Below, Eric So shares five key insights from his new book, The Collision: What AI Does to Us.
Eric is the Sloan Distinguished Professor of Global Economics and Behavioral Science at MIT’s Sloan School of Management, faculty co-director of MIT’s AI Executive Academy, and adviser to companies on effective AI integration.
What’s the Big Idea?
We spend enormous energy asking what AI can do for us, but far too little asking what it is doing to us. AI can make us more productive today while quietly making us less capable tomorrow. The challenge is to use AI to strengthen our thinking, not replace it.
Listen to the audio version of this Book Bite—read by Eric himself—in the Next Big Idea App, or buy the book.
1. We need to grapple with AI gravity.
I use the term AI gravity to describe the constant psychological, social, and economic forces pulling us toward more and more AI use.
Picture Magnus Carlsen and Fabiano Caruana competing for the 2018 World Chess Championship. They sat almost motionless, yet their heart rates rose high enough to concern a cardiologist, and their bodies burned extraordinary amounts of energy. High-stakes thinking is incredibly expensive work. Our brains evolved to prefer the easier cognitive path. Now we give that effort-averse brain a tool that promises to do the hard thinking for us. Of course we reach for it. Our biology practically demands it.
Competitive pressures in modern society make it worse. Imagine suspecting that your classmates or coworkers are using AI to move faster, produce more, and seem smarter. The mere possibility can trigger an arms race. Outsourcing to AI is not simply a character flaw. It’s a predictable response to AI’s gravitational pull.
Before reaching for AI, ask whether this is a task where AI helps you think better or where it thinks instead of you. That is the central implication of AI gravity: telling people to “think for themselves” is not a serious AI policy. We need evaluations that test unassisted capability, workflows that preserve independent first judgments, and incentives that reward learning and verification rather than throughput alone. Individuals can recognize gravity. Institutions must change the field producing it.
2. AI can make us productive today and less capable tomorrow.
On a May afternoon in Houston, Engine 51 arrived at a restaurant fire. A thermal imaging camera displayed 184 degrees at the entrance—manageable for a fire scene. The reading was accurate but incomplete. Above the ceiling, beyond the camera’s reach, fire had been consuming the wooden trusses for hours.
The crew entered. Smoke conditions deteriorated and the roof began to sag. Firefighter David Eckermann later recalled, “When we went into the building, the smoke didn’t match what I was expecting.” But whatever he sensed couldn’t overcome the confidence the camera created. Eight minutes later, the roof collapsed. Four firefighters died that day.
“If AI performs every early draft, diagnosis, or analysis, where will the next generation of experts come from?”
The technology had not malfunctioned. The camera worked as designed, and the firefighters followed their training. The deeper failure developed gradually. As thermal cameras became standard, departments spent less time teaching firefighters to read smoke without them. When veterans retired, some of that expertise retired with them. The tool improved immediate performance while quietly weakening the human capability surrounding it.
Organizations face the same trade-off with AI. Junior workers finish tasks faster while losing the apprenticeship through which judgment once formed. If AI performs every early draft, diagnosis, or analysis, where will the next generation of experts come from?
After Houston, departments introduced camera-blackout drills to preserve the skills firefighters needed when the technology went dark. Organizations adopting AI need their own version: identify the signature thinking skills people cannot afford to lose, then deliberately preserve opportunities to practice them. The alternative is to become strangers to our own expertise—confident in our tools and certain of our data, right up until the moment the roof collapses.
3. Build a cognitive gym.
Before industrialization, physical exertion was woven into ordinary life. Then machines replaced human muscle. Technology made physical capability less necessary for work, but no less necessary for health. Exercise became a deliberate reintroduction of resistance that modern life had removed. AI may now require us to do the same thing for our minds.
