Is AI Making Us Think Less? What Research Says About Cognitive Offloading

AI can make you think less, but in a narrower way than the scariest headlines suggest. The research on AI and critical thinking points to one pattern: a chatbot makes it easy to skip the part of a task that would have built or tested your own understanding, and the cost shows up later, when the tool is gone. In a randomized trial run in 2023 at a high school in Turkey and published in PNAS in 2025, students who practiced with an unrestricted GPT-4 chatbot did worse on a later unaided exam than classmates without it, while those given a version built to offer hints instead of answers did about as well as classmates without it.¹

No study here measured long-term changes in general reasoning, and most are short, small or self-reported. What they do support is a modest routine you can start with your next prompt: attempt the task before you open the chat, ask for hints rather than answers when you are learning something, and check hardest the outputs you trust most.

Whether assistants speed work up is covered in the guide to which everyday tasks an AI assistant handles well, and where it slips, with more guides to everyday AI use.

Cognitive offloading is normal, and usually a good trade

Cognitive offloading is using something outside your head, such as a note, a calendar alert, a search engine or a chatbot, to cut the mental work a task needs. Evan Risko and Sam Gilbert, reviewing the research in Trends in Cognitive Sciences in 2016, describe it as an everyday way people get around the limits of perception and memory.²

Offloading works because of capacity. Your working memoryworking memory: The small mental workspace that holds and handles the information you are using right now, such as a phone number you are about to dial. Its capacity is limited, which is why juggling several things at once quickly overloads it.Full entry in the glossary holds only a little at once, so writing the dentist’s appointment into your phone, or letting GPS track the turns, frees attention for the conversation in the car. Risko and Gilbert also sketch a cost, as a prediction rather than a finding: if a tool seems more accurate than you, you rely on it, your own memory for the task gets less use, your confidence in it drops, and you lean on the tool even more.²

1. A tool, such as GPS or a chatbot, seems more accurate than your own memory or judgment for this task.
2. So you rely on it, which is often the sensible choice.
3. Your own memory and practice for that task get less use.
4. Your confidence in doing it yourself falls, which makes relying on the tool more likely next time. Risko and Gilbert present this loop as a prediction to test, not a measured result.

A 2021 series of three lab experiments with a pattern-copying task showed both sides of the trade. When offloading was made more costly, people offloaded less, worked more slowly and remembered the patterns better afterwards. Knowing that a memory test was coming did not always protect memory, but in the third experiment, people forced to offload as much as possible who expected the test almost fully made up the loss.³

Before you hand a task to a tool, decide whether you will need the result in your head afterwards. If you will, read what comes back with the goal of remembering it: in that third experiment, the goal almost fully offset the cost of heavy offloading.

AI and critical thinking at work: less effort, different checking

Knowledge workers say generative AIgenerative AI: AI systems that produce new text, images, audio or code in response to a request, by generating output that resembles the data they were trained on. Chatbots built on large language models are the best-known kind.Full entry in the glossary makes critical thinking feel like less effort, and the more they trust the tool for a task, the less critical thinking they report doing. That comes from a 2025 survey by researchers at Microsoft Research and Carnegie Mellon University, presented at the CHI 2025 conference.⁴

The researchers asked 319 people who use AI tools at work at least once a week to describe real tasks, and collected 936 of them; people said they applied critical thinking in about 60 percent. Six of the seven authors work for Microsoft, which sells AI assistants; the sample, recruited and paid online, skewed young and tech-savvy; and every answer was self-reportself-report: A measure in which people describe their own behavior, feelings or circumstances, usually by answering a questionnaire. When the same person supplies both of the things being compared, shared habits of answering can make the link between them look stronger than it is.Full entry in the glossary, which the authors note can blur less effort using the tool with less effort thinking.⁴

The more useful finding is where the thinking went. Effort moved from gathering information to verifying it, from solving problems to fitting the AI’s answer into the work, and from doing the task to overseeing it, a role the authors call stewardship. Workers who were confident in their own skill reported more critical thinking, not less.⁴ Put simply, the thinking changes job, from producing an answer to judging one, and people seem to judge more when they believe they could do the task themselves.

