I love reading The Morning Paper where @adriancolyer translates computer science papers into English. With my keen interest in automotive software I took a special note of this excellent paper.
Only after deciding to write about this fascinating work did I notice that this was from my friend, the brilliant and witty Kirill Levchenko at sysnet.ucsd.edu.
I don’t have anything techincal to add to this fine work, but reading TMP’s take on it, I have to agree that the endgame here is one of two things. First, there could indeed be an AI race to constantly game the system. Currently there are some fairly crude, but entirely effective, heuristics to determine when the system is being watched (tested). But one could imagine a more sophisticated testing regime that required the wheels be turned randomly and the speed varied. An AI classifier could be conceived of which would still endeavor to cheat, and the race would be on.
The other option, and this seems like the just approach, is to not test the car occasionally (using extrapolation to find out how much it likely pollutes). No, the correct approach is to mandate full time pollution accounting. Your car should have a "measured polution" counter that just tallies it up and you get sent a bill accordingly.
The next topic, of course, would be cheating the pollution sensors. Don’t worry, there could be montoring sensors to ensure the correct functioning of the pollution sensors. And other turtles.