Artificial intelligence machines are getting so good, so quickly, that they’re poised to replace humans across a wide range of industries. In the next decade, we’ll see machines barge into areas of the economy that we’d never suspected possible—they’ll be diagnosing your diseases, dispensing your medicine, handling your lawsuits, making fundamental scientific discoveries, and even writing stories just like this one. Economic theory holds that as these industries are revolutionized by technology, prices for their services will decline, and society as a whole will benefit. As I conducted my research, I found this argument convincing—robotic lawyers, for instance, will bring cheap legal services to the masses who can’t afford lawyers today. But there’s a dark side, too: Imagine you’ve spent three years in law school, two more years clerking, and the last decade trying to make partner—and now here comes a machine that can do much of your $400-per-hour job faster, and for a fraction of the cost. What do you do now?

There is already some evidence that information technology has done permanent damage to workers in a large sector of the economy. This specifically applies to workers who are considered “middle skilled,” meaning that they need some training, but not much, to do their jobs.

Middle-skilled jobs include many that are generally recognized to be antiquated—secretaries, administrative workers, repairmen, and manufacturing workers, among others. Since the 1980s, across several industrialized nations (including the United States), the number of workers in these job categories has been rapidly declining (the pace of the decline increased greatly during the last recession). Instead, most job growth has been at the poles, in professions that require very high skills and earn high wages, and in the service sector, where most jobs require few skills and pay tiny wages.

David Autor, an economist at MIT who is the leading scholar of this phenomenon, calls it “job polarization.” Autor identifies a number of causes for the decline of middle-skilled work, including the decreasing power of unions and the declining federal minimum wage. He puts one factor above the rest, however: The rise of information technology.

Autor argues that middle-skilled jobs tend to have two factors in common—they are composed of lots of tasks that are both routine and geographically portable. What does a secretary do all day? He files, sorts, organizes, watches for calendar conflicts, and in other ways manipulates information. What does a tax preparer do? He asks you a series of questions, and performs some calculations based on your answers. These are all tasks that can be written in software—and, once there, they can be done faster, and more cheaply, by machines. And even when a computer can’t completely replace these middle-skilled jobs, it can make them easier to transfer to lower-wage humans—you still need a human being to answer tech support questions, but now you can hire someone in Andra Pradesh rather than Alabama. This decimation of middle-skilled work explains another unsettling trend in American business. New companies today are starting up with far fewer workers than in the past, and they’re staying smaller as they grow.