In 10 Minutes, We Learned to Give Up · Issue #163

Opening

Dear reader, Socrates worried about writing. He feared that once people started putting things down on paper, they’d stop bothering to remember. When the telegraph arrived, some predicted poetry would die. When calculators spread, people worried mental arithmetic would vanish. Then Google showed up.

It’s the same story every time. Every new tool brings the same fear — “this is going to wreck our brains” — and for the most part, the catastrophe never arrives. So when people say “AI is making us stupid” these days, it sounds like the same old script.

But when I went through the recent studies one by one, I couldn’t shake the feeling that this time might actually be different. Let me give you the conclusion up front. Unlike search engines or calculators, AI is the first tool that substitutes for thinking itself. And the early research is already sounding a small but unmistakable alarm.

🧩 Can We Keep Saying “It Turned Out Fine”?

In 2011, a Columbia University team ran a fascinating experiment. When people were told a piece of information but also told “you can look this up on a computer later,” they remembered where to find it better than they remembered the information itself. The researchers called this the “Google Effect.” We had quietly started outsourcing memory to Google.

Back then, too, worry and pushback were evenly matched. One camp argued that “Google is making us stupid” (a famous 2008 Atlantic cover story). The other argued that the time we saved from digging through libraries could go toward deeper thinking. Calculators followed a similar pattern: mental arithmetic may have weakened, but calculators let us tackle far more complex math.

That’s why so many people compare AI to the calculator. Sam Altman is one of them. But someone directly rejects that analogy: Nataliya Kosmyna of MIT, whose work is the most cited study on AI and cognitive ability. “You don’t fall asleep hugging a calculator and wake up with it,” she says. “You don’t pour your innermost worries out to a calculator, either.” The nature of the tool itself, she argues, is fundamentally different.

📊 What Happened in the Lab

Let’s start with Kosmyna’s team’s experiment. They split 54 participants into three groups to write essays: one group used a large language model¹ like ChatGPT, another used Google search, and the last group used nothing but their own heads. When they measured brain activity with EEG², the “brain-only” group showed the strongest neural connectivity, and the LLM group showed the weakest. What’s more, the people who wrote with AI couldn’t even properly quote the sentences they had just written, and they felt the least ownership over their own essays. And the gap only widened the more sessions they did.

This isn’t confined to the lab. A Wharton School team gave roughly 1,000 high school students in Türkiye an AI math tutor. One version was a plain ChatGPT-like tool; the other was a “guardrailed” version that offered hints instead of straight answers and incorporated teacher-made solutions and common-error notes. During practice, both groups did well — students using the guardrailed version got 127% more practice problems right. But once the AI was taken away for the actual test, the picture flipped. Students who had used the plain ChatGPT-style tool performed worse than students who had used no AI at all. The skill they’d briefly borrowed on the practice sheets simply wasn’t there when they had to stand on their own.

The chilliest finding comes from a recent study by Grace Liu’s team at Carnegie Mellon University. Participants solved fraction problems; one group could use AI help for the first 12 problems but had to solve the last 3 alone. The AI-assisted group did well on the first 12 — but on the last 3, they made more mistakes and gave up more often. It took just 10 minutes for this shift to appear. The researchers described today’s AI this way: “A mentor doesn’t just hand you answers. A mentor designs learning and puts your growth ahead of an immediate result. Today’s AI, by contrast, is a shortsighted collaborator optimized to give immediate, complete answers. It never once says ‘no.’”

It’s a familiar pattern. We’ve actually already been through something similar with GPS. A 2020 McGill University study found that the longer people had relied on GPS, the worse their spatial memory was when navigating without it. When researchers measured again 3 years later, those who had used GPS more in the interim showed a steeper decline in spatial memory.

What about creativity? A Georgetown University team compared more than 370,000 college application essays before and after ChatGPT’s arrival. Interestingly, essays touched by AI used more varied, ornate vocabulary. But the ideas inside them converged toward sameness. The language got richer while the thinking narrowed to a single point. Adam Green, who led the study, put it precisely: “Google helps me find what I was already looking for. AI decides what to look for on my behalf.”

Here we need to hit the brakes for a moment. Most of these studies have small samples, short observation windows, and many haven’t yet gone through peer review³. Liu herself draws a clear line: “Ten minutes of use doesn’t cause long-term brain changes or cognitive decline. What happens with repeated, long-term use is still an open question, and answering it requires long-term follow-up studies.” So it would be premature to conclude from today’s data that “AI is ruining our brains.” Still, it’s hard to dismiss the fact that different teams, running different tasks, keep picking up signals pointing in the same direction.

So What’s Actually Different This Time

Here’s the decisive difference between AI and every tool before it. In Green’s words, AI is the first technology that “comes up with ideas on our behalf.” Writing substituted for memory. Calculators substituted for computation. Google substituted for search. All of them handled the raw material of thought. But AI substitutes for thinking itself — for deciding what to think of and how to connect it.

