Dario Amodei is the CEO of Anthropic, a leader in artificial intelligence. For that reason, his comments in a recent interview grabbed headlines. Amodei argued that the next generation of AI systems could replace half of entry-level jobs and drive unemployment up to 20%. All of this could occur in the next five years, he said.

Recent data points seem to support these glum predictions. Mark Zuckerberg has said that AI will be as capable as a mid-level programmer by the end of this year. Microsoft announced thousands of job cuts this spring, with programmers disproportionately affected. In an announcement last week, Amazon’s CEO wrote: “It’s hard to know exactly where this nets out over time, but in the next few years, we expect [artificial intelligence] will reduce our total corporate workforce…”

If these data points are a sign of things to come, it would certainly be concerning. This outcome isn’t guaranteed, though. To illustrate why I see things differently, I’ll share a recollection from some years ago.

My father was a lawyer, and I remember visiting his office when I was a child. What was notable, in hindsight, is that there was no computer on his desk. There were no computers anywhere. To send correspondence, attorneys would speak into dictation machines. They would then hand the machines’ miniature tapes to secretaries, who would type up a first draft—literally on typewriters, with white-out to fix errors and carbon paper to make copies. The draft would be marked up in pen, then a final copy retyped, put in an envelope, stamped and mailed.

Because of the amount of work involved, each attorney in those days had a dedicated secretary. They spent their days typing and retyping documents like this. Today, of course, most everyone does their own correspondence, and as a result, there are far fewer of these jobs. But if we look back at historical data, we can see that the invention of the word processor and email didn’t cause any noticeable increase in unemployment. That’s for a few reasons. For starters, transitions like this occur slowly.

Also, new technologies are rarely a net negative. Instead, they tend to also create new jobs. Look at a law firm today, for example, and they’ll have far fewer traditional secretaries but significant numbers of people in IT, a field that didn’t even exist back in the days of typewriters and dictaphones.

The transition in office work is just the most recent change but by no means unique. Years ago, many people were employed as Morse code operators. Sam Altman, founder of OpenAI, has commented that in the past there were also large numbers of people employed as lamplighters, going down city streets each night lighting gas lamps. More significantly, farming used to be a major sector of the workforce. In 1900, 40% of Americans worked on farms. Today, it’s less than 2%.

These transitions were all significant, but none caused the sort of mass unemployment Dario Amodei is forecasting. At Amazon, in fact, the company has said it doesn’t expect AI to result in significant layoffs. Instead, the bulk of the headcount reduction is expected to occur through attrition—the normal course of employees changing jobs or retiring.

Amazon provides another useful data point on this topic. In its warehouses, the company now uses 750,000 robots. In theory, those robots have taken what would have been 750,000 jobs for people, but that’s not what the overall employment data show. Unemployment today is near the low end of where it’s been over the past 75 years.

How is it possible that technology has displaced jobs, and yet unemployment remains low? An economic principle known as the Jevons paradox can help us understand this. William Jevons was a British economist. In the 1860s, he observed that manufacturing plants had become more efficient in their use of coal and yet, counterintuitively, demand for coal was increasing.

The explanation: As manufacturers realized that they needed less coal to produce the same amount of output, they chose to expand their businesses into new areas, resulting in the use of more coal.

Over time, a second order effect kicked in. These gains in output led to wage increases and faster economic growth. That, in turn, further increased demand for coal.

The adoption of artificial intelligence might deliver the same positive effects, with greater productivity leading to higher wages and faster economic growth without any loss in employment.

Recent comments by the CEO of software vendor Box illustrate how the Jevons paradox might apply to AI. “AI is not replacing existing work that’s being done,” he wrote, “but adding new capabilities to the organization.” Box, in other words, won’t simply use AI to cut costs; they’ll use the technology to do more.

He adds: “This means that companies can simply now attack the kinds of problems that just never were economically feasible to solve before…Yes, there’s certainly opportunity to automate some of the work that we currently do to drive efficiency, but the vast majority of work that we will bring automation to is the work that we just never got around to in the first place.”

Interestingly, Amodei, despite his concerns about employment, sees some of these same benefits. In the same interview in which he made his comments about unemployment, he also described some of the potential positive effects AI might deliver: “Cancer is cured, the economy grows at 10% a year, the budget is balanced…”

The reality is that this is all an open question. In a May interview, Daron Acemoglu, an economist who recently won the Nobel Prize for his work on economic development, argues that AI will be able to replace only a fraction of jobs. But he adds, “it’s hugely uncertain, and it’s very difficult to know because it’s a very rapidly changing technology.”

That may be the best way to think about this question. It’s all very new and still uncertain. While Amodei worries about a potentially negative impact, that’s just a guess. Economic history suggests it very well may go the other way.​