Hacker News Reader

Comment by scott00 | Hacker News Reader

The data analysis is not straightforward. You would expect a driver's decision to use Autopilot to be correlated with the ability of the Autopilot to handle the conditions. So just comparing Autopilot accidents per mile to human accidents per mile won't cut it.

View on HN

My suggestion is to compare Tesla accidents per mile to non-Tesla accidents per mile. If people are being lulled into a false sense of security by AP then this should show up in the data.

A similar problem exists with that analysis. Tesla ownership is probably correlated to the amount of time/miles spent in driving conditions where Autopilot is effective.

You also add a whole bunch of other confounding factors if you're not comparing same-driver accident rates. Accident rates vary by location, city/highway driving mix, driver attributes (age, gender, credit score, etc), and probably many other things that may be correlated with Tesla ownership.

If so then at least, auto pilot is not harmful, is it?

It is still harmful in the sense that it gives a false sense of safety. See those lethal accidents where drivers were totally distracted.

How do you conclude that? GP’s premise amounts to “if the overall rate of human crashes per million miles is H and the autopilot rate is A, even if A is half of H, that doesn’t tell you anything about the relative safety if autopilot is engaged during a particularly safe subset of overall driving.”

Reply on news.ycombinator.com