[wanderer2323]

When you are training your search models you take editorial position all the time. There is no 'one single truth' of what the search results should be, let alone what the search-related artifacts (query suggestions, drill-downs, answers, etc) should be.

Real people go to work every day to label images as 'corresponding' or 'not corresponding' to a query, different people write guidelines for these labeling, other people curate which queries to label and which labeled results to train, etc etc.

In theory, all these people or at least their accumulated work produce some kind of 'neutral' result; in reality, a systemic bias on some of these levels can easily have an editorial effect that is impossible to prove.

[Balgair]

I'm certain that this is going to come up in anti-trust suits as a violation of Section 230. I can't see how it isn't editing the content that folk's see, despite the complexity of the ML. Judges don't like being told 'oh, it's too complex to understand, gee-shucks!' let us get away with it.

Surprisingly, such a ruling will have some philosophic ramifications about the ability for a computer to think and then edit. Though the case is unlikely to hinge on that semantic point, it will be talked about in lawyer bars.