"emergent features" Yes, exactly. Bigger picture: one problem with moving to give algorithms more control is that you can't always anticipate the behavior of

"emergent features" Yes, exactly. Bigger picture: one problem with moving to give algorithms more control is that you can't always anticipate the behavior of algorithms. Especially when the algorithm is being targeted by an attacker. You need to predict not just best-case or average-case behavior, you need to know all the possible worst-case behaviors. There's a direct analogy to computer security, where we have no answers to how to make systems hack-proof. Systems and algorithms are hard and we don't have a solution to prevent exploitation. Humans are much harder to exploit. If Russia tries to get Ukrainians banned from an event by appealing to humans, that's much more attacker-resistent than when Russia gets Ukrainians banned from Facebook by exploiting an algorithm.