The research I've seen is that for areas with good conditions year round they're currently safer than a subset of the population (those in at least one automotive accident in the past 5 years, or something like that).
Still not better than your average driver, but even if we start by replacing alcoholics and routinely poor drivers the net impact on safety goes up.
The proper comparison to be made should be between miles driven from autonomous vehicles vs. those of humans and then looking at the incident rates.