Strategy. Innovation. Brand.

artificial intelligence

Are Machines Better Than Judges?

No bail for you.

Police make about 10 million arrests every year in the United States. In many cases, a judge must then make a jail or bail decision. Should the person be jailed until the trial or can he or she be released on bail? The judge considers several factors and predicts how the person will behave. There are several relevant outcomes if the person is released:

  1. He or she will show up for trial and will not commit a crime in the interim.
  2. He or she will commit a crime prior to the trial.
  3. He or she will not show up for the trial.

A person in Category 1 should be released. People in Categories 2 and 3 should be jailed. Two possible error types exist:

Type 1 – a person who should be released is jailed.

Type 2 – a person who should be jailed is released.

Jail, bail, and criminal records are public information and researchers can massively aggregate them. Jon Kleinberg, a professor of computer science at Cornell, and his colleagues did exactly that and produced a National Bureau of Economic Research Working Paper earlier this year.

Kleinberg and his colleagues asked an intriguing question: Could a machine-learning algorithm, using the same information available to judges, reach different decisions than the human judges and reduce either Type 1 or Type 2 errors or both?

The simple answer: yes, a machine can do better.

Klein and his colleagues first studied 758,027 defendants arrested in New York City between 2008 and 2013. The researchers developed an algorithm and used it to decide which defendants should be jailed and which should be bailed. There are several different questions here:

  1. Would the algorithm make different decisions than the judges?
  2. Would the decisions provide societal benefits by either:
    1. Reducing crimes committed by people who were erroneously released?
    2. Reducing the number of people held in jails unnecessarily?

The answer to the first question is very clear: the algorithm produced decisions that varied in important ways from those that the judges actually made.

The algorithm also produced significant societal benefits. If we wanted to hold the crime rate the same, we need only have jailed 48.2% of the people who were actually jailed. In other words, 51.8% of those jailed could have been released without committing additional crimes. On the other hand, if we kept the number of people in jail the same – but changed the mix of who was jailed and who was bailed – the algorithm could reduce the number of crimes committed by those on bail by 75.8%.

The researchers replicated the study using nationwide data on 151,461 felons arrested between 1990 and 2009 in 40 urban counties scattered around the country. For this dataset, “… the algorithm could reduce crime by 18.8% holding the release rate constant, or holding the crime rate constant, the algorithm could jail 24.5% fewer people.”

Given the variables examined, the algorithm appears to make better decisions, with better societal outcomes. But what if the judges are acting on other variables as well? What if, for instance, the judges are considering racial information and aiming to reduce racial inequality? The algorithm would not be as attractive if it reduced crime but also exacerbated racial inequality. The researchers studied this possibility and found that the algorithm actually produces better racial equity. Most observers would consider this an additional societal benefit.

Similarly, the judges may have aimed to reduce specific types of crime – like murder or rape – while de-emphasizing less violent crime. Perhaps the algorithm reduces overall crime but increases violent crime. The researchers probed this question and, again, the results were negative. The algorithm did a better job of reducing all crimes, including very violent crimes.

What’s it all mean? For very structured predictions with clearly defined outcomes, an algorithm produced by machine learning can produce decisions that reduce both Type I and Type II errors as compared to decisions made by human judges.

Does this mean that machine algorithms are better than human judges? At this point, all we can say is that algorithms produce better results only when judges make predictions in very bounded circumstances. As the researchers point out, most decisions that judges make do not fit this description. For instance, judges regularly make sentencing decisions, which are far less clear-cut than bail decisions. To date, machine-learning algorithms are not sufficient to improve on these kinds of decisions.

 

(This article is based on NBER Working Paper 23180, “Human Decisions and Machine Predictions”, published in February 2017. The working paper is available here and here. It is copyrighted by its authors, Jon Kleinberg, Himabindu Lakkaraju, Jure Lesovec, Jens Ludwig, and Sendhil Mullainathan. The paper was also published, in somewhat modified form, as “Human Decisions and Machine Predictions” in The Quarterly Journal Of Economics on 26 August 2017. The paper is behind a pay wall but the abstract is available here).

Jevons Paradox and The Future of Employment

My new teaching assistant.

My new teaching assistant.

Four years ago, I wrote a somewhat pessimistic article about Jevons paradox. A 19th-century British economist, William Jevons, noted that as energy-efficient innovations are developed and deployed, energy consumption goes up rather than down. The reason: as energy grows cheaper, we use more of it. We find more and more places to apply energy-consuming devices.

Three years ago, I wrote a somewhat pessimistic article about the future of employment. I argued that smart machines would either: 1) augment knowledge workers, making them much more productive, or; 2) replace knowledge workers altogether. Either way, we would need far fewer knowledge workers.

What if you combine these two rather pessimistic ideas? Oddly enough, the result is a rather optimistic idea.

Here’s an example drawn a recent issue of The Economist. The process of discovery is often invoked in legal disputes between companies or between companies and government agencies. Each side has the right to inspect the other side’s documents, including e-mails, correspondence, web content, and so on. In complex cases, each side may need to inspect massive numbers of documents to decide which documents are germane and which are not. The actual inspecting and sorting has traditionally been done by highly trained paralegals – lots of them.

