Ajay Agrawal, Joshua Gans and Avi Goldfarb are all professors at the Rotman School of Management, University of Toronto, and with the school’s Creative Destruction Lab. They are the authors of Prediction Machines: The Simple Economics of Artificial Intelligence.
“AI is the next big thing,” your boss tells you. “Go out and get me some of that AI.” What do you do? Is AI something you can just go out and purchase? Is it like a robot you can buy and then just plug in and put it to work? To read the popular press on the coming AI revolution and the opportunity or threat that AI is coming for jobs, you could be forgiven for having that notion. But that is far from the reality of artificial intelligence today. Instead, as with many technologies, AI does one thing really well. You need to understand what that is and where it fits in before you can “get some AI.”
That one thing is prediction. Prediction is part of intelligence. We predict the day’s weather so as to decide what clothes to wear. We predict a person’s emotional state when we see a facial expression and decide what to say next. And we predict marketing effectiveness when we choose the words that will best persuade a customer to buy a product.
AI is often able to nail the prediction problem in ways that humans are not. AIs can quickly identify the content of images – so quickly that they can use your phone’s camera to tell that it is you before unlocking it. AIs can take words in French and translate them to English at speeds that translators cannot hope to achieve. They can even take long legal documents and identify sensitive information – tasks that can take paralegals hundreds of hours.
That’s great news. But that’s all the news. That is all AIs do. They use data to make predictions. We define prediction as using information you have to generate information you don’t have. So, for example, we call classification – such as taking a medical image of a tumour (information we have) and classifying it as malignant or benign (information we don’t have) – a type of prediction.
To make use of predictions, we often need to do more. It is all very well to have a prediction of the weather that tells you an umbrella might be necessary. But the AI does not know if the umbrella will fit conveniently into your bag or whether you can avoid the rain by ordering lunch in. To be sure, once we specify those missing areas of judgment, a clever engineer might fill them with a new app. But there are millions of decisions and equal numbers of variables to get those decisions right. Even the world’s entire stock of engineers cannot address them all.
Herein lies the challenge for business. When being sold a bill on AI, you need to ask yourself, “What uncertainty is that AI taking away?” Is it something that is really important to decisions you might or can make? Or is it something that is nice to know but that’s it. A fortune teller does you no good by telling you what will happen next week if there is nothing you can do about it.
AI offers the potential for the automation of thought. It forces us to reflect carefully on how people think and the parts of our thought process that machines can actually replace. In situations where prediction is the main challenge, there is great potential for automation. Anybody who has had the experience of teaching a teenager to drive knows that what causes you to shout in terror is not that the teenager does not know to brake when there is a car stopped in front of them, but instead, their inexperience in predicting whether a car in front of them is going to stop. Prediction machines excel at this.
In other cases, the predictions served up may be novel – so novel that we do not know what to do with them yet. When we don’t know things, we often play it safe rather than grapple with a decision. Businesses often present similar ads across many domains because they don’t know which individuals will see their ads. When AIs predict which individuals are most likely to view a particular ad, then tailoring becomes possible. But the AI does not know how to tailor and how individual types might respond to different ads. There is still a function for insightful people to make those judgment calls.
There is no such thing as “going to get some AI.” But there is such a thing as using the insight from data-driven predictions generated by an AI to enhance the performance of your organization. The rise of AI will further accentuate the performance advantage of those organizations that use data-driven decision-making versus those that don’t. In other words, avoiding data-driven decision-making is increasingly expensive. So, if you haven’t already, it might be time to get started with AI.