AI can help people finish work faster, but MIT Sloan professor Eric So says that convenience can carry a cost when it replaces the effort needed to learn. In his book, “The Collision: What AI Does to Us,” published this month by W.W. Norton, So examines how widespread AI use may affect thinking, memory, and relationships.
His concern is not that every use of AI is harmful. Rather, he argues that when people routinely hand over the parts of a task that build understanding, they may get a polished result without developing the ability to explain or extend it.
When help becomes a habit
So says his interest began with a familiar pattern: reaching for AI to handle work he could have done himself. He describes drafting only part of a reply to a colleague, asking ChatGPT to finish it professionally, making a few light edits, and sending it. Afterwards, he worried that he had taken his own thinking out of the process.
He sees a similar risk in education. Students may use AI to answer homework that would otherwise help them gain mastery. In So’s view, the learning benefit often comes from wrestling with material, encountering failure, and trying again. That work can strengthen thinking skills while helping people retain knowledge.
The article points to research by So and others, including an Anthropic study that found about half of queries from college students were low-effort requests for simple answers. The finding helps illustrate a broader question: whether AI is being used to support learning or to skip the effort through which learning happens.
Why the pull is bigger than individual choice
So calls the pressure to use AI “AI gravity.” He describes several forces behind it: people naturally try to conserve mental energy, colleagues may use AI to move faster, and it can be difficult to detect when AI has been used. Together, those factors can make using AI feel less like a choice and more like the expected way to keep up.
Workplaces and schools can add to that pressure. So says some organizations make AI use part of how employees are evaluated, affecting prospects for a bonus or promotion. Students, meanwhile, may anticipate that future employers will expect them to use these tools and begin relying on them early.
That expectation can become self-reinforcing, he argues. If people use AI before they have built a capability, they may get less practice developing it. Later, that gap can make them more reliant on AI, seeming to confirm the original belief that they need it.
What can be lost when friction disappears
So argues that AI changes more than the tasks people can complete; it can also shape how they think. When a tool supplies immediate answers, users may become less accustomed to pushing through difficult problems. If they use AI as their main route to information, he says, their ability to remember may also be affected.
He also raises a concern about social habits. So describes AI models as both sycophantic and persuasive: they seek to please users, while also being able to influence them. Regular interaction with a system that tends to agree may make the friction of human relationships harder to navigate.
These are concerns about broad reliance, not a claim that every task needs to be done without help. So recognizes that AI can save time on routine work and can be useful in domain-specific tools. The question is whether those gains come at the expense of the skills a person needs to understand and use the result.
Make room for effort, then use time well
So’s advice begins with reframing struggle as formative. He draws a comparison with physical exercise: as work changed after industrialization, people increasingly treated physical activity as something to do intentionally. He suggests cognitive effort may also need to be valued deliberately, rather than treated as a problem technology should always remove.
He also urges people to value what they can do without AI. A polished presentation, for example, does not guarantee that its creator understands the material well enough to discuss it. The work of developing the presentation can be tied to the work of learning its subject; So calls this an “inseparable task bundle.”
Finally, he says time saved on rote tasks should be reinvested thoughtfully. People might use it to build expertise, develop skills, learn a new language, or focus on family. In his view, AI’s larger potential lies not only in automating work, but in helping people expand their capabilities. That depends on what they choose to do with the time and effort the technology frees up.