> we know adversarial images could be created that will trick the system
It could maybe be used to perform a kind of DDoS attack on the human verification stage by increasing false-positives, but I doubt a non-child-porn image would fool a human into thinking it looks like child porn just because of some carefully-applied noise.
Also, in what conceivable way is this like Minority Report, ie three psychic humans, floating in a pond, hallucinating the future?