There are two ways to think about the same problem. Bot detection tries to incriminate: gather enough evidence to prove a given actor is synthetic. Human verification tries to exonerate: establish a positive signal that an actor is real. They sound symmetrical. They are not.

Incrimination is open-ended and adversarial. You are searching an unbounded space of "ways to look fake," against an opponent actively working to look real. Every new evasion expands the space you must cover. The compute cost trends upward without limit.

Incrimination compute is far greater than exoneration compute, so build the system that asks who is real, not who is fake.

Exoneration can be bounded. A strong positive proof of humanity (a verifiable, expensive-to-fake signal) closes the question rather than reopening it. Designing for verification instead of detection flips the economics from an endless chase into a finite check.