This article was cowritten with Caleb Watney, technology policy associate at the R Street Institute
Unsubstantiated driverless car hype may be annoying, but that shouldn't blind us to the real cost of unnecessarily delaying autonomous vehicle (AV) deployment.
Last week, after exploring new data from the California AV disengagement reports, Ross Marchand of the Taxpayers Protection Alliance argued that we should "put driverless cars back in the slow lane." California requires AV companies testing in the state to report each time a human operator takes over for a driverless car — an event otherwise known as a "disengagement." Marchand offers some interesting analysis, but ultimately reads far too much into a limited dataset and pushes for a restrictive policy prescription that would undermine public safety. The discussion is worth fleshing out because it reveals important limits to the "precautionary principle" mindset that is so common in AV discussions.
In 2017, Waymo — the self-driving car project formerly belonging to Google — reported driving over 350,000 miles on California roads with 63 total disengagements. Marchand claimed that, based on these data, Waymo's test vehicles are still not as safe as human drivers and that they are improving at a slower rate than those hyping AVs would have us believe. Further, he argued that until driverless cars can prove they are safer than human operators, we should keep them off public roads — and instead test them on expensive private tracks.
There are a few glaring issues with this argument. First, it overestimates how applicable and reliable the California disengagement data really are. As many commentators have pointed out, disengagement data are a poor measure of AV progress. Not only are disengagement reports an apples-to-oranges comparison across vehicle manufacturers who use different definitions, strategies and road conditions for testing, but Marchand drills down by comparing particular disengagement subcategories, leaving him with sample sizes of less than 20 — several orders-of-magnitude too small to make meaningful comparisons. Furthermore, comparing disengagements to would-be fatalities is problematic given that a safety driver's presence enables testing in regions that the vehicle is still learning to handle.
Marchand also left aside the successful testing and deployment of Waymo's fully driverless cars in Arizona. Since November 2017, hundreds of AVs have been providing free taxi services in the suburbs of Phoenix without any safety driver in the front seat. To date, there have been no reported accidents or fatalities. This suggests what we've known all along: these companies already face a host of legal, political, economic, regulatory and publicity pressures that incentivize them to prioritize safety in AV deployment. They know that every bump, scrape and crash will make headlines (regardless of who is at fault) and will slow or — if it's serious enough — completely derail their path to market. Waymo obviously feels confident enough to take its hands off the wheel, and so far has been right. Why rip AVs off the roads when no one has been harmed?
Marchand's larger argument against AV testing on public roads provided a textbook example of the precautionary principle in practice. Simply put, the precautionary principle requires innovators to prove that their new technology will not harm society, rather than placing the onus on regulators and litigators to demonstrate that an innovation actually causes harm.
And to that point, Marchand fails to specify what exactly the harm of public testing has been. Public testing has not unleashed mass fatalities on society, or even mass fender-benders. Rather, it appears to be speeding up the feedback loop of better data and more-rigorous test environments, leading to faster improvements in autonomous technology.
As a society, we can't afford to wait until we are 100-percent certain that driverless cars are statistically safer than humans before letting them on the roads. As a report from RANDhighlighted, it could take several decades to accumulate enough miles on private test courses to know beyond a shadow of a doubt that AVs are safer than their human counterparts. Relying on Marchand's precautionary principle approach would mean waiting decades while nearly 40,000 people die on our roads every year. Regulatory delay of this magnitude could, conservatively-speaking, cost tens of thousands of lives.
That's not to say private test courses don't have a role to play in AV development. Indeed, Waymo already operates an extensive test track in Arizona where operators take real-world scenarios and experiment with hundreds of possible variations. This hybrid approach combines the advantages of real-world testing and private test courses. But forcing all AV testing onto private test tracks cuts off the real-world data necessary for this complementary approach and substantially raises the barrier to entry for new competitors.
To be clear, we should avoid over-hyping the progress made in AV development. Carefully taking into account the safety data will be a key part of this effort. But halting all real-world AV deployment is a heavy-handed "solution" desperately in search of a problem.