We live in a culture obsessed with eliminating friction: seamless, effortless, instant. But productive struggle is not a bug in human cognition. It is how memory, understanding, and mastery form. A preliminary MIT study offers striking evidence. Researchers followed 54 people as they wrote essays unaided, with search, or with ChatGPT. Brain connectivity declined as external assistance increased. Minutes after writing, 83 percent of ChatGPT users struggled to recall a direct quotation from their essays. None of the attempted quotations were accurate.
In a small crossover group, participants who began with AI still showed reduced engagement after it was removed. Those who struggled unaided first and added AI later retained stronger connectivity. The pattern is clear: struggle first, assistance second.
“Productive struggle is not a bug in human cognition.”
That is the purpose of a cognitive gym: deliberately restoring the intellectual resistance that AI removes from ordinary life. Ask what the task is for. If the objective is execution, efficiency may be appropriate. If the objective is mastery, removing the struggle defeats the purpose. Using AI to build a skill is like bringing a forklift to the gym. Yes the weights go up and down, but it misses the point entirely.
4. Beware of the Dream Team fallacy.
We often assume that combining a capable human with a sophisticated machine will produce better decisions. Human judgment plus machine precision. The best of both worlds. I call this the Dream Team fallacy.
Imagine a doctor examining an X-ray, then receiving an AI assessment. Now the doctor must interpret the image, evaluate the AI, and decide how much weight to give each. The human brain isn’t designed to process multiple streams of diagnostic information simultaneously without triggering predictable shortcuts.
When the assessments agree, the doctor may mistake agreement for independent confirmation and overlook their shared evidence. When they disagree, the AI’s answer can become an anchor, pulling the doctor away from a sound initial judgment. Advanced technology doesn’t just enhance capabilities. It enhances confidence. And enhanced confidence, in the wrong circumstances, becomes enhanced stubbornness.
Human-machine partnerships aren’t combinations; they’re reactions. And the nature of that reaction depends on design choices that often seem trivial. One prescription is to deliberately add friction: require people to form their own judgment before revealing the AI’s recommendation, then give them time to examine any disagreement.
We are creatures of habit, prone to shortcuts, quick to anchor on first impressions. The problem isn’t that we have human quirks. The problem is that we too often design as if we don’t. The ultimate promise of this new era is a partnership that makes us think better. One built for the humans we actually are, not the ones we wish we were.
5. AI can scale learning—or merely the appearance of it.
In 343 BCE, Philip II offered to rebuild Aristotle’s destroyed hometown if Aristotle would tutor his son Alexander. For centuries, personalized education remained largely the privilege of people who could afford their own Aristotle.
In 1984, psychologist Benjamin Bloom quantified its value. Students receiving one-on-one tutoring performed two standard deviations better than students in conventional classrooms. Bloom called the challenge the two-sigma problem: we knew personalized tutoring worked; we could not provide every student with an Aristotle. AI may change that. But access does not guarantee learning.
“What we cultivate, we keep.”
Turkish high school students received unrestricted AI access while practicing mathematics. Their performance rose 48 percent. On a closed-book exam without AI, however, they scored 17 percent below students who had practiced without it. Their confidence remained high. They had borrowed capability, not built it.
In Nigeria, public-school students used a freely available AI system in pairs for 12 sessions over six weeks. Teachers received brief training, and prompts asked students to explain their reasoning rather than simply giving answers. The students gained the equivalent of one and a half to two years of regular schooling, at a cost of $ 48 per student.
The difference was not access to AI. One system optimized for completing the work. The other preserved the questioning and productive struggle through which learning happens. AI does not solve the two-sigma problem by making answers cheaper. It solves it by making good questions, timely hints, and personalized feedback scalable.
Whether we get a tutor or an answer machine depends on how the interaction is designed—and how we choose to use it. AI may finally give everyone an Aristotle, but only if we ask it to teach us, not answer for us. What we cultivate, we keep. What we outsource, we lose.
Enjoy our full library of Book Bites—read by the authors!—in the Next Big Idea App:
