Daron Acemoglu, the Massachusetts Institute of Technology professor who recently won the Nobel Prize in economics, worries that artificial intelligence will worsen income inequality and not do all that much for productivity. His friend and colleague David Autor is more hopeful, believing that AI could do just the opposite.

New research from Aidan Toner-Rodgers, an MIT doctoral student, challenges both Acemoglu’s pessimism and Autor’s optimism. Both professors are raving about it.

“It’s fantastic,” said Acemoglu.

“I was floored,” said Autor.

Neither Autor nor Acemoglu is changing his mind on AI. But the research by Toner-Rodgers, 26 years old, is a step toward figuring out what AI might do to the workforce, by examining AI’s effect in the real world.

Many economists, including Autor and Acemoglu, have looked at how earlier technologies have reshaped the labor market. But while this understanding of the past provides important context, how AI will affect the economy is difficult to tease out: Will it be like the gasoline-powered internal combustion engine, which transformed entire industries, boosting growth, creating vast categories of new work and lifting millions of Americans into new, more productive, better-paying jobs? Or the zeppelins of the 1920s and 1930s, which people thought would be world changers and are now a nostalgic afterthought?

The zeppelin didn’t exactly take off.

To figure out where AI might fit, economists need careful studies of its use in today’s workplace. Toner-Rodgers’s paper does just that. His work examines the randomized introduction of an AI tool to 1,018 scientists at a materials-science research lab.

The discovery and creation of new materials—from the invention of Bakelite in the 1900s to Kevlar in the 1960s—has historically been a time-consuming process of trial and error. Scientists, after identifying what properties they would like a compound to have, then come up with ideas of what the chemical structure of a new compound might look like. Then they start testing out compounds, hoping to hit on one that works. 

AI tools that have been trained on the structure of existing materials can make the discovery process significantly shorter and less expensive. Scientists specify the characteristics they would like a compound to have and the AI tool generates recipes that the scientists can then evaluate.

“Maybe the most exciting thing about AI is that it could accelerate scientific discovery and innovation,” said Toner-Rodgers. “This would be a huge benefit.”

The child of two California schoolteachers, Toner-Rodgers was a basketball-obsessed kid who played guard for Macalester College in Minnesota. He took his first economics class as a freshman and got hooked. After a stint working at the Federal Reserve Bank of New York, he entered MIT in 2023. 

The lab that Toner-Rodgers studied randomly assigned teams of researchers to start using the tool in three waves, starting in May 2022. After Toner-Rodgers approached the lab, it agreed to work with him but didn’t want its identity disclosed. 

Aidan Toner-Rodgers 

What Toner-Rodgers found was striking: After the tool was implemented, researchers discovered 44% more materials, their patent filings rose by 39% and there was a 17% increase in new product prototypes.  Contrary to concerns that using AI for scientific research might lead to a “streetlight effect”—hitting on the most obvious solutions rather than the best ones—there were more novel compounds than what the scientists discovered before using AI.

Toner-Rodgers was a bit surprised himself. He had thought at best it would have just kept up with the scientists on novel discoveries. “You could have come up with a bunch of lame materials that are not actually helpful,” he said.

These gains in scientific innovation could lead to gains elsewhere, since new inventions can lead to unexpected developments down the line. The gasoline-powered internal combustion engine was developed to power automobiles, but when used in tractors, it transformed farming. The share of U.S. agricultural employment fell from 20% in 1930 to 6% in 1960, notes Harvard University economist David Deming, but over that period agricultural output grew fourfold.

Acemoglu said he was “somewhat surprised” the lab experienced such a marked increase in productivity, which, if repeated enough throughout the economy, would cut against his pessimistic take. In a recent paper, he estimated that AI would boost U.S. gross domestic product over the next decade by a total of about 1%.

The lab is just one, specific example. Acemoglu notes that unlike large language models such as ChatGPT that people are trying to apply broadly, the lab’s AI tool was built specifically for materials discovery.

Gasoline-powered tractors transformed farming.

The scientists that Toner-Rodgers studied also held advanced degrees in chemistry, physics and engineering. Those are skills that might have enabled them to adopt AI more easily. 

Acemoglu still worries that AI could widen income disparities. Toner-Rodgers’s paper suggests one way that might happen.

He found that researchers who were already the most successful at discovering compounds were even more successful with the AI tool, while other scientists didn’t benefit as much. Because an individual’s compensation tends to be tied with their productivity, that augurs for increased income inequality.

Looking at the recipes that the tool suggested, the top scientists tended to correctly identify the ones that were most likely to succeed, and move forward with those first. The scientific output of researchers in the top 10th in terms of past success increased by 81%.

But other researchers were worse at picking winners and spent a lot of time evaluating potential compounds that didn’t work out. Those in the bottom third registered little improvement at all.

Autor still thinks that AI could reduce income inequality as long as workers are properly trained. A paralegal with AI and sufficient training, for example, might perform many of the tasks now relegated to experienced lawyers—and thus make more money.

Otherwise “it’s like sending people up in an airplane without putting them in a flight simulator first and telling them, ‘Oh, sorry, you crashed, I guess you’re a bad pilot,’” he said.

David Autor and Daron Acemoglu have both studied how earlier technologies reshaped the labor market.

One last thing Toner-Rodgers found about the lab’s AI tool: The scientists didn’t like it all that much, with 82% reporting reduced satisfaction with their work.

While many AI optimists believe the technology will reduce the number of tedious tasks people have to perform, the scientists felt that it took away the part of their jobs—dreaming up new compounds—they enjoyed most. One scientist remarked, “I couldn’t help feeling that much of my education is now worthless.”

Perhaps as they get more familiar with AI, the scientists will get happier with using it, but there are no guarantees.

“A key, creative part of the process was automated,” said Toner-Rodgers. “People just might be unhappy with that permanently.”