Three economists, Ajay Agrawal, Joshua Gans, and Avi Goldfarb, teamed up to write Prediction Machines: The Simple Economics of Artificial Intelligence. Maybe it’s not surprising, then, that their book boils down to three foundational insights:
- New developments in artificial intelligence don’t actually bring us intelligence but instead a critical component of intelligence — prediction.
- The real significance of which, at least for economists, is that AI makes prediction radically less expensive.
- When we have more and cheaper prediction, other complements become more valuable — especially judgment.
The rest of their book largely applies these three ideas so that business people, primarily, can better understand and leverage the ways AI is reshaping their industries. Let’s consider an overview of their reasoning.
That the really significant outcome of artificial intelligence is better, and especially cheaper, prediction is easily the book’s most interesting idea. And prediction, for the authors, is both simple and profound. Definitionally, it’s simple: “Prediction is the process of filling in missing information. Prediction takes information you have, often called ‘data,’ and uses it to generate information you don’t have.” But it’s also profound because prediction is such a foundational input. “Better prediction means better information, which means better decision making.”
It turns out the range of what the authors view as AI prediction activities is far-reaching indeed. They open their book with an especially arresting example. Your child is doing homework alone in another room when you hear a question, “What’s the capital of Delaware?” You start to think, Baltimore . . . too obvious . . . Wilmington . . . not a capital. “But before the thought is complete, a machine called Alexa says the correct answer: ‘The capital of Delaware is Dover.’ Alexa is Amazon’s artificial intelligence, or AI, that interprets natural language and provides answers to questions at lightning speed.”
But what Alexa was really doing, according to the authors, was prediction. Specifically, Alexa was “taking the sounds it heard and predicting the words the child spoke and then predicting what information the words were looking for. Alexa doesn’t ‘know’ the capital of Delaware. But Alexa is able to predict that, when people ask such a question, they are looking for a specific response: ‘Dover.’” Similarly, when you type a few words into the Google search bar, their algorithm predicts the results that most likely reflect what you want to know.
But here’s an example that stretches the bounds of prediction rather dramatically. Up till recently, autonomous vehicles were limited to highly predictable, controlled environments — until engineers reframed navigation as a prediction problem. “Instead of telling the machine what to do in each circumstance, engineers recognized they could instead focus on a single prediction problem: ‘What would a human do?’” Now various companies are testing autonomous vehicles on city streets and highways around the country.
From an economics viewpoint, though, what the authors find especially interesting is not that AI does (many kinds of) prediction well, it’s that it does prediction so cheaply. Google, for example, made search cheap, as does Alexa in a different form. In fact, what really interests the authors about technological change in general is that it makes things cheap that were once expensive.
As examples, they point out that what is economically significant about the computer revolution is that it made arithmetic cheap, in turn allowing us to discover more and more applications for newly-cheap (arithmetic) computing. Just like, previously, with the invention of electricity and the incandescent bulb, we found more and more uses for cheap light.
Now, they contend, AI will be economically significant precisely because it again takes something valuable — prediction — and makes it much cheaper. Which means we are going to start using a lot more prediction, and “we are going to see it emerge in surprising new places.”
Does that mean widespread job loss, as cheap AI prediction takes over for humans? No, for at least two reasons. First, prediction is just one of several components of decision making, and some of the others, like judgment, especially play to human strengths.
As an example, the authors point to the advent of automatic teller machines (ATM) — designed to automate tellers out of existence. What actually happened, though, was that the most routine aspects of tellers’ jobs were automated, freeing them to take on higher-value customer service roles. Their new tasks — talking to customers about their banking needs, advising them on loans, and working out credit card options — were more complicated, and required a great deal more subjective judgment. Arguably, these new jobs made much better use of the particular strengths of humans.
Second, even in the realm of prediction, both machines and humans have their respective advantages.
In 2016, a Harvard/MIT team of AI researchers won the Camelyon Grand Challenge, a contest that produces computer-based detection of metastatic breast cancer from slides of biopsies. The team’s winning deep-learning algorithm made the correct prediction 92.5 percent of the time compared with a human pathologist whose performance was at 96.6 percent. [But] the researchers went further and combined the predictions of their algorithm and a pathologist’s. The result was an accuracy of 99.5 percent. That is, the human error rate of 3.4 percent fell to just 0.5 percent. Errors fell by 85 percent.
It turns out the human and the machine were good at different aspects of prediction. The human pathologist was usually right when saying there was cancer. In fact, it was unusual to have a situation in which the human said there was cancer but was mistaken. In contrast, the AI was much more accurate when saying the cancer wasn’t there. The human and the machine tended to make different types of mistakes. “By recognizing these different abilities, combining human and machine prediction overcame these weaknesses, so their combination dramatically reduced the error rate.”
These sorts of human-machine collaborations are, according to the authors, where our future is headed. Both the customer service specialists who used to be tellers, or the pathologist who, with help from artificial intelligence, now makes dramatically better biopsy assessments, would tell us, I suspect, that we don’t really have to worry about that future — too much.