New data released last week by the California Department of Motor Vehicles offered bright news for fans of autonomous vehicles. Based on so-called “disengagement reports” submitted by developers of the technology, self-driving cars are getting better, and they are getting better faster than before.
The reports track how many miles a vehicle travels between “disengagements” – that is, when the human operator took control away from the automated system piloting the vehicle. For Waymo—the developer formerly known as Google—there was a disengagement, on average, every 5,128 miles during the 2016 reporting period, compared to every 1,244 miles in 2015. Nissan drove far fewer miles than Waymo, but also saw its results improve dramatically. Its operators disengaged every 246 miles last year, compared to every 14 miles in 2015.
But before we start popping the Champagne corks—and particularly, before we start laying down money on which company’s stock to buy—it bears examining whether these mandatory reports tell us anything useful. Rather than giving the public extra confidence in autonomous technologies, the reports—which are ripe for misinterpretation—could actually undermine it.
What’s immediately clear is that there’s no way to compare the 11 reports on offer from the DMV on an apples-to-apples basis. Different developers test different types of technology under different conditions with different human testers who each have unique thresholds for what scenarios demand disengagement. That’s a huge amount of variation.
The big question is whether these sorts of reporting requirements create perverse disincentives that discourage firms from addressing particularly challenging operational scenarios. A company that chooses to test a large number of miles driven in urban centers, surrounded by traffic and pedestrians, likely will report more frequent disengagements than one that chooses to focus its testing on freeways.
In practice, the California DMV’s reporting requirements penalize firms that undertake tasks with a higher degree of difficulty by subjecting them to additional reputational risk. Predictably, in the wake of the disclosures, much already has been made of the “poor” performers, including calls for robust intervention in the name of safety. At this juncture, the last thing public policy should do is discourage efforts to tackle the hardest problems likely to confront self-driving technologies.
The argument for mandatory disclosure of test results rests on an assumption that touting the results will increase public confidence in what remains a radical new technology, thus helping to hasten its adoption. But that ignores the opposite risk – the degree to which ill-informed public scrutiny actually could serve to undermine that confidence.
The answer is not to more stringently regulate what kinds of reports developers have to file. Uniform reporting requirements are celebrated by bureaucrats, but inevitably advantage some developers over others. Despite the best of intentions, public choice theory tells us that where rules can be shaped to tilt the playing field, it’s usually the biggest incumbent firms that will wield the greatest influence. When it comes to an emerging technology as potentially revolutionary as autonomous vehicles, we should guard against the potential for cronyism and regulatory capture as firmly as we can.
States like Nevada and Michigan have recognized the best way to improve public confidence in self-driving vehicles is not by mandating reports, but by letting that public experience safe and effective deployed vehicles. Only in deployment will people build the concrete confidence needed to adopt self-driving technologies. The federal government—specifically the National Highway Traffic Safety Administration, which has shown great restraint to this point—should continue to do the same.
To be clear, it is fair to say the California DMV reports indicate that meaningful, if uneven, progress is being made toward our self-driving future. We should celebrate those results. But we also should be leery of expanding reporting requirements or imposing them in other jurisdictions. There is always a cost to regulation, even when it involves an apparently benign data call.
Image by Lightspring