Comment by yshklarov

That observed fatality rate, 2, is too low to give a precise estimate of the mean. If we assume fatalities follow a Poisson process, then it's plausible (at 5% significance level) that the true expectation is only 0.36 deaths, not 2.

So it's more useful to look at serious injury statistics as a proxy. For those, Waymo says that there is a significant reduction.