Is AI Making Us Worse Thinkers? Early Research Suggests It Might Be

It's too soon to say AI is frying our brains. But the studies so far, on memory, persistence, and critical thinking, don't look reassuring.

My family is in its natural state when we're arguing. Not just about what should happen, but about what already did.

Growing up, before my grandparents finally wired their home for the new millennium, their full set of encyclopedias was the only way to settle a dispute. Plenty of arguments went unresolved, especially about anything that happened after the books were printed. Then Google arrived. All that access didn't just shrink the gap between what we could know in seconds and what we'd have to wonder about for months. It changed how the world remembers.

Here's the thesis, up front: AI probably won't fry anyone's brain overnight. But the early evidence suggests that handing our thinking to a chatbot can quietly weaken the things we most want to keep sharp: memory, persistence, and the ability to solve a problem on our own. The science isn't settled. The experiment, though, is already running on all of us.

Key Takeaways

• An MIT study found people who used generative AI to write essays performed worse over time than those who used Google or no aid at all.
• The risk looks steepest for people learning a skill for the first time, not experts adding AI to work they already know.
• The fix isn't abstinence. It's deciding, deliberately, which skills to keep independent and which you're fine offloading.

What the "Google effect" taught us

Researchers gave the pattern a name back in 2011: the "Google effect." When people knew they could look something up again easily, they remembered where to find the information better than they remembered the information itself. "We are becoming symbiotic with our computer tools," the researchers wrote, "remembering less by knowing information than by knowing where the information can be found."

The reaction split, and it should sound familiar. An Atlantic cover story asked whether Google was "making us stupid." Others argued the opposite: that Google democratized knowledge and traded hours in library stacks for supercharged thinking. So who was right? Sort of both, and that's the uncomfortable part. We didn't get measurably smarter or freer for deep work. Instead, the boundary between work and home dissolved, and many of us now work longer hours than before the internet existed.

Why AI isn't just another calculator

Every new tool has triggered the same fear, and the calculator comparison is the one AI's boosters reach for most, Sam Altman included. Nataliya Kosmyna, the MIT researcher behind one of last year's most-cited studies on AI and cognition, calls that comparison a fallacy. "You don't fall asleep and wake up with a calculator," she says. "You don't talk to the calculator about everything you have in your mind."

That's the difference in one line. Socrates worried writing would rot our memories. Some thought the telegram would kill poetry. The calculator was supposed to erase mental math. None of those collapsed the way the doomsayers feared, though each did shift things gradually, across decades, alongside smartphones, looser grading, and school budget cuts. Fewer than 10% of people read poetry now. The share of 12th graders scoring at least basic math proficiency has fallen from its 2013 high, per the National Assessment of Educational Progress. AI is different in kind, not degree. It's the first tool that doesn't just fetch what you thought of. It thinks of things for you.

What early AI studies found

The research is new, much of it not yet peer-reviewed, and the samples are small. But the findings keep pointing the same direction. Here's the evidence stack, study by study.

MIT, essay writing (Kosmyna). People given generative AI to write essays performed worse over time than those who used Google or wrote with no aid at all. Limitation: measures a specific writing task, not general intelligence.

Wharton, math tutors, 1,000 Turkish high schoolers. Two AI tutors: one a standard ChatGPT-style answer machine, one with guardrails that gave hints and teacher-sourced explanations instead of direct answers. Students did better with both while they had access. But once the tools were taken away, the group that used the standard ChatGPT-style tool did worse than students who never had AI at all. Limitation: short-term test, single subject.

McGill, GPS and spatial memory (2020). The longer people leaned on GPS, the worse their spatial memory when they had to navigate without it. A three-year follow-up with a small sample found the decline got steeper. Limitation: small follow-up cohort, correlational.

Carnegie Mellon, math persistence (April 2026 preprint, Grace Liu). Participants solved fraction problems. One group had an AI that would hand over answers for the first 12 questions, then had to do the last 3 alone. The AI group solved more of the first 12, but were more likely to skip the final 3 and less likely to get them right. The effect showed up after just 10 minutes. "Current AI systems are fundamentally short-sighted collaborators," the authors wrote, "optimized for providing instant and complete responses, without ever saying no." Limitation: 10 minutes tells us nothing about long-term change, and Liu says so herself.

Georgetown, creativity in 370,000+ college essays. Human-written essays carried more genuinely new ideas than AI-assisted ones, though AI essays used more unique language. So AI can sharpen the wording in a single piece while flattening the range of ideas across a whole group. Limitation: application essays are a narrow creative form.

Correlation, not proven causation. That distinction matters, and none of these studies closes it. What's striking is that so many independent designs, aimed at different skills, keep surfacing the same shape: help while it's there, a dip once it's gone.

Who is most at risk?

The clearest dividing line isn't age, it's whether you learned the skill before you met the tool. "There is a high risk if young people never learned this critical thinking, because they have the convenience of an AI thinking for them," says Michael Gerlich of the Swiss Business School. There's a real difference between spending a decade coding and then adding AI to your work, versus only ever learning to "vibe code."

But experience doesn't make anyone immune. Skills need upkeep. A 1971 study of pilots found they held onto hand-eye coordination fairly well after four months without flying, yet their cognitive skills slipped: recalling procedural steps, visualizing the plane's position, running mental calculations. "Any skill that we acquired as humans does need to be retained and retrained," Kosmyna says. Use it or lose it applies to the expert and the novice both. The novice just has less to lose it from.

What we still don't know

Honesty check: the alarming studies are also thin. Ten minutes of AI use, as Liu is quick to say, "definitely wouldn't do some long-term brain or cognitive decline." What happens after months or years of daily reliance is a genuinely open question, and answering it needs longitudinal studies that no one has finished yet.

We also can't cleanly separate AI's effect from everything else already reshaping our brains. IQ scores climbed about three points a decade through the 20th century, then people started testing lower across several categories between 2006 and 2018, with the steepest drops among 18-to-22-year-olds, the most digitally native adults measured. Is that AI? Smartphones? Something else entirely? We don't know. Attention spans have shrunk too, but researchers think that's a habit, not a neurological rewiring. We can retrain focus, if we're willing to remove the distractions we've built into our own days.

How to use AI without outsourcing your thinking

There's no clean rule here, but the researchers I spoke to converge on one principle: choose, on purpose, what you let the machine do. "Prioritize the skills that we want to keep independently," Liu says, "and then the skills that we're OK with outsourcing and offloading." A few practical starting points, drawn from what these experts actually practice:

• Guard the blank page. Adam Green of Georgetown's Laboratory for Relational Cognition worries most about the creative thinking "we develop through practice." Draft first, then bring AI in to sharpen. Don't let it ideate before you have.
• Prefer scaffolding over answers. The Wharton tutor that gave hints, not solutions, didn't leave students worse off. Ask AI to coach you toward an answer, not to hand you one.
• Keep some things fully manual. Kosmyna says she "proudly" doesn't use large language models in her personal life. Pick your equivalent: navigation without GPS sometimes, math in your head, a first draft with nothing open but a cursor.
• Notice when you're skipping the hard part. The CMU study's warning sign was people quitting on difficulty. If AI is the reason you didn't push through, that's the moment to close the tab.