While many people think of artificial intelligence as an automation tool, David Orter, an economics professor at the Massachusetts Institute of Technology, says it’s best to view it as a collaboration tool that enhances rather than replaces employee skills.
“For airline pilots, we really want them to be able to fly the plane manually. We don’t want them to be completely dependent on the autopilot. So it’s important whether their skills degrade,” he said.
In a recent episode, “MIT CSAIL Alliances” Podcast, Autor and MIT Sloan Principal Investigator We investigated the impact of AI on work, the future of work, and productivity. Here are five insights from their discussion.
AI does not necessarily improve productivity
Mr Thompson said there were “very mixed” messages about the impact of AI on productivity. He pointed to a 2025 study that looked at experienced open source developers creating updates for software libraries. Developers who used generative AI wrote code faster compared to the control group, but took 19% longer to complete the entire task. It took a while to create the prompt, check the output, and wait for the model to work.
However, Thompson noted that the developers in the study believed that the AI tools had improved speed by at least 20%. “This is not to say that I think AI will make everyone less productive,” he said. “But I think this tells us that there’s going to be… a lot of friction.”
Automation has different impacts
In a recent paper, Autor and Thompson considered how automation changes the value of labor. Tools like spell check and autocorrect have placed more emphasis on the more advanced skills of proofreaders by automating relatively simple tasks, Thompson says. Wages for proofreaders who remained on the job market rose. The same may eventually apply to programmers who have expertise in using software development tools beyond just basic coding, Auter added.
However, wages tend to decline as automation is applied to the more specialized tasks of the role. That’s what happened to taxi drivers when encyclopedic knowledge of city streets and optimal routes became available via smartphone GPS. “Now I can come in from the street and drive a taxi pretty well,” Thompson said.
This shows that it’s not just about whether the job is automated or not. The bigger question is: Will this technology automate your support tasks and allow you to do what you’re really good at more efficiently, or will it commoditize specialist tasks so that anyone can do them without you? Autor said.
AI Executive Academy
Face-to-face at MIT Sloan
Register now
Collaboration is the goal
Automation ultimately works best as a collaboration tool by introducing capabilities that humans lack, Autor said.
He used the example of CheXbert, an AI model that can analyze and label radiology reports. Two-thirds of radiologists are just as effective in making diagnoses based on X-rays alone. One reason for this is that “we can train this machine on 10 times the number of labeled scans that a human can see in a lifetime,” Autor said.
But when radiologists used CheXbert, they performed worse than if they had acted alone, especially when the AI was uncertain about what the X-rays were showing. While humans have access to valuable information such as medical history, AI does not, and the tool is not designed for collaboration. “This isn’t a limitation of AI; it’s a challenge of designing AI in a way that works effectively with human capabilities,” Auter says.
It’s important to keep humans updated
The more accurate an AI model needs to be, the more expensive it becomes. Going from 80% to 90% accuracy is a significant jump in cost, and going from 90% to 99% is even more expensive, Thompson said.
“I think a lot of companies are entering a world where they say, ‘I want to fully automate,’ and then realize it’s too expensive to do that,” Thompson said.
Instead of fully automating processes where errors are costly, consider keeping humans in the loop to review the output of AI models when making important decisions.
AI and our working hours
AI helps data centers cool and communicate
Autor and Thompson concluded by highlighting two of the most effective AI use cases they’ve seen.
- AI voice call is in progress Used by Chinese courier services Helping hearing-impaired drivers better engage with customers. This reduced customer complaints and increased driver productivity and wages.
- Google uses deep learning to automate data center cooling, which accounts for a significant portion of a center’s power budget. AI systems can learn various factors, such as wind flow, to reduce the amount of power needed to cool equipment.
neil thompson I am the chief researcher of MIT Computer Science and Artificial Intelligence Laboratory and MIT Initiative on the Digital Economydirector of MIT Future Tech Laboratory researching technological innovation and corporate strategy.
david orter He is a professor of economics at MIT. MIT Stone Center on Inequality and Shaping the Future of Work. His research examines the impact of technological change on labor markets.
