Annie Turner explains why AI must not only reflect the world as perceived by those who create it if transport is going to become for everybody in the era of automated vehicles.
Artificial Intelligence (AI) is foundational to automation, which in turn underpins IoT. The volume, velocity and variety of data that will be generated by all kinds of transport, but most obviously connected and automated cars, will be far beyond the capabilities of humans to deal with.
Many people are rightly concerned about the misuse of AI – deliberate or otherwise – which is why the European Union has published its AI Ethics Guidelines on Monday. One area that really scares me is the deadly potential for unconscious bias against women – the automotive industry is more than a century old, but vehicles are far more dangerous for women than men after all this time.
McKinsey and Cisco reckon that about 250 million cars will be connected to the internet by 2020, providing new features from smart sensors to big-data enhanced geo-analytical capabilities, and connected car services will be worth about $40 billion a year.
Preventing accidents is one of the biggest motivations for connected and automated vehicles, but consider this*, although men are more likely to be involved in a car crash than women, when involved in a car accident:
- women are 47% more likely to be seriously injured;
- women are 71% more likely to be moderately injured; and
- women are 17% more likely to die.
It’s all about how cars are designed and for whom. For instance, women tend to sit much further forward than men because they are typically shorter. They have to be physically closer to the pedals and to see over the dashboard.
The car industry describes this as sitting “out of position” – that is, women sit differently from the default position designed for men. This means women tend to suffer worse internal injuries in frontal collisions, and the angle of women’s knees and hips are different too, making their legs more vulnerable.
In rear-end collisions, women are more susceptible to whiplash due to having about a third as much muscle in their necks as men and being generally lighter than men, they are thrown forward with greater force. This is exacerbated by seats in modern cars, which are firm to support men’s weight, but they don’t help absorb the shock of an accident for lighter women.
There is no suggestion that any of this is deliberate: years ago, I had a tall male driving instructor who could not grasp that it was impossible for me with my British size 5 (European 38, US size 7) feet to keep my heel on the floor and depress the clutch because he had size 11 feet.
Why does all this matter?
There is a great danger that such mistakes could be transposed to, multiplied and even amplified in the AI used to control multiple aspects of all modes of transport. This is because it is inevitable that AI will be built around how those doing the building perceive the world.
An illustration of this is that AI is good at facial recognition of white men, but poor at recognising women of any colour and non-white men: as an article in Wired last August noted, only 12% of “leading Machine Learning researchers” are women. The European Commission is doing well with 22 women out of the group of 52 experts it has assembled to advise on AI, but the group is overwhelmingly white.
The travel environment controlled by AI will be far more complex from a tech point of view than the one we inhabit today. This means the relationship between lots of things and events will be very much harder to predict or comprehend, meaning a series of biased, algorithm-based assumptions could be compounded: after all, a crash is not a single impact, but a rapid sequence of events.
Another factor is that AI will be deployed inside and outside of vehicles based on what they are like now – just like the earliest cars looked like horse-drawn carriages without the horses – but automated vehicles will change beyond recognition in time. The AI will need be in step with these changes and how humans behave when vehicles turn into pods that provide a driverless mobile sitting room or office or hotel.
We are still improving how AI combines and interprets visual, kinetic and geo-spatial information provided by sensors and other devices in terms of accuracy, speed and consequent action. While autopilot cannot foresee all circumstances and prevent all accidents, failures should be a tiny fraction of those they avert unnoticed.
After a fatal accident last year in a Tesla car, the company stated, “There are about 1.25 million automotive deaths worldwide [annually]. If the current safety level of a Tesla vehicle were to be applied, it would mean about 900,000 lives saved per year.”
We need to make sure that everyone benefits.
* Taken from Invisible Women: Exposing Data Bias in World designed for Men, by Caroline Criado Perez, published April 2019.
The author of this blog is Annie Turner, editor of IoT Now Transport