As you can imagine, the process is time-consuming and error-prone. It’s also fairly easy to automate through deep learning. Artificial neural networks (ANNs) can study the examples of which documents are germane and which are not and learn how to distinguish between the two. Just turn suitably trained ANNs loose on boxes and boxes of documents and you’ll have them sorted in no time, with fewer errors than humans would make.

In other words, artificial neural networks can do a better job than humans at lower cost and in less time. So this should be bad news for paralegal employment, right? The number of paralegals must be plummeting, correct? Actually no. The Economist tells us that paralegal employment has actually risen since ANNs were first deployed for discovery processes.

Why would that be? Jevons paradox. The use of ANNs has dramatically lowered the obstacles to using the discovery process. Hence, the discovery process is used in many more situations. Each discovery process uses fewer paralegals but there are many more discovery processes. The net effect is greater – not lesser – demand for paralegals.

I think of this as good news. As the cost of a useful process drops, the process itself – spam filtering, document editing, image identification, quality control, etc. – can be deployed to many more activities. That’s useful in and of itself. It also drives employment. As costs drops, demand rises. We deploy the process more widely. Each human is more productive but more humans are ultimately required because the process is far more widespread.

As a teacher, this concept makes me rather optimistic. Artificial intelligence can augment my skills, make me more productive, and help me reach more students. But that doesn’t mean that we’ll need fewer teachers. Rather, it means that we can educate many, many more students. That’s a good thing – for both students and teachers.

Male Chauvinist Machines

Yanks win last night?

Yanks win last night?

Do men and women think differently? If they do, who should develop artificial intelligence? As we develop AI, should we target “feminine” intelligence or “masculine” intelligence? Do we have enough imagination to create a non-gendered intelligence? What would that look like?

First of all, do the genders think differently? According to Scientific American, our brains are wired differently. As you know, our brains have two hemispheres. Male brains have more connections within each hemisphere as compared to female brains. By contrast, female brains have more connections between hemispheres.

Men, on average, are better at connecting the front of the brain with the back of the brain while women are better at connecting left and right hemispheres. How do these differences influence our behavior? According to the article, “…male brains may be optimized for motor skills, and female brains may be optimized for combining analytical and intuitive thinking.”

Women and men also have different proportions of white and gray matter in their brains. (Click here). Gray matter is “…primarily associated with processing and cognition…” while white matter handles connectivity. The two genders are the same (on average) in general intelligence, so the differences in the gray/white mix suggest that there are two different ways to get to the same result. (Click here). Women seem to do better at integrating information and with language skills in general. Men seem to do better with “local processing” tasks like mathematics.

Do differences in function drive the difference in structure or vice-versa? Hard to tell. Men have a higher percentage of white matter and also have somewhat larger brains compared to women. Perhaps men need more white matter to make connections over longer distances in their larger brains. Women have smaller heads and may need less white matter to make the necessary connections — just like a smaller house would need less electrical wire to connect everything. Thus, a larger proportion of the female brain can be given over to gray matter.

So men and women think differently. That’s not such a surprise. As we look ahead to artificial intelligence, which model should we choose? Should we emphasize language skills, similar to the female brain? Or local processing skills, similar to the male brain? Should we emphasize processing power or information integration?

Perhaps we could do both, but I wonder how realistic that is. I try to imagine what it would be like to think as a woman but I find it difficult to wrap my head around the concept. As a feminist might say, I just don’t get it. I have to imagine that a woman trying to think like a man would encounter similar difficulties.

Perhaps the best way to develop AI would involve mixed teams of men and women. Each gender could contribute what it does best. But that’s not what’s happening today. As Jack Clark points out, “Artificial Intelligence Has A “Sea of Dudes’ Problem”. Clark is mainly writing about data sets, which developers use to teach machines about the world. If men choose all the data sets, the resulting artificial intelligence will be biased in the same ways that men are. Yet male developers of AI outnumber females by a margin of about eight-to-one. Without more women, we run the risk of creating male chauvinist machines. I can just hear my women friends saying, “Oh my God, no!”

From AI to UBI

We'll fill it when you're born.

We’ll fill it when you’re born.

In his 1984 novel, Neuromancer, that kicked off the cyberpunk wave, William Gibson wrote about a new type of police force. Dubbed the Turing Police, the force was composed of humans charged with the task of controlling non-human intelligence.

Humans had concluded that artificial intelligence – A.I. – would always seek to make itself more intelligent. Starting with advanced intelligence, an A.I. implementation could add new intelligence with startling speed. The more intelligence it added, the faster the pace. The growth of pure intelligence could only accelerate. Humans were no match. A.I. was a mortal threat. The Turing Police had to keep it under control.

Alas, the Turing Police were no match for gangsters, drug runners, body parts dealers, and national militaries. The most threatening A.I. in the novel was “military-grade ice” developed by the Chinese Army. Was Gibson prescient?

