You could design an AI to employ a similar heuristic, though perhaps it would be pattern-matching against a designated model human, rather than against itself. But the edge cases show that you need better heuristics than that

I don't see how this is different from having the AI recognise a teacup. We don't actually know how we do it. That's why it is difficult to make a machine to do it. We also don't know how we recognise people. "Presuming that observed similarities etc." isn't a useful description of how we do it, and I don't think any amount of introspection about our experience of doing it will help, any more than that sort of thinking has helped to develop machine vision, or indeed any of the modest successes that AI has had.