There are so many myths about autonomous driving and safety in the public debate. We're going to look at four of them here from a US perspective.
January 16, 2025
© iStock/Liliya Filakhtova
The US Department of Transport says that “human error is a factor in 94 percent of all fatal crashes”. Quite regularly this finding is inappropriately modified to “94 percent of fatal crashes are due to human error.” This is then used to imply: autonomous vehicles (AV) will be almost 20 times safer (only suffer from that remaining 6 percent of crashes) if we get rid of those oh-so-unreliable human drivers.
To be sure, impaired and distracted human driving are problems. But once we take a closer look at what is behind those numbers, the picture gets much clearer. Sometimes the driver does cause the crash. But many other times something goes wrong, and the driver is faulted for fail to avoid a crash caused by that something else. If the driver fails to avoid a theoretically avoidable crash, then the driver’s imperfection is counted in that 94 percent. This is not at all the same as the driver having caused the crash, and this number does nothing to prove that AVs will necessarily do better. AVs will have to similarly mitigate risk from external factors. And we should not forget that AVs will make mistakes in judgement that lead to at-fault crashes as well.
So the question is: Can an AV really do better? The answer so far is they will potentially be better – but nobody knows how it will turn out. The question that matters is when AVs will be at least as good as human drivers at compensating for problems that arise during driving, while making fewer driving mistakes themselves. One might say that robots won’t drive drunk, but they lack common sense.
Anyone invoking the 94 percent statistic as a compelling safety reason to adopt AVs is pumping the technology rather than encouraging an informed discussion about safety.
Autonomous vehicle companies regularly argue against requiring them to conform to existing industry safety standards. But on closer inspection, these objections to ISO 26262, ISO 21448, and ANSI/UL 4600 vanish into thin air. I would like to take a closer look at three of the rhetorical objections against using safety standards:
“Not a perfect fit” amounts to saying that only a bespoke safety standard for a particular company’s technology can be considered instead of getting a safety standard “off the rack” so to speak. But the use of the word “perfect” here is typically disingenuous. A standard that is a 99.999 percent fit is not a 100 percent fit, and therefore not “perfect.” The question should be whether safety standards are usable in practice. Moreover, all three safety standards mentioned offer significant flexibility and permit conformance strategies to be adjusted to suit if a reason is documented why that makes sense.
No “single” standard does everything required for safety because the standards are purposefully written to cover different parts and work together as a team. The industry teams who write these standards have liaisons to avoid overlap as well as shared voting committee members across ISO 26262, ISO 21448, ANSI/UL 4600, and other relevant standards. One single standard for everything in an AV would simply be too big a chunk to bite off at once. Given the existence of pre-AV automotive safety standards, creating a one-stop-shopping AV standard would violate standards organization operating procedures due to the impermissible overlap between standards that would result.
The statement that safety standards were not written “specifically” for AVs is plainly incorrect and beside the point. What matters is which standards get the job done, not whether they also might help make non-autonomous vehicles safe. The scope of ISO 26262 is for all vehicles, including relevant aspects of AVs. The initial version of ISO 21448 was just for driver assistance systems, but the currently issued version explicitly covers autonomous vehicles. Without question, ANSI/UL 4600 has been specific to autonomous vehicles from the start. Other standards specifically intended to cover AV-related technology, such as ISO 8800 are also becoming publicly available. Useful, highly relevant international industry consensus standards are there. The problem is getting AV companies to conform to them.
AV companies loudly proclaim that safety is number one and that their vehicles are or will be safer than human vehicles. They justify public road testing because they say they need to collect and analyze data from that testing to attain the promised safety. But, many companies treat everything about safety while they are doing that testing as a secret. Later they might publish selected statistics about how things turned out after the fact. But they might only publish numbers that make them look good.
From a public policy point of view, states and municipalities are providing AV testers and for-profit AV operators with free access to a shared public resource in the form of public roads and human test subject road users. Given that AV technology is still immature, AV companies are planning to profit massively from access to this public resource, while putting other road users at potentially elevated risk. In exchange for the privilege of accessing that shared public resource, they should have an obligation to provide data demonstrating they are not subjecting the general public to undue risk, especially while the technology is still maturing.
The important thing is that metrics should be relevant to public safety rather than technology maturity. For example, reporting how often an AV blocks fire trucks proceeding to an emergency, or how often an AV runs red traffic signals has a compelling public interest, but says nothing about the details of the autonomy technology.
But if an AV test platform regularly suffers near hits, blocks emergency equipment or violates traffic laws the company has a vested interest in suppressing the information to avoid embarrassment. The proprietary secret ploy is used more to hide public road operational problems than it is to protect legitimate secrets about the inner workings of autonomous vehicle technology.
While we hope it will eventually be true, there is not yet any proof that AVs based on current technology will ever be safer than human drivers for severe injuries and fatalities – which is the core of the industry safety promise. Any harm now is weighed against only the hope of future benefit, not the certainty of future benefit.
The safety benefits of AVs are aspirational and targeted for some time in the future. Every year, that time seems to get further away. Given the track record of promises and delays, nobody really knows how far in the future.
Moreover, there is no real proof to show that AVs will ever be safer than human-driven vehicles, especially with human-driven vehicles becoming ever safer themselves over time due to the addition of passive and active safety technology. The relevant basis of comparison is a new robotaxi against a new conventional car equipped with the latest safety systems – not a new robotaxi compared against an “average” 10+ year old car with previous-generation safety technology.
It seems reasonable to expect AVs will reach a point where they can be acceptably safe in appropriately restricted operational design domains (ODDs). But the “when” still remains a question mark, as does whether the net safety benefit will be compelling. Or there will be a net safety benefit at all.
It is inappropriate to use an aspirational goal of perhaps someday saving lives as an excuse to do real harm to real people in the near term by cutting safety corners while scrambling to mature the technology.
The author, Prof. Dr. Philipp Koopman, is Deputy Chairman of the advisory board of the Fraunhofer Institute for Cognitive Systems IKS.
