A Tesla FSD Beta tester has conducted a Turing-like test with version 2020.48.35.7 of FSD. The tester, known as AI DRIVR on YouTube asked his wife to tail his Tesla on 5 routes. At the end of each route, she had to guess if it was her husband or FSD in control of the vehicle.
FSD did well but gave itself away on a handful of occasions:
7:49 – There was some weaving between parked cars.
18:25 – There was a minor swerve around a parked car, but FSD probably got away with that one.
20:50 – FSD had trouble with stopping at the right place before an intersection.
However there was a real nice, human-like right hand turn at 29:23.
The final score is at 30:15
This exercise throws up some interesting questions. Is the aim of FSD to fool other human drivers? Probably not, but it is a bi-product of a well-functioning self-driving system. Should we let other human drivers know that a machine is operating the vehicle next to them? Would humans be more tolerant of self-driving vehicles? Some might because they know they are being recorded from several angles.
On a note about the test itself, it became a reverse-Turing test when the human was in control as he wanted to ‘fool’ his wife into thinking he was a machine (10:25 in the video).
We expect to see many more exercises like this. They may not be very scientific, but they are interesting nonetheless.