The goal is improving safety. To improve, autopilot needn't be safer than the best driver, it only needs to be safer than the average driver. Even if adoption depends on proving autopilot is better than good humans, if it fails this test, then people just don't use it and safety is no worse than it is today.
> autopilot is a big unknown
We really just need to know whether Teslas get into more accidents or more severe accidents than other vehicles (modulo some adjusting for demographic differences among drivers). If Teslas contribute to fewer/less-severe accidents, then it demonstrates that they're safer overall. Whether they get there via autopilot or other means is an implementation detail. In any case, Tesla isn't just going to ship unsafe software because it would destroy their brand (every Tesla incident gets 1000x more scrutiny than other brands).