Last year in the UK, 1,695 people were killed in road traffic accidents. A further 28,100 were seriously injured, and 106,207 injured in more minor ways.
Although our roads are considerably safer than they were just a few decades ago – and far safer than those in countries like Portugal (twice our rate of road fatalities), the US (three times higher) and Saudi Arabia (15 times higher) – they’re still a pretty dangerous place.
Those are some figures to bear in mind when thinking about self-driving vehicles, which were mentioned in the King’s Speech. The Government plans to allow mostly-autonomous vehicles like buses and delivery vans to operate on British roads by the end of the decade, via an Automated Vehicles Bill.
A lot of people worry about being run over by an unfeeling, uncaring robot-run vehicle. But self-driving vehicles could hardly have a worse record than that of human drivers, whose errors cause untold misery on the roads every year.
When I say “mostly-autonomous”, I’m referring to the fact that these vehicles would be a level 4 on the Automation Scale produced by the Society of Automotive Engineers, which goes from 0 (a fully manually-operated vehicle) to 5 (a fully autonomous one with no input from humans whatsoever).
A lot of vehicles you might’ve been in or driven are already at levels 1 to 3, which involves aspects like cruise control, lane centering, and automatic emergency braking. Those are features that make driving easier, but don’t take control away from humans so don’t count as “self-driving”.
Why aren’t we at level 4 or 5 yet? Elon Musk, in his capacity as the owner of electric vehicle company Tesla, has been promising that fully self-driving vehicles will be on the roads “next year” every year since 2014, and we still haven’t seen them.
That’s because the technology turned out to be a lot tricker to develop than first imagined. As any driver will tell you, the road can be a very unpredictable place.
Getting self-driving vehicles to react appropriately to pedestrians, cyclists, or animals suddenly moving into the road, while at the same time navigating complex junctions, roundabouts, inclement weather, emergencies, and unexpected changes to the road layout, hasn’t been easy.
In some cases, an abundance of caution and a fear of causing accidents means that self-driving vehicles have gone too far in the opposite direction, shutting down at the slightest hint of trouble, which can itself cause accidents on the road.
The tech writer Timothy Lee recently noted a case where a self-driving car “got rear-ended after it slammed on its brakes to avoid hitting a bag blowing in the wind”. It shows how the AI technology to operate driverless vehicles in the “Goldilocks zone” – where they run efficiently but also don’t run anyone over – isn’t yet there.
Lee looked at dozens of such accident reports from two of the self-driving car companies, Waymo and Cruise. Waymo, he estimated, had a crash for every 60,000 miles driven. In comparison, human drivers in the US (by one estimate at least) have an accident every 165,000 miles. On those numbers, humans look much better.
However, for “major” crashes, Waymo’s rate appeared to be lower than that for humans: Lee estimates that human drivers have around eight serious crashes in two million miles (the total number of miles Waymo’s self-driving cars have driven to date), whereas Waymo had only three or four. “The evidence for better-than-human performance”, Lee says, “is starting to pile up”.
That, it should be said, is under fairly limited circumstances. The self-driving cars aren’t doing the exact same routes humans do, so it’s not a perfect like-for-like comparison. Nevertheless, some of the uses are impressive: earlier this year, it was reported that Waymo and Cruise were having success in running trials of driverless cars in some US cities, for example driverless taxis taking people home from Phoenix airport in Arizona.
This time a year ago, the challenge of creating an AI model that produced useful text, solved complex problems, and usefully contributed to all kinds of jobs seemed very far off, too. And then, with the advent of ChatGPT and other AI models, it suddenly happened.
It’s perfectly possible that, like chatbots, the AI models required to run truly self-driving vehicles will soon get over the threshold where they become useful, rather than mere curiosities. The discussion, though, is now one about incremental progress, and is a far cry from the wildly overhyped pronouncements from Musk and others over the past 10 years.
In terms of making that technology safer, it’s great news that the UK Government intends to make the necessary legal changes to allow more self-driving vehicles. Simulations and tests in highly-controlled environments are all very well, but you can’t make the vehicles safer, or even gather the data required to make them safer, if you aren’t able to test them in the real world.
It’ll be interesting to watch the reaction. In my view self-driving vehicles are very likely to end up safer than human drivers, especially with AI advancing as quickly as it has been in the past year.
But it might only take a few well-publicised, caught-on-camera instances of a horrible accident involving a self-driving vehicle on British roads to turn public opinion against them. It’s already the case that videos of weird self-driving behaviour in road trials regularly go viral, and Uber sold off its self-driving car unit in 2020 after a very prominent and tragic fatality.
But we should always remember what we’re comparing them to: although self-driving vehicles are never going to be 100 per cent safe, human drivers, prone as we are to distraction, tiredness, intoxication, and all sorts of cognitive slipups, are hardly perfect either.
