I am skeptical of claims that autonomous vehicles would decrease congestion even without taking into account this paradox (i.e., even if the density of cars on the road remained the same). All the claims seem to reference the same study [1], but per the abstract they are just simulating one circular lane of traffic. No "lane changes, bottlenecks, merges, or changes in grade". Intuitively, it seems like once you add all these other factors in, autonomous vehicles won't help much unless deployed on a wide scale; for example, it doesn't seem like one driver can do much to speed up the merging of cars ahead of it.