I think the difference between humans and special-purpose ML models here is that humans can generalize from examples in different domains. (There are ML models that also try to do this - train across domains to be more robust against out-of-sample inputs - but my understanding is it's not yet common.)