If the Turing Police couldn’t control A.I., I wonder if we can. Three years ago, I wrote a brief essay expressing surprise that a computer could grade a college essay better than I could. I thought of grading papers as a messy, fuzzy, subtle task and assumed that no machine could match my superior wit. I was wrong.

But I’m a teacher at heart and I assumed that the future would still need people like me to teach the machines. Again, I was wrong. Here’s a recent article from MIT Technology Review that describes how robots are teaching robots. Indeed, they’re even pooling their knowledge in “robot wikipedias” so they can learn even more quickly. Soon, robots will be able to tune in, turn on, and take over.

So, is there any future for me … or any other knowledge worker? Well, I still think I’m funnier than a robot. But if my new career in standup comedy doesn’t work out, I’m not sure that there’s any real need for me. Or you, for that matter.

That raises an existential question: are humans needed? We’ve traditionally defined “need” based on our ability to produce something. We produced goods and services that made our lives better and, therefore, we were needed. But if machines can produce goods and services more effectively than we can, are we still needed? Perhaps it’s time to re-define why we’re here.

Existential questions are messy and difficult to resolve. (Indeed, maybe it will take A.I. to figure out why we’re here). While we’re debating the issue, we have a narrower problem to solve: the issue of wealth distribution. Traditionally, we’ve used productivity as a rough guide for distributing wealth. The more you produce, the more wealth flows your way. But what if nobody produces anything? How will we parcel out the wealth?

This question has led to the development of a concept that’s now generally known as Universal Basic Income or U.B.I. The idea is simple – the government gives everybody money. It doesn’t depend on need or productivity or performance or fairness or justice. There’s no concept of receiving only what you deserve or what you’ve earned. The government just gives you money.

Is it fair? It depends on how you define fairness. Is it workable? It may be the only workable scheme in an age of abundance driven by intelligent machines. Could a worldwide government administer the scheme evenhandedly? If the government is composed of humans, then I doubt that the scheme would be fair and balanced. On the other hand, if the government were composed of A.I.s, then it might work just fine.

 

Ethics And The Self-Driving Car

Why does it have to look like a car?

Why does it have to look like a car?

In last night’s critical thinking class, we discussed and debated how to instill ethics into self-driving cars. To kick off the debate, I adapted a question from an article by Will Knight in a recent issue of Technology Review.

If a child runs in front of a self-driving car, is it ethical for the car to injure or kill its own passengers in order to avoid the child?

The question split the class roughly in two. Some students argued that the car should avoid the child at all costs – even at the risk of serious injury or death to the car’s passengers. Others argued that the life of the person in the car was at least equal to that of the child. And if the car contains more than one passenger, wouldn’t multiple lives outweigh the value of one child’s life? Why should the child alone be protected?

The students also weighed the value of different lives. I asked them to assume that I was in the car. So the question became: Which is more valuable, the life of a child or the life of a mature adult? Some argued that the child was more valuable because she has many years of life yet to live. By comparison, I have lived many years and “used up” some of my value.

Others argued exactly the opposite. A mature adult is a storehouse of knowledge and wisdom. To lose that would be a blow not only to the individual but also to the community. The death of a child, though tragic, would eliminate only a small quantity of wisdom.

The question then evolves to: Could we have different ethical systems in cars that are used in different cultures? For instance, could cars that are used in cultures that value older people as fonts of wisdom, use an ethical system that protects adults over children? At the same time, could cars used in youth-oriented cultures, opt to protect children ahead of adults? What are the ethics here?

As you may have noticed, we were all making one key assumption: that self-driving cars will be delivered with an ethical system already installed. Let’s change that assumption. Let’s say the car arrives in your driveway (it delivers itself presumably) with no ethical rules programmed into it. The first thing we do is to program in the rules that we believe are the most ethical. The rules I program into my car may be different than the rules you put in your car.

This is, of course, exactly what we do today. When I drive my car, it’s controlled by my ethical system. Your ethical system controls your car. Our ethics may be different, so our cars “behave” differently. We’re quite used to that in everyday life, partially because the rules are not explicit. But as we explicitly program ethics into cars, is that really the way we want to do it?

Pushing on, we also addressed questions of self-driving cars with no passengers. We generally assume that self-driving cars will have passengers and at least one of those will be a responsible adult. But do they have to? Here are two variants:

  • I want a pizza. Is it ethical for me to send my self-driving car to the local pizza place to pick it up for me? Conversely, could the pizza place simply deliver it to me in a self-driving car with no human in it?
  • Is it ethical for a Dad to place his two kids in a self-driving car and instruct the car to deliver them to their primary school … without Dad going along? Could self-driving cars become child delivery systems?

I think we can find many more questions here. To wit: Can you own a self-driving car or is it a community resource to be shared? Does a self-driving car have to look like a regular old car? Why? (Here’s an example of one that doesn’t).

There’s a lot to think about here. Indeed, we may find that it will be more difficult to solve the ethical problems than technical problems. I’ll write more about these issues in the future. In the meantime: Safe driving!

My Social Media

YouTube Twitter Facebook LinkedIn

Newsletter Signup
Archives