I get your point, I think along those lines quite often myself.
As per the sabotage part, bad input data that does not accurately get labeled as such until way too late in the “AI learning cycle” is I think the way to go. Lots and lots of such bad input data. How we would get to that point, that I don’t know yet, but it’s a valid option going forward.
> How we would get to that point, that I don’t know yet, but it’s a valid option going forward.
Chaos engineering. As a modern example, all this gender identity stuff wreaks absolute havoc on credit bureau databases.
Tomorrow, we'll have people running around in fursuits to avoid facial recognition. After that, who knows.
Sabotage will be the name of the game at that point. Find ways to quietly confuse, poison, overwhelm and undermine the system without attracting the attention of the monitoring apparatus.