This is true. A useful difference between this form of parameter optimization and the unalignability of humans is that it is possible to force constraints on these parameters. E.g. San Francisco’s speed cameras run in collection mode to obtain baselines and then in enforcement mode where they have to demonstrate reduction in speeding. California mandates minimum yellow times.

Flaws exist but we can constraint-engineer to minimize adversarial optimization of parameters. Improvements allow us to asymptotically approach our goal optimum. Human discretion, on the other hand, is fluid and dodges all attempts at enforcement.

We require many human agents solely because transforming unstructured vision space into structured judgment was impossible in the past. Currently, in a sufficiently constrained environment (roadways) we are able to solve sufficiently constrained problems (speeding, running a red light).

One hopes that we are able to solve this problem in more spaces as time goes by.