—
Ask anyone who has relied on turn-by-turn directions for a few years to sketch a map of their own neighborhood, and the gaps tend to show up quickly. Researchers have a name for this habit of handing mental work to an outside tool. They call it cognitive offloading, and decades of studies on GPS units and calculators offer a useful baseline now that conversational AI can take over reasoning and writing too. Those studies also raise a practical question about whether every AI tool carries the same risk, or whether AI designed to strengthen human thinking rather than replace it might lead somewhere different. The evidence so far points to an answer that is more specific than the headlines suggest.
Offloading is an old habit with mixed results
What counts as cognitive offloading
Psychologists Evan Risko and Sam Gilbert described cognitive offloading in a 2016 review in Trends in Cognitive Sciences as using physical action to reduce the mental demands of a task. Writing a grocery list counts. So does tilting your head to read sideways text. Offloading in this sense is ordinary and often smart, because it frees limited working memory for other things.
What happens to information we hand off
The catch is that whatever gets offloaded tends not to be stored internally. A 2011 study in Science by Betsy Sparrow, Jenny Liu, and Daniel Wegner found that people who expected a computer to save information were less likely to remember the information itself. They were better at remembering where it was kept. The brain adjusted its filing system to fit the tool.
Calculators and the question of timing
Calculators show that the outcome depends on timing. A widely cited 1986 meta-analysis by Ray Hembree and Donald Dessart in the Journal for Research in Mathematics Education found that calculators used alongside regular instruction generally did not erode paper-and-pencil skills, and in many cases improved them. The notable exception was fourth grade, where sustained calculator use appeared to hurt basic skills. Children at that age were still building the arithmetic fluency the device replaced. Offloading a skill you already have works very differently from offloading one you have not yet learned.
What heavy GPS use did to spatial memory
The clearest example of skill erosion comes from navigation. In a 2020 study published in Scientific Reports, McGill University researchers Louisa Dahmani and Véronique Bohbot measured the lifetime GPS use of 50 regular drivers and tested several aspects of their spatial memory in virtual navigation tasks. Drivers with more GPS experience had worse spatial memory when they had to find their own way.
The team retested a smaller group about three years later. Those who had used GPS more in the meantime showed a steeper decline in spatial memory that depends on the hippocampus, the brain region that builds mental maps. The heavier users also did not report a weaker sense of direction to begin with. That detail suggests the GPS use came first and the memory decline followed, rather than people with poor navigation skills turning to GPS to compensate. The authors acknowledge the follow-up sample was small, so the finding is a strong signal rather than settled fact.
The mechanism is easy to picture. Turn-by-turn directions let a driver follow a sequence of instructions without ever forming a sense of how places relate to each other. The skill stops getting exercised, so it weakens.
The same pattern is now appearing with conversational AI
AI use and critical thinking scores
Studies of generative AI are newer and smaller, but they point the same way. In a January 2025 paper in the journal Societies, Michael Gerlich of SBS Swiss Business School surveyed and interviewed 666 people. He found a significant negative correlation between frequent AI tool use and critical thinking scores, and the relationship ran largely through cognitive offloading. Participants aged 17 to 25 relied on AI the most and scored the lowest. Gerlich has been careful to point out that the study shows correlation rather than causation.
Confidence and effort among knowledge workers
Researchers from Microsoft Research and Carnegie Mellon University reached a related conclusion at the CHI 2025 conference. Hao-Ping Lee and colleagues surveyed 319 knowledge workers who use generative AI at least weekly and gathered 936 examples of real work tasks. Workers with more confidence in AI reported less critical thinking. Workers with more confidence in their own abilities reported more. The researchers also found that AI changed the kind of thinking people did, moving it away from producing work and toward checking and integrating what the AI produced.
Brain activity during AI-assisted writing
The most widely discussed study came from the MIT Media Lab. Nataliya Kosmyna and colleagues used EEG to record brain activity in 54 participants who wrote essays over several months. One group used ChatGPT, one used a search engine, and one used no tools. The ChatGPT group showed the weakest neural connectivity and the brain-only group the strongest. In a final session, participants who switched from ChatGPT to writing unaided showed weaker engagement and weaker recall, a lingering effect the authors called cognitive debt. The paper was released as a preprint, and Kosmyna has pushed back on coverage that framed it as proof of brain damage.
Skill atrophy, not clinical decline
None of this research shows cognitive decline in the clinical sense. What it shows is closer to skill atrophy, the predictable result of not practicing something. That distinction matters because it ties the effect to how a tool gets used, and how a tool gets used depends partly on how it was built.
Does the AI’s design change the outcome?
If all AI use led to the same dependency, design would be irrelevant. One of the stronger studies on the question suggests otherwise. Hamsa Bastani, Osbert Bastani, Alp Sungu, and colleagues ran a field experiment with nearly 1,000 high school math students in Turkey, published in PNAS in 2025. One group used a standard ChatGPT-style interface the researchers called GPT Base. Another used GPT Tutor, which ran on the same model with prompts built to safeguard learning by offering hints rather than complete answers. A control group had no AI.
During practice sessions, both AI groups pulled well ahead of the control group. Grades rose 48 percent with GPT Base and 127 percent with GPT Tutor. Then access was removed for an exam. Students who had practiced with GPT Base scored 17 percent lower than students who never had AI at all. The GPT Tutor group largely avoided that drop. The researchers traced the difference to students leaning on the unrestricted version as a crutch.
Why tool behavior matters
The underlying model was the same in both groups. What changed was how the tool behaved, and that was enough to change what students learned. Lee’s team arrived at a similar view from the workplace side, arguing that generative AI tools should be built to support critical thinking by addressing users’ awareness, motivation, and ability. Read alongside the calculator and GPS research, the lesson is that dependency is shaped by what a tool asks of the person using it.
One example of the design approach
A small number of developers are now building with that variable in mind. Vesela, a conversational AI app from Gray Sky AI, is one example. The company frames its goal around how a person thinks once a conversation ends, aiming for more clarity and agency, instead of around the answer delivered. In practice the app often responds to a dilemma with a question that sends the user back to their own reasoning, and its stated design principle keeps the user’s judgment at the center. No published study has yet tested whether this approach produces the protective effect seen with GPT Tutor. For now it stands as an example of a design philosophy that treats dependency as a problem to solve, which is where the research says the work should begin.
What the evidence supports right now
Taken as a whole, the research does not support the claim that AI is making people less intelligent. It supports something narrower. Skills that go unpracticed weaken, and tools that do the practicing for us speed that process. The effect appears strongest when people are still learning a skill or trust the tool more than themselves. It also seems to grow when a tool hands over finished answers by default.
Where the evidence is still thin
Much of the AI research is still early. Gerlich’s and Lee’s studies are correlational and rely partly on self-report. The MIT study is small and appeared before peer review. The Bastani experiment is the strongest causal evidence available so far, and it happens to be the one showing that design can largely prevent the harm.
Habits that keep the skill in use
Bohbot has offered a simple fix for GPS users. She suggests using it on the way to a new place and switching it off for the trip home. The same logic carries over to AI. Working through a problem first and bringing in the tool second keeps the skill in use, and tools that respond with questions do some of that work on their own. Neither habit requires giving up AI. Both keep the person doing the part of the work that builds the skill.
—
