Wouldn't it be fairly simple to run tests for awhile and give the thumbs up if the average, shall we say, carnage rate is lower than with human drivers? Seems like it would be a no-brainer.
And when there are inevitably problems (or even deaths) caused by the AI systems, it's not like car companies don't currently have liability for malfunctions in their designs. Seems it would be the same...
You obviously don't work in a highly regulated industry :-)
Things that seem like "no brainers" to software developers are absolute show stoppers to regulatory bureaucracies like the Department of Transportation.
It's not impossible and it will certainly happen eventually, but it will take years of proposals and studies for a new set of regulations to evolve that these cars will have to follow. Liability is a separate issue that the car companies are used to dealing with. Changing minds at the US DOT (for example) will be a lot harder.
And when there are inevitably problems (or even deaths) caused by the AI systems, it's not like car companies don't currently have liability for malfunctions in their designs. Seems it would be the same...