Ideally Tesla's marketing would say it's safer than drunk, sleepy, and reckless drivers, though it might not sell as many Autopilots.
I also imagine that cars will learn from both human and each others' driving patterns over time, which (under effective guidance) should enable an enormous improvement in a relatively short period of time.
However, I take the other interpretation. I don't think crashes are randomly distributed. That that scenario is absolutely perfect is really the problem because of humans. All a crash takes is a second of attention lapse at a really bad time. And in such perfect circumstances we get bored and take road safety for granted as opposed to driving at night or navigating around some tricky curves. And that's a perfect scenario to end up getting yourself killed in. And more importantly here, that's also the absolutely perfect scenario for self driving vehicles who will drive under such conditions far better than humans simply because it's extremely trivial, but they will never suffer for boredome or lack of attention.
It means this will kill people but fewer people than humans would and they have actual data that backs this assertion up.
The benchmark is not an alert, cautious, and sober driver because that's often not how actual people drive. So right now it's often safer to drive other times it really is safer to use autopilot, net result fewer people die.