What 2 MIT experts are thinking about AI and work

While many people think of artificial intelligence as an automation tool, MIT economics professor David Autor said it’s best to see it as a collaboration tool that amplifies employee skills instead of replacing them. 

“For airline pilots, we really want them to be able to fly a plane manually. We don’t want them to be fully dependent on autopilot. So it matters whether their skills decline,” he said.

In a recent episode of the “MIT CSAIL Alliances” podcast, Autor and MIT Sloan principal research scientist Neil Thompson explored AI’s impact on jobs, the future of work, and productivity. Here are five insights from their discussion. 

AI does not always boost productivity

Thompson said there are “very, very mixed” messages about AI’s effect on productivity. He pointed to a 2025 study that looked at experienced open-source developers writing an update to a software library. Developers who used generative AI wrote code faster but took 19% longer to complete the overall task compared with the control group. Writing prompts, checking outputs, and waiting for the model to do its work took up time.

Yet developers in the study thought the AI tool had increased their speed by at least 20%, Thompson noted. “This is not that I think that everyone is going to be made less productive by AI,” he said. “But I do think this is telling us … that there are going to be a bunch of frictions.”

Automation has varied impacts

In a recent paper, Autor and Thompson looked at how automation changes the value of labor. By automating relatively easy tasks, tools like spellcheck and autocorrect placed greater value on the more advanced skills of proofreaders, Thompson said. For proofreaders that remained in the job market, wages increased. The same may ultimately be true for programmers with expertise in using software development tools rather than just basic coding, Autor added.

Yet when automation is applied to a role’s more expert tasks, wages tend to fall. That’s what happened to taxi drivers when their encyclopedic knowledge of city streets and optimal routes became available via GPS on smartphones. “Someone can now walk in off the street and drive a taxicab pretty well,” Thompson said.

This shows that it doesn’t just matter whether a job is exposed to automation. The bigger question is, “is this technology going to automate your kind of supporting tasks, allow you to work more efficiently on things that you’re really good at, or is it going to take your expert tasks and commoditize them so anyone can do them without you?” Autor said. 

AI Executive Academy

In person at MIT Sloan

Register Now

Collaboration is the goal 

Automation ultimately works best as a collaborative tool when it introduces capabilities humans lack, Autor said. 

He used the example of CheXbert, an AI model that can analyze and label radiology reports. It does as well as two-thirds of radiologists when making diagnoses based on X-rays alone, in part because “you can train this machine on 10 times as many labeled scans as a person could look at in their lifetime,” Autor said. 

But radiologists performed worse when they used CheXbert than when they acted on their own, especially when the AI was uncertain about what the X-ray indicated. Humans have access to valuable information, such as clinical history, that the AI does not, and the tool is not designed for collaboration. “It’s not a limitation of AI. It’s a challenge of designing it in a way that collaborates effectively with human capacities,” Autor said.

It’s important to keep humans in the loop

AI models get more costly the more accurate they need to be. There’s a steep jump in the cost of going from 80% to 90% accuracy, Thompson said, and even more going from 90% to 99%. 

“Many businesses I think are getting into this world where they say, ‘I want to fully automate,’ and then they realize that it’s just too expensive to do that,” Thompson said. 

Instead of fully automating a process where a mistake is costly, consider keeping a human in the loop to review an AI model’s output when it’s making important decisions.

AI can help with data center cooling, communication 

Autor and Thompson closed by highlighting two of the most effective AI use cases they’ve seen. 

• AI voice calls are being used by courier services in China to help deaf or hard-of-hearing drivers better engage with customers. This has reduced the volume of customer complaints while improving driver productivity and wages.
• Google uses deep learning to automate cooling in data centers, which represents a significant portion of the centers’ electricity budget. The AI systems are able to learn about different factors, such as wind flow, to decrease the amount of power needed to cool equipment.

Neil Thompson is a principal research scientist at MIT Computer Science and Artificial Intelligence Laboratory and the MIT Initiative on the Digital Economy, and director of the MIT FutureTech lab. He studies technological innovation and firm strategy. 

David Autor is a professor of economics at MIT and co-director of the MIT Stone Center on Inequality and Shaping the Future of Work. His research explores labor-market consequences of technological change.