MIT’s Eric So warns AI shortcuts can weaken the skills behind the work

A polished answer can conceal a missed chance to learn. In an October 15, 2026 interview with MIT News, MIT professor Eric So argued that heavy reliance on AI can strip away the effort that builds thinking skills. His proposed response is not to abandon AI, but to use it in ways that preserve that effort.

The concern began with an ordinary exchange. A colleague asked So for help, and he started writing a reply. Half a sentence in, he turned to ChatGPT to finish it professionally. He lightly edited the result and sent it. What unsettled him afterward was how little of his own thinking remained in the process.

He told interviewer Peter Dizikes that he had also seen students hand AI homework meant to help them master a subject. Those observations helped prompt The Collision: What AI Does to Us, published by W.W. Norton this month. So, a professor of global economics and behavioral science at MIT Sloan, examines how the tools affect both individual decisions and the incentives surrounding them.

The pull he calls “AI gravity”

So’s explanation goes beyond a lack of willpower. He describes a combination of the desire to conserve mental energy, pressure to keep up with other AI users, and difficulty detecting when someone has used the technology. Together, he calls those forces “AI gravity”: a persistent pull toward letting the tool do more.

That pull can become institutional. So says some organizations make AI use part of employee evaluations, tying it to bonuses or promotions. He also sees students anticipating that future employers will expect AI fluency. His concern is a self-reinforcing cycle: using AI before developing a capability can leave someone more dependent on it later.

The learning mechanism at the center of his argument is repeated effort: wrestling with material, making mistakes and trying again. Instant assistance can remove those opportunities, he argues, while making relief from difficulty feel normal. These are So’s arguments about AI use and learning, rather than results from a new experiment announced in the interview.

We need to recast how we think about struggle, as something that is formative rather than a problem to be solved.

Eric So, speaking to MIT News

Keeping the learning inside the task

For So, a polished presentation illustrates the difference between producing something and understanding it. He describes seeing Claude-made slides presented by people who had not sufficiently learned the material. Preparing slides and being able to discuss their substance are linked, he argues; automating the first can mean missing the preparation needed for the second.

His practical response rests on four principles, laid out in the interview:

1. Embrace productive struggle. Treat the effort of learning as something worth preserving. So compares this with deliberately exercising after industrialization reduced the physical activity required to earn a living.
2. Value what you can do without AI. Judge more than the finished output: the understanding developed while making it matters when you need to explain, discuss or use the work.
3. Reinvest the time saved. So acknowledges that AI can handle rote work faster. He urges putting the freed-up time toward building expertise, learning a language, cultivating skills or being with family.
4. Use AI as a tutor, not an answer machine. So argues that tailored assistance can support learning, provided the user accepts the effort a tutor asks for instead of seeking only a finished answer.

The unresolved challenge in that last recommendation is motivation. So sees AI as capable of helping people learn, but says users must actively seek the friction of tutoring. His advice asks people to choose a more demanding interaction even while the personal and workplace pressures he describes reward reaching for AI’s immediate help.