We’ve all seen technology proceed like greased lightning.
In my lifetime, we’ve gone from typewriters to internet-enabled laptops. We’ve seen smartphones go berserk and enormous progress in survival rates for cancer. These fields have transformed. It is tempting to predict more exponential change in the field you’re most excited by.
But caution is needed.
The fields in which we see progress are affected by survival bias. We don’t see the frustrated scientists trying and failing to revolutionise other fields. Look around you and much is as it was 100 years ago. I’m sitting at a wooden chair at a wooden table, wearing woolen socks and leather shoes. The alphabet is the same as it was, and so is my keyboard layout. There’s a clock on the wall telling me the time with two rotating hands. I just got over a common cold. I’m eating brown rice and snowpeas. It could be 1850 – if not for the macbook.
So not everything is on the brink of revolution. Which is why I have to pull back on my former enthusiasm for autonomous cars. I admit I was focused on the potential upsides – in traffic, in accidents, in parking, and on the successes Google has had with its autonomous car program. Google is backing the project, appointing the old head of Ford. But even Google fails sometimes, as with Google Wave.
“The benefits are so great that we will force ourselves to accept them, even with a few risks,” I told myself.
But then I started thinking about the development path, and I became significantly cooler on the chances of success.
Autonomous cars will only break through once they are trusted.
Humans set a very high bar for risk in situations where they perceive they are not in control. (This is why people object to tiny risks of living downwind of a polluter and won’t let their kids walk to school, but still eat chips and drive fast.)
Autonomous cars won’t just have to prove they are safer than humans at driving, but much safer – for car occupants, other road users, pedestrians, wildlife and pets.
THE LONG TAIL
Computer operated cars are probably already better than humans at driving in car traffic on freeways and on busy roads. Humans are dreadful at mundane repetitive tasks that require paying attention.
But car crashes can happen in odd moments.
This is where humans excel. We dominate computers at dealing with problems we never saw before.
Humans will remain best at dealing with things like:
- A big black garbage bag blows onto the road but we know we needn’t swerve as we can tell it is light by the way it moves.
- Kangaroos are on the side of the road so we better slow down because they often jump in front of the car.
- It’s Saturday afternoon, there’s just been a football match, some sort of fight is happening on the side of the road, and you know someone could easily step out into traffic as part of the brawl.
Many serious crashes occur in scenarios that are in the long tail of distributions. Machine learning will not cover them all, so there will remain a few scenarios (I predict on the basis of statistics alone) in which autonomous cars continue to perform predictably worse than humans despite the best efforts of programmers.
Other types of software can launch with “beta phases” where failure is embarrassing, but not catastrophic. But the testing that will have to happen before any serious real world traffic experiments involving autonomous cars will be enormous. Google’s experiments driving round California are good, but still limited in scope and scale.
A few high-profile crashes will be enough to set a very high technical and legal bar for autonomous cars. The concept of surrendering ones life to a machine is a staple of science fiction because it irritates a real issue in human psychology – control.
MANY OPPORTUNITIES FOR SETBACK
It is not just technical problems that can hold up autonomous cars indefinitely. Political, road engineering, PR and software challenges will impede getting autonomous cars to the point where people trust them and forgive their mistakes.
For just one example, the FBI is opposed to driverless cars, according to a brand new report. Solving that will be tricky. And when it is solved another impediment will arise.
There’s a lot of fail points. I suspect – again on the basis of pure statistics – one will resolve into a big sticking point for a long time to come.
SUCCESS WON’T LOOK LIKE SUCCESS
Cars will continue to have more and more sensors and autonomous capabilities. But during this time, non-autonomous cars will continue to be sold.
Traffic will be mixed for at least the next 50 years. Some freeways and highways will perhaps be autonomous-only. But not places where there are pedestrians, bicycles, shops, parking, and of course traffic lights. So the benefits of full autonomy will not be realised for a very long time. Don’t hold your breath.
The upside of the failure of the fields about which we are most excited is that we might get blindsided by a revolution in a field where we didn’t expect any improvement. Nano technology, GM foods, high-speed trains, smell-o-vision: any of these could be the one in which a breakthrough happens that turns out incredibly positive.