We’re talking about the future of work. Could you start by explaining how, in general, you think about the future of work and what you were looking for in your choice of books?

In my view, to think clearly about the future of work, we have to begin by looking at the past: at inherited ideas about the impact of technology on work and traditional views on how we ought to respond. By reflecting on what we got right and wrong, we can better prepare ourselves for the future. My first two choices reflect this: one book captures a popular view of machine capabilities; the other, a common plan for responding to the challenge of automation. My remaining three choices focus on the different problems that I think we will face in the years to come.

When I started thinking about the impact of technology on work, almost a decade ago, the sort of book that I would have begun with was The New Division of Labor, based on the fascinating work done by Frank Levy and Richard Murnane—that’s my first choice.

This book contains an early articulation of a view of machine capabilities that has been dominant in economics for some time. It is based, in part, on a particular understanding of how machines must operate: that they must somehow copy the way human beings think and reason in order to outperform them. If you want to automate a task, the argument goes, then you have try and uncover the rules that human beings follow when they perform that task, and capture them in a set of explicit instructions for a machine to follow. If you can capture how human beings perform a task in a set of rules, then that task can be automated. But if you cannot, then that task is likely to be out of reach of machines.

“I don’t think we’re taking seriously enough the threat of a world where there’s not enough well-paid work for people to do”

This view is not confined to economics. It was also popular in the field of artificial intelligence at one point. I know this because my dad, with whom I co-authored a book called The Future of the Professions back in 2015, wrote his doctoral thesis on artificial intelligence and the law back in the 1980s at Oxford University. Almost forty years ago, he was part of the vanguard, trying to build systems that could solve legal problems. And importantly, the philosophy of what he was doing back then was very similar to what these economists were elaborating some years later; he thought that if you wanted to build a system to perform a legal task, you had to sit down with a human lawyer, get her to explain to you how it was she performed that task, then capture that explanation in a set of explicit rules for a machine to follow.

In economics, this view of machine capabilities gave rise to a very influential distinction—between so-called ‘routine’ tasks and ‘non-routine’ tasks. A task is ‘routine’ if a human being can articulate how she performs it; a task is ‘non-routine’ if she cannot. And the thought was that we can only automate ‘routine’ tasks, because those are the only ones for which we can readily articulate and write explicit rules for a machine to follow. This distinction is now widespread outside the academic world. Think how often people commenting on the future of work might claim that machines can only perform tasks that are ‘repetitive’ or ‘rules-based’ or ‘well-defined’; those are just different words for ‘routine’. Conversely, they might say that machines struggle with tasks that are ‘difficult to specify’ or ‘complex’; these are other ways of saying ‘non-routine’. The New Division of Labour, then, captures quite nicely an early account of a view of machine capabilities that has been dominant in economics, artificial intelligence, and wider life for some time.

Recently, though, a problem has emerged with this view of machine capabilities. What do the tasks of driving a car, making a medical diagnosis, and identifying a bird at a fleeting glimpse have in common? Well, these are all tasks that, until not too long ago, leading economists thought could not readily be automated. Yet today, all of them can be to an extent. Almost all major car manufacturers have driverless car programmes. There are countless systems that can diagnose medical problems. And there is even an app developed by the Cornell Laboratory of Ornithology that can tell you what bird you are looking at if you take a quick photo of it.

Why did economists think these tasks were out of reach of machines? Because they are all ‘non-routine’ tasks, on their definition that I described before. Ask a doctor, for instance, how she makes a medical diagnosis, and she is going to struggle to tell you exactly what rules she is following. She might be able to give you a couple of rough bits of advice, but in the end she will struggle. Instead, she will appeal to things like gut reaction, instinct, experience, and judgment. It is difficult to articulate how we use these faculties, and so it was thought it is very hard to automate tasks that required them: if a human being cannot explain how she performs a task, the argument goes, where do we begin in writing a set of rules for a machine to follow?

“Machines are increasingly able to perform tasks in fundamentally different ways to human beings, without having to copy our reasoning or thinking processes”

So, what went wrong? Because of recent advances in processing power, data storage capability and algorithm design, machines are increasingly able to perform tasks in fundamentally different ways to human beings, without having to copy our reasoning or thinking processes. The result is that it now matters far less that we struggle to articulate how we perform particular activities. Take the system recently developed by a team of researchers at Stanford that can tell whether or not a freckle is cancerous as accurately as leading dermatologists. How does it work? It is not trying to copy the judgment or the intuition of a human doctor. Instead it has a database of 129,450 past cases, and it’s running what is essentially a pattern recognition algorithm through those cases, hunting for similarities between them and the particular photo of the troubling freckle in question. It is performing the task in an un-human way, based on the analysis of more possible cases than any human doctor could hope to review in their lifetime.

The deeper consequence of these technological advances is that the ‘routine’ verses ‘non-routine’ distinction, which shapes how many people still think about automation, is actually far less useful than it was in the past. More and more ‘non-routine’ tasks are being taken on by machines. That is what fascinates and troubles me in my writing and research.

When was The New Division of Labour written?

The paperback came out 2005, so 16 years ago. I don’t think the ‘routine’ versus ‘non-routine’ distinction is ever explicitly made in that book—that may have come a little later in the literature—but it is one of the earliest places in economics that we see this emphasis on ‘rules’, and this preoccupation with trying to capture the ones that human beings follow when trying to automate a task. That is why it is such a valuable book.

I wrote A World Without Work, in part, in response to the apparent breakdown of this distinction. I find it remarkable that so many activities, which we had previously thought were out of reach of machines, are increasingly being automated. And I don’t think we’re taking seriously enough the threat of a world where there’s not enough well-paid work for people to do, because of these remarkable technological advances.

Even though in some sense it has been taken over, what conclusions do the authors draw about the future of work, given their assumptions about technological development?

The broader literature that has built up around this book tends to support an optimistic view about the future of work. It encourages us to think that there is some big realm of activity—namely all these ‘non-routine’ tasks—that we cannot automate. My fear is that this is no longer the case. However, rather than try to identify new boundaries to the capabilities of machines, I argue in my work that a more useful way to think about new technologies is that they are relentlessly—gradually, but relentlessly—becoming more capable, encroaching on tasks we once thought only human beings could do.