The first case of a pedestrian death caused by a self-driving vehicle has provoked an understandably strong reaction. Are we witnessing the questioning of the Emperor’s new clothes? Have we been living through the latest and greatest wave of hype around AI and its performance? Will self-driving cars come off the road for a generation, or even forever? Or, on the contrary, will we quickly come to accept that fatal accidents are unavoidable, as we have done for so long in the case of human drivers, after all?

It does seem to me that the dialogue around self-driving cars has been a bit unsatisfactory to date. There’s been a surprising amount of discussion about the supposed ethical issues; should the car save the lives of the occupants if doing so involves killing a greater number of bystanders? I think some people have spent too long on trolley problems; these situations never really come up in practice, and ‘try not to have an accident at all’ is probably a perfectly adequate strategy.

More remarkable is the way the cars have been allowed to move on to public roads quite quickly. No doubt the desire of the relevant authorities in various places to stay ahead of the technological curve has something to do with this. But so far as I know there has been almost no sceptical examination of the technology by genuinely impartial observers. The designers have generally retained quite tight control, and by and large their claims have been accepted rather uncritically. I’ve pointed out before that the idea that self-driving machines are safer than humans sits oddly with the fact that current versions all use the human driver as the emergency fall-back. There may in fact have been some tendency to treat the human as a handy repository for all blame; if they don’t intervene, they should have done; if they do intervene, then any accident is their responsibility because the AI was not in control at the moment of disaster.

Our confidence in autonomous vehicles is the stranger because it is quite well known that AI tends to have problems dealing with the unrestricted and unformalised domain of the real world. In the controlled environment of rail systems, AI works fine, and even in the more demanding case of aviation, autopilots have an excellent record – although a plane is out in the world, it normally only deals with predictable conditions and carefully designed runways. To a degree roads can be considered similarly standardised and predictable, of course, but not all roads and certainly not the human beings that frequent them.

It can be argued that AI does not need human levels of understanding to function well; machine translation now turns in a useful performance without even attempting to fathom what the words are about, after all. But even now it has a significant failure rate, and while an egregious mistranslation here and there probably does little harm, a driving mistake is another matter.

Do we therefore need a driving test for AI? Should autonomous vehicles be put through rigorous tests designed to exploit likely weaknesses and administered by neutral or even hostile examiners? I would have thought something like that would be a natural requirement.

The problem might be whether effective tests are possible. Human drivers have recognisable patterns of failure that can be addressed with more training. That may or may not be the case with AIs. I don’t know how advanced the recognition technology being used actually is, but we know that some of the best available systems can behave in ways that are weirdly unpredictable to human beings, with a few pixels in an image exerting unexpected influence. It might be very difficult to test an AI in ways that give good assurance that performance will degrade, if at all, only gradually and manageably, rather than suddenly and catastrophically.

In this context, footage of the recent accident is disturbing. The car simply ploughs right into a clearly visible pedestrian wheeling a bike across the road. It’s hard to see how this can be explained or how it can be consistent with a predictably safe system. I hope we’re not just going to be told it was the fault of the human co-pilot (who reacted too slowly, but surely isn’t supposed to have to be ready for an emergency stop at any moment?) –  or worse, of the victim.


Original Content Source