[sokoloff]

Right. I suspect that the net effect is that the Tesla software transforms these what would have been close calls into…still close calls where the Tesla software gets credit for a “save”.

If a Tesla “saved” a driver 5 times in 20K miles, my first question is always going to be “how many collisions did they have in the prior 20K miles in their other car?”

[staktrace]

This is a good point, and I agree. If there was a close call you can't actually say for sure if it would have been a collision without the software.

[sokoloff]

This effect has a large influence on the overall A/B analysis. If pre/post test analysis suggests that Tesla saved you from only 0 to 1x collisions vs a previously estimated 4x-5x collisions, your tolerance for newly introduced collisions would naturally be much lower.