In the long run, I think it's unavoidable that AI-type systems continue to improve, while humans don't, so this will become a harder and harder problem. One he

In the long run, I think it's unavoidable that AI-type systems continue to improve, while humans don't, so this will become a harder and harder problem. One helpful approach would be to separate out "why CAPTCHA" into preventing abuse (through high volumes) and "guaranteed one (or small number) per person" from "am I interacting directly with a live human", and using different things for each. The naive solution to a lot of this is identity -- if FB profiles are "expensive" to create, especially old ones with lots of social proof, you can use something like FB connect. However, there are a lot of downsides to this (chief being centralization/commercial control by one entity, which might be a direct competitor; secondarily, loss of anonymity overall.) One interesting approach might be some kind of bond -- ideally with a ZK proof of ownership/control, and where the bond amount is at risk in the case of abuse, but it's not linked to identity.