* Looking at GitHub history biases against people whose lives are too busy to code in their free time (e.g. maybe they have kids) and who weren't lucky enough to have a previous job that touched a lot of open source.
* Looking at past work experience misses brilliant junior programmers coming out of college.
* Don't even think about using ML for this, that doesn't solve biases, it just hides them in a black box.
It seems like there's no right answer here.