WHY is productivity growth low if information technology is advancing rapidly? Prominent in the 1980s and early 1990s, this question has in recent years again become one of the hottest in economics. Its salience has grown as techies have become convinced that machine learning and artificial intelligence will soon put hordes of workers out of work (among tech-moguls, Bill Gates has called for a robot tax to deter automation, and Elon Musk for a universal basic income). A lot of economists think that a surge in productivity that would leave millions on the scrapheap is unlikely soon, if at all. Yet this year’s meeting of the American Economic Association, which wound up in Philadelphia on January 7th, showed they are taking the tech believers seriously. A session on weak productivity growth was busy; the many covering the implications of automation were packed out.

Recent history seems to support productivity pessimism. From 1995 to 2004 output per hour worked grew at an annual average pace of 2.5%; from 2004 to 2016 the pace was just 1%. Elsewhere in the G7 group of rich countries, the pace has been slower still. An obvious explanation is that the financial crisis of 2007-08 led firms to defer productivity-boosting investment. Not so, say John Fernald, of the Federal Reserve Bank of San Francisco, and co-authors, who estimate that in America, the slowdown began in 2006. Its cause was decelerating “total factor productivity”—the residual that determines GDP after labour and capital have been accounted for. Productivity has stagnated despite swelling research spending (see chart). This supports the popular idea that fewer transformative technologies are left to be discovered.

Others take almost the diametrically opposed view. A presentation by Erik Brynjolfsson of MIT pointed to recent sharp gains in machines’ ability to recognise patterns. They can, for instance, outperform humans at recognising most images—crucial to the technology behind driverless cars—and match dermatologists’ accuracy in diagnosing skin cancer. Mr Brynjolfsson and his co-authors forecast that such advances will eventually lead to a widespread reorganisation of jobs, affecting high- and low-skilled workers alike.

Productivity pessimism remains the norm among official forecasters, but more academics are trying to understand how automation may affect the economy. In a series of papers, Daron Acemoglu of MIT and Pascual Restrepo of Boston University present new theoretical models of innovation. They propose that technological progress be divided into two categories: the sort that replaces labour with machines; and that which creates new, more complex tasks for humans. The first, automation, pushes down wages and employment. The second, the creation of new tasks, can restore workers’ fortunes. Historically, the authors argue, the two types of innovation seem to have been in balance, encouraged by market forces. If automation leads to a labour glut, wages fall, reducing the returns to further automation, so firms find new, more productive ways to put people to work instead. As a result, previous predictions of technology-induced joblessness have proved mostly wrong.

However, the two forces can, in theory, fall out of sync. For example, if capital is cheap relative to wages, the incentive to automate could prevail permanently, leading the economy to robotise completely. The authors speculate that, for now, biases towards capital in the tax code, or simply an “almost singular focus” on artificial intelligence, might be tilting firms towards automation, and away from thinking up new tasks for people. Another risk is that much of the workforce lacks the right skills to complete the new-economy tasks that innovators might dream up.

These ideas shed light on the productivity paradox. Mr Brynjolfsson and his co-authors argue that it can take years for the transformative effects of general-purpose technologies such as artificial intelligence to be fully felt. If firms are consumed by efforts to automate, and such investments take time to pay off, it makes sense that productivity growth would stall. Investment has not been unusually low relative to GDP in recent years, but it has shifted away from structures and equipment, towards research-and-development spending.

If research in automation does start yielding big pay-offs, the question is what will happen to the displaced workers. Recent trends suggest the economy can create unskilled jobs in sectors such as health care or food services where automation is relatively difficult. And if robots and algorithms become far cheaper than workers, their owners should become rich enough to consume much more of everything, creating more jobs for people.

The risk is that without sufficient investment in training, technology will relegate many more workers to the ranks of the low-skilled. To employ them all, pay or working conditions might have to deteriorate. If productivity optimists are right, the eventual problem may not be the quantity of available work, but its quality.