AI Makes Thinking Effortless, and Researchers Say That’s the Cost

Daily use of generative AI climbed to 95.7% among people tracked in a new three-wave study on human problem-solving, up from 76.2% when researchers started measuring. Confidence in double-checking the machine’s answers fell over the same stretch, and accuracy on the hardest problems in the study stayed capped at 50%, no matter how often people leaned on the tool. It is the newest data point in a fast-growing pile of research asking whether AI is making people dumber, and so far, the findings are not reassuring.

The old worry, dating back to a 2011 study, was that Google made people lazy researchers, content to remember where a fact lived instead of the fact itself. AI raises a sharper version of the same question. A chatbot generates a finished thought and hands it over ready-made, a step beyond simply pointing someone toward stored information, and a run of studies published over the past year, from the Massachusetts Institute of Technology’s (MIT) Media Lab to Microsoft’s own research division, suggest that convenience carries a cost measured in memory, persistence and the shrinking will to check the machine’s homework.

The Frictionless Trap

The clearest picture so far comes from MIT’s Media Lab, where a team led by researcher Nataliya Kosmyna wired 54 students from five Boston-area universities into electroencephalography (EEG) headsets, a setup that tracks brainwave activity, and followed them through four months of essay writing. One group wrote with ChatGPT doing the heavy lifting. A second used Google search for research. A third worked from memory alone, no tools at all.

By the end, the chatbot group showed the weakest brain connectivity of the three, the thinnest memory of what they had supposedly just written, and a fading sense that the essay was even theirs. Eighty-three percent of them could not accurately quote a single line from the essay they had turned in minutes earlier. Take the tool away and have the same students write unaided, and the gap did not close: the effects lingered even after ChatGPT use stopped. Researchers labeled the pattern cognitive debt, borrowing brainpower now and paying interest on it later. The paper, posted through MIT Media Lab’s Your Brain on ChatGPT study, put its own conclusion bluntly: over four months, LLM users “consistently underperformed at neural, linguistic, and behavioral levels.”

Some researchers now argue that the industry’s push to strip away friction, the very thing that makes these tools feel so good to use, is the same design choice weakening the muscle they are meant to support. That concern has pushed a number of AI researchers past just critiquing the chatbot format. Some are walking away from conversational chatbots to build so-called world models, systems designed to simulate physical reality rather than just predict the next word in a reply.

The Verification Gap Widens

A study posted this year tracked the same group of people across a three-wave longitudinal study on AI, metacognition and what its authors call the verification bottleneck, and found the trade-off sharpening as AI use spread rather than leveling off.

• 76.2% to 95.7% – the jump in daily AI use recorded across the study’s three waves.
• 65.0% to 59.1% – the decline in participants’ confidence that they could correctly verify AI’s answers to complex problems.
• 46.7% to 50.0% – the range where actual accuracy on the study’s hardest problem stayed stuck the entire time, even as perceived efficiency rose.

Microsoft’s own researchers, working with Carnegie Mellon University, found a related pattern among knowledge workers rather than students. A survey of workers in business, education, arts, administration and computing found that those who most trusted the accuracy of AI assistants thought less critically about those tools’ conclusions.

The data shows a shift in cognitive effort as knowledge workers increasingly move from task execution to oversight when using GenAI. Surprisingly, while AI can improve efficiency, it may also reduce critical engagement, particularly in routine or lower-stakes tasks in which users simply rely on AI, raising concerns about long-term reliance and diminished independent problem-solving.

The Microsoft and Carnegie Mellon University researchers wrote that in a paper presented at the CHI Conference on Human Factors in Computing Systems, a study that had not yet completed formal peer review.

Who Loses the Most Ground

Age shows up as a dividing line in more than one dataset. Michael Gerlich, head of the Center for Strategic Corporate Foresight and Sustainability at the Swiss Business School, is one of the researchers studying this risk most closely. In a survey of 666 participants, Gerlich found a significant negative relationship between how often people used AI and how well they scored on critical thinking measures, with cognitive offloading, outsourcing mental effort to a tool, doing the mediating. Younger participants fared worse than older ones in that same research: those with heavier AI dependence in the sample also scored lowest on critical thinking.

A separate paper, a preprint titled AI Assistance Reduces Persistence and Hurts Independent Performance, found a similar shape of trade-off in a different setting: performance improves while the tool is available, then drops once it’s taken away, a dynamic its authors say now shows up across nearly every reasoning domain they’ve tested, feeding concerns about overreliance and deskilling. One of the researchers behind that work, Grace Liu, a machine learning PhD candidate, has told Business Insider the effect showed up after just ten minutes of AI use in her experiment, and that whether it compounds after years of everyday use remains an open question. Business Insider has also reported on a Wharton School of the University of Pennsylvania study that gave AI math tutors to 1,000 high school students in Turkey; those who used a standard, ChatGPT-like tool did worse once it was taken away than students who never had AI access for the work at all.

Researchers keep pointing to the same kinds of settings where the pattern takes hold.

