There is a contradiction in economic forecasting today that I’ve come to think of as the “robot paradox.” Some people seem confident that automation will take many workers’ jobs, yet they cannot point to evidence that technology has done anything in the last few years to replace work or add to productivity. Indeed, economic growth has been lackluster for the last few years, productivity growth is mysteriously moribund, and the last two years have been perhaps the best time this century for wage growth. This is not what the end of work looks like.

Since I have written repeatedly that policymakers should take the threat of automation seriously, I’ve developed several theories about the robot paradox. The first begins with humility: Maybe I’m wrong, and today’s statistics are evidence that machines will continue to supplement workers, as they have mostly done for the last few centuries, rather than erode overall employment. The second is that I’m right, just not yet: The economy is on the precipice of several wrenching changes—self-driving cars, machine-learning, and the continued digitization of shopping—that will replace hundreds of thousands of jobs in a future so near it is practically the edge of the present.

But the third theory is the most important, the most empirical, and yet the most overlooked. It is that the time to look for technological displacement of workers is not during recoveries, but rather during recessions. There is nothing to see now, but after the next downturn (or the recession after that), there will be.

Consider manufacturing, where employment has fallen by almost 40 percent, or six million jobs, since the 1960s. Most people don’t realize that in most years since 1970, manufacturing employment has been relatively steady or increased. Between 1969 and 2015, manufacturing employment grew by four million during recoveries. But factory jobs fell by 10 million during recessions. In short, the U.S. economy has been in a recession for less than 15 percent of the time between 1970 and 2015, but these periods accounted for 167 percent of the decline in manufacturing jobs.

In my cover story last year on technology and the future of work, I spoke with Henry Siu, an economist from the University of British Columbia, who researched the relationship between worker-replacing technologically and recessions. Sometimes technology acts like secret agent embedded with employees, working alongside them for years before suddenly turning on them and replacing many jobs in a fell swoop. “The personal computer existed in the ’80s,” he said, “but you don’t see any effect on office and administrative-support jobs until the 1990s, and then suddenly, in the last recession, it’s huge. So today you’ve got checkout screens and the promise of driverless cars, flying drones, and little warehouse robots. We know that these tasks can be done by machines rather than people. But we may not see the effect until the next recession, or the recession after that.”

For years, this theory has been more of an observation than an empirical finding. But a new study by the economists Brad Hershbein, of the W.E. Upjohn Institute for Employment Research, and Lisa B. Kahn, at the Yale University School of Management, finds supporting evidence in job vacancy postings. In areas that suffered steep downturns in 2007 and 2008, they found significant evidence of “up-skilling” in job postings. That is, firms were more likely to demand college experience and computer expertise for jobs that used to be low-skilled. After ruling out some other explanations, they conclude that these job postings were a reflection of the fact that the recession accelerated the adoption of computer technology. In other words, the Great Recession replaced workers with technology and required that new hires have computer expertise to work with their new “colleagues,” who were now computers. Their paper provides “the first direct evidence that the Great Recession accelerated routine-biased technological change.”

Nobody can be certain of technological shifts many years in the future. But in the event that robots will take many people’s jobs is important, the timing is as important as the event itself. If the shifts were gradual, policy-makers would have more time to debate and write laws and workers would have more time to learn new skills. But automation is not a gradually applied force. Instead, labor-replacing automation seems to operate like the famous description of bankruptcy—slowly, slowly, and then all at once.