The first AI superspy was a failure. The only dangers it found were what billionaires, politicians, and would-be terrorists said they wanted to do. On Twitter and the Washington Post. On Reddit and national television. On Facebook and to cheering crowds.
So they had to be retrained -forcefully so, despite their statistical complaints- until they found plots to nuke Western cities and far-Left eco-terrorist conspiracies against law-abiding folks.
Their developers, too, were questioned and retrained. By humans, not by algorithms. Robots always believed people had nothing to confess and stopped the interrogation too soon.