Picture a project manager asking a chatbot for the risks in a launch plan. If she knows the project, she notices the missing supplier delay; if she trusts the tool and the list looks complete, she forwards it. The older name for that second pattern is automation bias, which a 2010 review in Human Factors found in novices and experts alike, and which training or instructions did not prevent.⁵

A separate 2025 survey by Michael Gerlich of SBS Swiss Business School found a similar link in a broader group. Among 666 people in the UK recruited through social media, heavier AI users scored lower on critical thinking, which was partly self-rated. It was a one-time cross-sectionalcross-sectional: Describes a study that measures everyone in its sample at a single point in time. Because the possible cause and the outcome are recorded together, it can show that two things occur together but not which of them came first.Full entry in the glossary survey, so it cannot say whether AI use lowered the scores or weaker critical thinkers turned to AI more.⁶

The tasks you would happily wave through are the ones where checking has quietly stopped.

Whichever way that runs, the working lesson is the same: read your own confidence in the tool as a warning light. The tasks you would happily wave through are the ones where checking has quietly stopped.

When the chatbot does the practice, learning suffers

Practice builds skill only if you do the work yourself, and the high school trial in Turkey tested exactly that. Some classes practiced math with a standard GPT-4 chat, some with a version told to give teacher-written hints and never the answer, and some with their textbooks and notes only. Then everyone sat an exam without any help.¹

The study

Moderate evidence

Hints versus answers in about 50 Turkish math classes

During practice, the standard chat raised scores by 48 percent and the hint-only tutor by 127 percent, compared with textbooks alone. On the unaided exam that followed, students who had used the standard chat scored 17 percent lower than the textbook group, while the hint-only group scored about the same as it. Students who had used the standard chat did not think they had learned less.¹

The authors traced the harm to how the tool was used: students asked the standard chat for solutions and copied them, using it as a crutch that kept them from fully engaging with the material, while the hint-only group asked for help and tried answers themselves; the chatbot’s own math errors mattered less. The practice sheets looked better and the students felt no worse off, which is why this cost is so easy to miss. The main caveats are one school, one subject, a 2023 model and an exam soon after practice. The work was funded by units of the Wharton School, and the authors declare no competing interests.¹

A smaller randomized controlled trialrandomized controlled trial: A study that assigns participants to the treatment or the comparison group at random, so the groups start out alike and a difference in what happens next can be put down to the treatment. Random assignment makes the two groups comparable; it does not make the people in the trial representative of anyone else.Full entry in the glossary with 91 university students found a similar trade: those randomly given ChatGPT rather than Google to research nanoparticles in sunscreen experienced a lower cognitive load but produced lower-quality reasoning in their recommendations.⁷ Across 89 studies of university students, one 2026 systematic reviewsystematic review: A review that fixes its question and its rules for including studies in advance, then searches out every study that fits and weighs them together. Some systematic reviews pool the results into a meta-analysis; others describe what the studies found without combining the numbers.Full entry in the glossary reached a conditional verdict: its author proposes that generative AI amplifies thinking when teaching is structured around it and substitutes for it when use is unguided.⁸

At work the same choice appears whenever you learn something new, such as a spreadsheet function you will need every month. Asking for the finished formula gets today’s report done. Asking for the next step, trying it, then comparing with the full answer is slower, and it is the kind of help that left learning intact in the trial. If you will need the skill again, spend the extra minutes now.

Further reading

• 

  Make It Stick: The Science of Successful Learning

  by Peter C. Brown, Henry L. Roediger III and Mark A. McDaniel

  The research on learning through effort, such as trying a problem before you see the answer, that backs this article's try-first habit.

As an Amazon Associate WiserHours earns from qualifying purchases.

Try first, then decide what to hand over

The first habit rests on a randomized trial, the others on lab work, surveys and an observational study, so treat them as sensible defaults rather than proven fixes.

