Hacker News Reader

Comment by tptacek | Hacker News Reader

I think the whole point of the article is that the strengths and weaknesses of human and AI drivers are different, and that the differences are exploitable by adversaries, and may in fact be so easily exploitable that "adversaries" could include bored teenagers.

View on HN

Bored teenagers can (and sometimes do) "exploit" real drivers. I've seen kids throwing rocks, and bricks off of overpasses, walking on the sidewalk lunge as if they are about to jump into the road as a joke, etc... I've also seen people slam on their brakes and/or swerve into other lanes to avoid empty plastic shopping bags blowing around, etc...

The invisible rope prank comes to mind.[0] I think there are a whole class of pranks to which AI will not be susceptible to. I assume the LIDAR can or will be able to clearly detect no such obstruction in the roadway (perhaps even if it does exist). However there are certainly many other hacks that will consistently fool AI, at least for a while.

[0]: https://youtu.be/G_pAcIjqcuY

He goes off on a tangent on the theory of the mind and saying that a computer might think a cloud is a truck.

In order to get an AI to think a cloud is a truck, you'd have to somehow reinforce the negative numerous times, essentially brain-wash the AI into thinking a cloud was a truck. Guess what else can be brain washed?

The author holds humans in far too high of regard here. Clearly given the number of accidents and road deaths a year, humans aren't very good at intuiting others' actions.

Bored teenagers are one annoyance. A greater annoyance is where automated car behavior is exploited to the detriment of inside occupant(s).

Think Brazil and how kidnapping and ransoming are not unheard of events. Unless these cars have an override, gangs and other criminals can exploit this "juvenile prank".

I remember a documentary that described a course for business executives in Brazil on evasive driving techniques for fleeing from kidnappers†. I imagine that all of these techniques are much riskier than any of a car's ordinary expected behavior. You could view this as another high-stakes variant of the trolley problem that several articles have identified as implicit in programming a car to respond to accident risks: is your self-driving car willing to drive in dangerous ways and violate traffic laws to escape from a dangerous situation? Is it willing to try to intentionally harm an assailant in self-defense? Is it willing to allow a passenger to override safety measures to do any of these things?

† To avoid giving the wrong impression about how Brazilians see crime risks in their country, I should say that I know many Brazilians and I don't know anyone who has ever taken this kind of course. I think it's easy for media to seize on and highlight exciting risks and exciting precautions: for example, I once also heard foreign media claim that "many" people in São Paulo commute by helicopter to avoid kidnappers on the ground. But friends in São Paulo cast severe doubt on this story, saying that the ultra-rich mostly like helicopters because they're fast and prestigious. :-)

The sensor logs from the trolled AI system can show likely troll/malicious intent from the human. In case there's a dispute over who's at fault (& responsible for costs).

I will pack such a sensor log in my emergency aid kid and use it to reanimate dead people.

I'm sorry I don't understand this. I was talking in context of a human driver trying to force an accident against a machine-driven car. Is passenger death less of a risk when both drivers are human?

Reply on news.ycombinator.com