[Veserv]

The most damning thing is that the most advanced version, with the most modern hardware, with perfectly maintained vehicles, running in a pre-trained geofence that is pre-selected to work well [1] with trained, professional safety drivers, with scrutinized data and reporting average a upper bound of 40,000 miles per collision (assuming the mileage numbers were not puffery [3]).

Yet somehow they claim that old versions, using old hardware, on arbitrary roads, using untrained customers as safety drivers somehow average 2.9 million miles per collision in non-highway environments [2], a ~72.5x difference in collision frequency, and 5.1 million miles per collision in all environments, a ~175x(!) difference in collision frequency, when their reporting and data are not scrutinized.

I guess their most advanced software and hardware and professional safety drivers just make it 175x more dangerous.

[1] https://techcrunch.com/2025/05/20/musk-says-teslas-self-driv...

[2] https://www.tesla.com/fsd/safety

[3] https://www.forbes.com/sites/alanohnsman/2025/08/20/elon-mus...

[3.a] Tesla own attorneys have argued that statements by Tesla executives are such nonsense that no reasonable person would believe them.

[tim333]

That's not really like for like. You are comparing "level 4" where the car is supposed to do everything to driver assist where the driver is supposed to take over if things go off track.

I'm not sure what the guys in the taxis with their hands on the arm rest do. I guess they have a button that either stops the car or connects it to a remote control operator?

[randoglando]

It's because of selection bias. In the older vehicles, customers won't turn on autopilot if they think it won't handle the situation. So, they turn it on highways and easier paths.

[AnthonyMouse]

There is also another possibility.

People often don't report minor accidents. Someone scrapes a pole without causing enough damage to hit their insurance deductible, are they going to file a police report? Mostly not. And then the older number had that in it and the newer one doesn't.

But the number for human drivers works like the old number. They're dividing miles driven by reported accidents. On top of that, they're using the average -- by miles -- which isn't the same as the median, and in particular that will over-represent drivers who drive the most miles, who are disproportionately professional drivers.

[fragmede]

Let's be real. The real bar is drunk humans. I can scream and yell and not be friends with anybody who would put other people in danger like that, and they'll still drive drunk. FSD is good enough that it'll detect the driver's fallen asleep and pull over and park the car. Tesla can't talk about that for obvious legal reasons, but that's already saved many lives. Unfortunately we can't know those stats for comparison and holy shit people need to not drive drunk, but DUI laws don't cure addiction.