The stats from Tesla are misleading (“winner’s bias”) because they are limited to a set of scenarios where FSD can work and there is a “safety net” - a human who has to take over when the FSD simply gives up.
We have to compare FSD against humans on the same road, under the same conditions and every FSD failure is a crash, as every human failure is a crash. Unfortunately, that data is not available.
Yes, FSD is basically only available to function when its very safe conditions. Humans would have many less accidents if you only tracked the time they were driving in the same very safe conditions.