I'll leave it as an exercise to the reader to consider what the police might use as a proxy/heuristic to figure out what group a suspect is in.

I agree that this is a danger. Fortunately, with additional data available, data scientists will be able to show any biases more clearly. If it happens with a rating system, it was happening before the rating system, except now there is the benefit of more data.

Similarly, you could also show if people from Group A (the more weighted group) also rate the police lower, despite similar treatment as those of Group B. The data will cut both ways.

EDIT>> Nvm, you were using "weighted" to mean "the weight of their vote", not "the weight applied to their vote."

One example of this is "voter fraud". It is possible to commit voter fraud, and so certain parties want strong identity guarantees to vote. But voter fraud isn't happening in practice, and stringent ID requirements actually suppress minority votes.

It is better to implement the simplest possible system first, observe it, and fix problems later as needed, after careful analysis of the running system.

But not the ordinary YAGNI. You're Absolutely going to need some way of dealing with low-quality input.

You really should wait and see before you invest a ton of time and money into solving problems that end up not being problems.

If you are actually operating under the assumption that the output doesn't matter at all, then implementing the review system in the first place is a waste of effort, let alone any features thereof. But assuming that we do care about the output, its quality does matter.

>There's also no evidence that there would be enough poor quality inputs to make a meaningful impact on scores.

In pretty much the same way that there's no evidence that any given new website needs to worry about sql injection.