• Classrooms, where guardrail-free tutoring bots let students skip the struggle that builds a skill in the first place.
• Entry-level jobs, where new hires may never practice fundamentals older colleagues learned before AI existed.
• Everyday research, where a generated summary replaces the friction of reading a primary source start to finish.

Did the Loudest Warning Overstate Its Case?

The MIT paper did not stay a quiet academic release for long. It swiftly went viral, inciting alarmist headlines and backlash. Outlets compressed a preprint running past two hundred pages into blunt declarations, among them that “ChatGPT erodes critical thinking skills” and that using the tool can “rot your brain.” Kosmyna’s own lab pushed back on that framing, asking journalists specifically to avoid describing the results with words like terrifying, stupid, dumb, brain rot or harm.

The findings came out ahead of formal peer review, and the lead author has said that was deliberate. She worried that waiting quietly through review would let policy get set in the meantime. “What really motivated me to put it out now before waiting for a full peer review is that I am afraid in 6-8 months, there will be some policymaker who decides, ‘let’s do GPT kindergarten.’ I think that would be absolutely bad and detrimental,” Kosmyna told Time magazine, according to The Hill.

Other academics flagged a design problem in the experiment itself. The change in neural connectivity of the brain-only group over the first three sessions was likely the result of becoming more familiar with the study task, a phenomenon known as the familiarisation effect. As study participants repeat a task, they grow more efficient and their strategy adapts. The group that moved from AI to working unaided, meanwhile, only got one crack at working solo instead of three, which limits how far the comparison can be pushed; to fully justify the study’s claims, those participants would need three unaided sessions too. The topline numbers still stand. The one-line version, that ChatGPT rots your brain, oversold what a 54-person pilot study could prove on its own.

A Familiar Panic, With One New Variable

None of this is the first time a new tool has triggered a memory panic. Socrates worried that writing would hollow out memory. Calculators were supposed to end mental math. Search engines got the same treatment about two decades ago, and that one came with data: a 2011 study found frequent search engine users didn’t retain information gleaned through online searching as well as people who learned it offline, remembering instead where to find it again.

What’s different this time, several researchers argue, is scope. A search engine helps someone find an answer they already meant to look for. A chatbot generates the reasoning itself and hands over a finished product, which one MIT Media Lab study described as “excessive reliance on AI-driven solutions” contributing to “cognitive atrophy” and shrinking critical thinking abilities, a harder habit to build resilience against than remembering a search term.

Researchers Are Testing Friction as the Fix

If frictionless design is the mechanism, the proposed fixes increasingly involve putting some friction back. An increasing volume of research indicates that overdependence on AI tools for cognitive tasks results in observable skill deterioration, a phenomenon researchers call cognitive offloading. One paper on the psychological risks of emotionally attuned AI companions, a paper describing what it calls the Cognitive Atrophy Hypothesis, argues that systems built to keep users engaged and satisfied end up quietly wearing down the resilience they are supposed to support.

Its proposed countermeasure, which the authors call Stoic Architectures, would have an AI system deliberately introduce difficulty once it detects a user leaning on it too heavily, intentionally shifting from comfort to challenge to protect user resilience rather than mere satisfaction. It’s a familiar shape of problem. Oton Technology has reported before on how nutrition apps built around gamification carry their own hidden behavioral risks, another instance of software optimized for engagement producing side effects nobody designed on purpose.

What’s missing across nearly every study published so far is time. Researchers studying AI chatbots and cognitive health have called for longitudinal work that follows the same people for years, not months, to separate a temporary dip from something that doesn’t come back.

The daily-use number will keep climbing past 95%. Nobody yet has a solid number for how many people still trust themselves enough to check the machine’s work.

Frequently Asked Questions

Is there proof that AI use permanently damages the brain?

Not yet. The longest published trial, MIT’s four-month essay-writing study, found effects that lingered after participants stopped using ChatGPT, but the lead researcher released the results ahead of formal peer review and has pushed back against reductionist interpretations of the findings.

What is cognitive debt?

Cognitive debt is the term MIT Media Lab researchers used to describe what they measured in the ChatGPT group of their essay-writing study, brainpower borrowed in the moment that shows up later as weaker memory and brain connectivity.

Does using AI always hurt critical thinking?

No. Michael Gerlich, whose 666-person survey found the clearest link between AI use and falling critical thinking scores, has suggested AI is not inherently harmful to cognitive ability, arguing that how a tool gets used matters as much as whether it gets used at all.

Who is most at risk of losing skills to AI?

Younger, heavier users score worst so far. Gerlich’s research found younger participants showed both higher AI dependence and lower critical thinking scores, and separate research on persistence describes a related deskilling risk for people who lean on AI before they’ve built a skill independently.

How is this different from the old worry that Google made us dumber?

Search engines mainly changed where people store memories, letting them recall a source instead of a fact. AI chatbots generate the reasoning itself, which researchers using the extended mind thesis describe as the tool becoming an active contributor to thinking rather than a passive filing cabinet the way a search engine or notebook works.

What can people actually do about it?

Researchers say the fix isn’t necessarily quitting AI outright, but being deliberate about which skills to protect and practice independently, since offloading small, repetitive tasks can free up effort for bigger ones without automatically costing a person the brainpower needed to do them.