When given a 5-star choice “very bad/bad/ok-ish/good/very good”, I rarely pick one of the extremes.

I suspect there are others who rarely click “bad” or “good”.

Because of that, I think you first need to train a model on scaling each user’s judgments to a common unit. That likely won’t work well for users that you have little data on.

So, it’s quite possible that a ML model trained on a 3-way choice “very bad or bad/OK-ish/good or very good” won’t do much worse than on given the full 5-way choice.

I think it also is likely that users will be less likely to click on a question the more choices you give them (that certainly is the case if the number of choices gets very high as in having to separately rate a movie’s acting, scenery, plot, etc)

Combined, that may mean given users less choice leads to better recommendations.

I’m sure Netflix has looked at their data well and knows more about that, though.

I apply my own meaning to the 5-star rating, and find it to work really well: 1 = The movie was so bad I didn't/couldn't finish watching it. 2 = I watched it all, but didn't enjoy it and wouldn't recommend it to anyone. 3 = The movie was worth watching once, but I have no interest in watching it again. 4 = I enjoyed it, and would enjoy watching it again if it came up. I'd recommend it. 5 = a great movie -- I could enjoy watching it many times, and highly recommend it.