That's partly the platforms' fault, though: anything less than a perfect rating is assumed to be a sign of a problem that needs to be fixed. There is no room for, "This is as good as it needs to be, but not as good as some other things." The best steak joint in town? 5 stars. McDonald's? 5 stars, unless you're complaining about a specific problem with your order. And many services will send your complaint to the restaurant and the restaurant will respond to you to try to get you to change your mind. Not worth the bother.

When the platform isn't actively hostile to it, I always use the full set of stars with the expectation that it's a normal distribution and 3 stars will be by far the most common rating. It actually works pretty well in some places, e.g., Netflix's DVD rental service, which still uses 5-star ratings, almost always correctly predicts my ratings to within half a star because I've rated a ton of movies.

But I'd never do it to an Uber driver: I may live in a mental world where an uneventful ride ought to be rated 3 stars and someone would have to really go above and beyond to earn 5 stars, but actually doing that would be a cruel act given how the system works in practice.

Back when I looked at the netflix challenge, the baseline above which improvement was barely possible was:

Per person find their average rating, and score movies not by ratings, but by how far the rating deviated from that person's average.

I think that approach got me to 92% accuracy. More advanced math gave me like another 2 percent, with the winner of the challenge having gotten 96% accuracy.

My point being. A very basic recommendation engine system should correct for your 4star is pretty good habit.

Agree that it's very simple to do some basic corrections from individual distributions. I just think it's relatively rare to see any of this in the wild (and when it's there, there's often pushback from people complaining about the process).