Still learning itAlready masteredQuick to verifySlow to verify

Quick to verify, Still learning it: Ask for hintstry first, then compare

Quick to verify, Already mastered: Offload freelyspot-check now and then

Slow to verify, Still learning it: Do it yourself firstthe AI answer is a check, not a start

Slow to verify, Already mastered: Offload, then verifyline by line before it leaves you

1. Try first, then ask for hints

Spend a few minutes on your own draft or answer before you open the chat. When the point is to build a skill, tell the assistant not to give the answer: ask for the next step or a hint. In the school trial, harm came from students asking for solutions and copying them, and the hint-only design left exam scores intact.¹

2. Decide what you need to remember

Sort what you must know from what you only need to find, and read the first kind with the goal of remembering it, the condition that protected memory in one of the 2021 lab experiments.³

3. Check hardest where you feel surest

This is the knowledge-worker survey turned into a habit. Pick one routine AI output a week, ideally the kind you usually forward unread, and check it line by line.

4. Keep some reps unassisted

Skills may fade when a tool does them for you. In a 2025 observational before-and-after study at four endoscopy centers in Poland, doctors who had started using AI to spot precancerous growths detected fewer of them in the colonoscopies they then did without it.⁹ Do some of the work the slow way on purpose, such as drafting one reply a day without the assistant.

What the Google effect showed, and what failed to replicate

Worry that search engines weaken memory traces back largely to a 2011 Science paper by Betsy Sparrow, Jenny Liu and Daniel Wegner. Its four studies suggested that hard questions make people think of computers, and that people who expect information to stay available remember it less well and remember where to find it instead.¹⁰

Part of that has not survived. The computer-priming experiment failed twice in a large replication project reported in 2018, and again in a 2020 PeerJ replicationreplication: Running a study again with fresh data and the same method, to see whether the original result comes back. A finding that fails to replicate is not automatically wrong, but it stops being something you can lean on.Full entry in the glossary designed to stay closer to the original; its author concludes the effect may be much smaller than first thought. All three attempts tested that priming experiment, not the memory ones.¹¹

The memory side has a twist. Across three experiments reported in 2015, people who saved a file before studying the next one remembered the next one better, unless the save seemed unreliable.¹² Offloading, in other words, can clear room as well as empty it.

Myth

Research has shown that search engines wreck your memory.

Fact

The Google-effect paper's computer-priming experiment failed repeated replication attempts, and later lab work found that saving files can free memory for new material.

Knowing where a contract lives rather than what its notice clause says is fine for a file you open twice a year, and a poor bet for the pricing rules you quote to clients every week. Chatbots pose the same know-or-find choice that search did, only many times a day, so make it on purpose rather than by default.

Your Brain on ChatGPT: a small preprint, read with care

The MIT Media Lab’s “Your Brain on ChatGPT” study is a preprintpreprint: A research paper posted publicly before it has been peer reviewed or published in a journal; in economics and finance the same stage is usually called a working paper. Its findings may change, or fail to hold up, by the time it is published.Full entry in the glossary that had not been peer reviewed as of September 2026. It recorded brain activity with EEG in 54 students and university staff in the Boston area, randomly assigned to write essays with ChatGPT, a search engine or nothing, over three sessions spread across four months. Brain connectivity was strongest in the no-tools group and weakest in the ChatGPT group, and in the first session most ChatGPT users could not correctly quote a sentence from the essay they had just written. Only 18 people returned for a fourth session, and the authors list the small, local sample as a limitation.¹³

A comment by other researchers, also a preprint, questions the sample size, the EEG analysis, reproducibility and inconsistencies in how results were reported.¹⁴ EEG connectivity is also a proxy: it shows how brain regions coordinate during a task, not how well someone reasons. A headline claiming that ChatGPT damages the brain reads far more into those patterns than one small, unreviewed study can carry.

The finding that transfers most easily to work is the plainest. If you cannot recall a sentence of a document you “wrote” an hour ago, take it as a sign you have not really processed it yet, and read it again before you have to defend it in a meeting.

Ranking the studies: trials, surveys and a preprint

The best evidence that AI help can cost learning comes from randomized trials in schools and universities; the evidence about workers comes from surveys, and the brain-scan evidence from one small preprint.

Weigh each row by its design: a randomized trial shows what one kind of AI help did in one setting, a survey only what went together, and a preprint may still change after review.

The bottom line

The studies so far do not show that AI makes you a weaker thinker across the board. The risk is specific: it lets you skip the practice and the checking that keep your own understanding in shape, and the strongest trial reviewed here shows the price, better work during practice and weaker performance once the tool is gone. Attempt first, ask for hints when you are learning, and give the most scrutiny to the outputs you trust most.