The problem is that some people haven’t developed this ability yet in the first place. Michael Gerlich of Swiss Business School warns: “There’s a real risk that younger generations never learn critical thinking to begin with. Because AI conveniently thinks for them, they may never develop the capacity at all.” Even people who already have the skill aren’t safe. A 1971 study of pilots found that even after long stretches without flying, hand-eye coordination held up fine — but the “cognitive” skills, like recalling procedures and mentally tracking the plane’s position, dulled quickly. Muscle memory sticks around. The muscle of thought atrophies if you stop using it.

Google’s precedent is worth chewing on, too. Did Google actually make us smarter or freer? Not really. The line between work and life blurred, and attention spans shortened. IQ scores, which had risen by about 3 points every 10 years throughout the 20th century, began falling across several measures after 2006 — and the steepest drop was among 18-to-22-year-olds, precisely the most digitally native generation.

There’s one hopeful clue, though. Researchers see shortened attention spans less as a permanent change in brain structure and more as a matter of habit. That means if we clear away the distractions we ourselves have installed, we can retrain ourselves to focus for longer stretches. If the ability hasn’t vanished but has simply gone unused, then there’s just as much room to bring it back.

Oswald’s Lens

Honestly, I’m a bit skeptical of both sides on this one.

As someone who’s worked with data, I think the “AI is ruining our brains” headline still rests on thin evidence. You can’t declare humanity’s future based on experiments with a few dozen participants, 10-minute observation windows, and pre-peer-review papers. Fear always sells faster than data. But at the same time, the optimism of “calculators were fine, so this will be too” strikes me as lazy. When different studies keep pointing in the same direction, that’s more likely a pattern than a coincidence.

Building AI tools myself, I’ve confirmed one thing over and over. Today’s AI is designed to give “immediate, complete answers.” That’s the product’s core virtue — users don’t wander, don’t get stuck, and get their result right away. But learning sits exactly on the other side of that. It’s in wandering, getting stuck, and making mistakes that the muscle of thought gets built. As Liu’s team put it, AI never says “no.” That’s not a flaw in AI — it’s a design choice. And design choices can be changed.

So the conclusion I’ve landed on isn’t “whether or not to use AI.” It’s consciously choosing what to keep in my own hands and what to hand off. Kosmyna says she uses the AI she built only for research, and proudly doesn’t use language models in her personal life. Green says, “respect the blank page.” I try to switch AI off, too, for that specific moment when I’m drafting something. I might lose a bit of speed in the output, but I want to keep the feel of actual thinking as my own.

Closing

Let me sum up. First, the fear that “new technology ruins our brains” has always existed, and it has mostly proven wrong. Second, but AI is the first tool to substitute for thinking itself, beyond search and calculation, and early studies are already sounding small alarms about persistence, creativity, and critical thinking. Third, it’s still too early to draw firm conclusions — which is exactly why we need to decide now what abilities we want to keep as our own.

Just for today, why not pick one problem you’re stuck on and see it through to the end without AI? The exact point where you want to give up is, in fact, where the muscle of thought gets built.

💬 Is there an ability you’ve already handed over to AI? I think I’ve completely surrendered my sense of direction to GPS. If there’s something you’d say “I will never hand this off to AI,” tell me why in the comments — I might use it for a future issue.

💬 Share your thoughts on the question above in the comments · 📨 If you know someone who’s been saying “my mind isn’t what it used to be,” quietly pass this along

References & Further Reading

Primary sources

Nataliya Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task”, MIT Media Lab, 2025.
Hamsa Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics”, PNAS, 2025.
Grace Liu et al., “AI Assistance Reduces Persistence and Hurts Independent Performance”, 2026 (preprint).
Louisa Dahmani & Véronique Bohbot, “Habitual use of GPS negatively impacts spatial memory during self-guided navigation”, Scientific Reports, 2020.
Adam Green et al., “The Link Between Diverse Words and Original Ideas Is Weakening in the AI-Era College Admissions”, PsyArXiv preprint (Georgetown University), 2026.
Michael Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking”, Societies, 2025.

Background

Betsy Sparrow, Jenny Liu & Daniel Wegner, “Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips”, Science, 2011. : This is the original paper on the “Google Effect.”
Nicholas Carr, “Is Google Making Us Stupid?”, The Atlantic, 2008. : Reading this feels like déjà vu for today’s debate — it’s 17 years old, yet it reads the same now.

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

Large language model (LLM): AI trained on vast amounts of text to generate sentences, like ChatGPT. It’s the core engine behind what we commonly call “generative AI.” ↩
EEG (electroencephalography): A method that attaches sensors to the head to measure the brain’s electrical activity. It lets researchers see how actively different regions are working together. ↩
Peer review: The process by which experts in the same field verify a paper’s methods and conclusions before formal publication. A paper that hasn’t yet gone through this process is called a “preprint.” ↩