By Sophie Dickins
An Intelligence We No Longer Understand
Artificial intelligence is no longer the stuff of science fiction. While early prototypes struggled to form a sentence or beat a beginner at chess, today’s AI is highly sophisticated. In 2017, artificial intelligence can defeat the world’s best human chess player and write articles fit for print. In fact, the Washington Post's “robot journalist” has published more than 850 articles this year. AI has now proved complex mathematical theorems, its answers published in prestigious scientific journals. There’s just one issue - no mathematician in the world can explain why its answers are correct. They know AI is right, but they can’t say how.
The most sophisticated forms of AI are known as “black box” artificial intelligence. It is now possible to give a computer program a question, feed it huge amounts of data, and receive an answer which we know is correct, but cannot fully understand. Much like a human brain, AI has a complex neural network whose workings are hidden and constantly evolving. Because of this neural network, AI can come up with its own answers without allowing us to see its brain - that is, the complex factors and calculations that allowed it to reach its conclusion. These calculations are too complex for human minds to understand.
This isn’t particularly important when it comes to abstract mathematical proofs, but AI is now being used in situations that previously required human judgment, and its proposed solutions are making many people uncomfortable. For example, artificial intelligence has already made its way into the justice system. An AI was taught to calculate the likelihood of an inmate violating his parole conditions, and was subsequently tasked with applying its newfound knowledge. The AI was asked to calculate the relative probability of 500,000 paroles violating the terms of their paroles using all available data on them. This meant that the artificial intelligence program decided the relative weights that should be placed on dozens of different variables by employing an algorithm that generated the probability to be used. The parole AI might calculate that a 51-year-old man with no prior offenses has a 12% chance of violating his parole, whereas a 22-year-old man who has already been convicted three times has an 88% chance. Thus, parole would be granted to the first man but not the second. This is simply a more mathematically sound alternative to the parole system. Instead of a human being deciding how likely a person is to re-offend based on limited information, an AI can draw on hundreds of thousands of cases to reinforce its decision.
The problem with this is that when AI was used in this exact scenario, it assigned a weight to race. It found that it was statistically more likely for a black man to violate parole than a white man, and therefore recommended granting black men parole less often. This is a form of discrimination that is currently illegal.
Is A Consistent Bias Preferable?
AI may base its decisions on factors that are out of one’s control, but human beings do the same. Attractive people get shorter sentences than unattractive people. Judges who are hungry give harsher sentences than those who just ate lunch. The human mind is inconsistent in a large number of ways, many of which are due to our underlying biology. Unlike AI, we cannot be impartial, coldly calculating machines. But we might actually be worse.
AI consistently assigns probabilities to events based on hundreds of thousands of past occurrences. A human judge decides to release someone based on their perception of that individual - their feelings towards that specific person, their appearance, their race or their behaviour. Both AI and humans assign importance to things that people cannot fully control. At least AI is consistent, whereas humans are affected by multiple cognitive biases.
A More Equal World, Or A More Accurate One?
For nearly a century, our society has moved towards equality through the elimination of biases and discrimination. Under current laws, women must be paid the same amount as men. People of all races, genders and sexualities must be treated equally, and one cannot be treated preferentially on the basis of their age.
However, it is possible to find legal exceptions to this rule. For example, car insurance rates consider factors that are not under a person’s control: their age and their gender. A male driver between the ages of 18 and 25 might pay car insurance premiums twice as high as a woman in her forties. It is clear that some young men who are responsible drivers are being unfairly punished by being lumped together with the more irresponsible members of their age group. But statistically, they are a lot more likely to drive over the speed limit, drive intoxicated, and crash their cars. So all young men pay the price. Notice there’s no public outcry, or plead to reduce young men’s unfairly high car insurance premium - as a society we have deemed this injustice acceptable.
As AI enters more aspects of our lives, these kinds of discriminatory behaviors will become more common. Artificial intelligence makes it possible to calculate risk profiles with a higher degree of certainty and a larger sample size than ever before. But is that something we want? Is it acceptable to deny a black man parole, at least in part, because of the color of his skin?
Most would say no. But the alternative seems equally unreasonable. If we release people with higher risks of recidivism, ignoring statistical factors that point to their likelihood of reoffending, we are putting members of the community at risk. There are parallels to be drawn in many other fields. Could governments selectively tax people who eat sugary foods because they are more likely to become ill and incur healthcare costs? In the event of a crash, what if the algorithm of a self-driving car decides that it is better to preserve the life of a CEO rather than the lives of two children? Although AI might be more equitable, determining a person’s value using mathematics is difficult to swallow.
Whether or not you are comfortable with artificial intelligence determining where you work, how much things cost and even whether you are allowed a place in society, we are clearly moving in that direction.
Comments