It's quite a bit different in that AIs don't make such chains of reasoning. They just "notice" correlations between inputs and training data with no regard to m

It's quite a bit different in that AIs don't make such chains of reasoning. They just "notice" correlations between inputs and training data with no regard to mechanism. We might construct a chain of causation between a microexpression -> social status -> stress -> blood pressure -> life expectancy, but the AI doesn't and doesn't care. It's about as likely to judge by the lighting of the photo as the presumed character of the subject. Such AI judgement may be as or more racist than human judgement, given that they could make correlations based on immutable characteristics. In a sense they're only as biased as the training data, but that may also be true of many human racists.