You ask ChatGPT to draft the email, summarize the report, and pick which option sounds smarter, and you accept the answer without quite checking it. A 2025 study of 666 people found that the more often you do that, the lower your critical thinking scores tend to run, with a habit called cognitive offloading sitting in the middle of the link. The relationship is correlational, not proof that AI rots your brain. It is still real, statistically significant, and worth taking seriously today.
Most coverage of this research split into two useless camps. One side ran panicked headlines about AI making us stupid. The other waved it away as a study that proves nothing. Both miss the finding. The danger is not that you offload work to a machine. You already offload to calculators, maps, and your phone's contacts list, and you are fine. The danger is offloading the one thing you should keep, which is the reasoning.
What the study actually found
Frequent AI tool use correlated with weaker critical thinking, and cognitive offloading explained most of that connection. Michael Gerlich, a researcher at SBS Swiss Business School, surveyed 666 participants in the United Kingdom across three age bands, measured how often they used AI tools, how much they offloaded mental tasks, and how they scored on critical thinking. The pattern was consistent: heavier AI users offloaded more thinking, and heavier offloaders scored lower. Offloading was the mediator, the bridge that connected tool use to the dip in scores.
One detail matters for anyone managing a young team or raising kids around these tools. Younger participants leaned on AI more heavily and posted lower critical thinking scores than older participants. Gerlich could not say whether that gap comes from age, from growing up with the tools, or from something else, but the split was clear in the data.
There was also an education effect worth holding onto, because it points to the way out. Higher educational attainment tracked with stronger critical thinking regardless of how much someone used AI. Read that carefully. Education did not erase the offloading risk. It tracked with a baseline of thinking that held up under it. The plausible reason is not the diploma itself. It is the practiced habit of questioning what you read, and that habit is trainable. That is the optimistic finding buried under a gloomy headline.
Here is what most write-ups of this study skip: the whole sample was British. Every one of the 666 people lived in the United Kingdom, split across ages 17 to 25, 26 to 45, and 46 and older. That single fact should change how hard you lean on the numbers. A pattern measured in one country, at one moment, with people self-reporting their own AI use, is a strong hint about human behavior, not a law of it. The signal is worth acting on. The precision is not worth quoting to two decimal places.
A note on the paper's credibility, because skeptics will ask. The journal published a formal correction on 10 September 2025, and it was narrow: Table 4 had been duplicated from Table 3, and the corrected table was reissued. The author confirmed the scientific conclusions were unaffected. The headline finding stands. As of June 2026, no retraction or failed replication has surfaced.
The study did not show AI makes you dumber. It showed that the more thinking you hand off, the less thinking you appear to do on your own.
Correlation is not the same as cause
This is where honesty matters. The study measured a correlation, so it cannot prove which way the arrow points. Maybe heavy AI use weakens critical thinking. Maybe people who already think less reach for AI more. Maybe a third factor, like time pressure or crushing workload, drives both at once. A single survey taken at one point in time cannot untangle that. Treat the result as a loud signal to change how you work, not as a verdict that AI damages your mind.
A separate study from Microsoft and Carnegie Mellon, presented at CHI 2025, points the same way from a different angle. The researchers, led by Hao-Ping Lee, surveyed 319 knowledge workers about 936 real AI use cases. Their finding: the more people trusted the AI's output, the less critical thinking they reported applying. The more they trusted their own ability instead, the more they thought, though at a higher mental cost. Two studies, two methods, one shared theme. Confidence in the tool tends to quiet your own scrutiny.
Cognitive offloading is not the villain
Offloading is one of the smartest things the human brain does. You stopped memorizing phone numbers because your phone holds them. You do not compute square roots by hand. Writing itself is offloading: putting memory onto paper so working memory has room to reason. None of that made anyone less intelligent. It freed capacity for harder problems. Socrates worried that writing would ruin memory, and in a narrow sense he was right, but the trade bought humanity the ability to build on ideas across generations. Every useful tool moves some load off the mind. The question is always which load.
So the real question is not whether to offload. It is what to offload. There is a healthy division of labor hiding inside the research that almost no coverage names directly: offload recall, keep reasoning. Forgetting a fact, a file name, or last week's decision is pure friction. Letting a tool reach a conclusion you never examined is where the cost actually shows up.
Recall versus reasoning
Recall is retrieval: what did we agree, where is that document, what were the numbers. Hand it off and you lose nothing, because reproducing it from memory adds no insight. Reasoning is judgment: is this argument sound, does this conclusion follow, what are we missing. Hand that off and critical thinking atrophies, because the skill only grows through the act of doing it.
The trap is that AI blurs the line between the two. Ask it to find a fact and it answers, fine. Ask it to recommend a strategy and it answers in the same confident tone, and the two outputs feel identical even though one is recall and the other is reasoning you just skipped. The Microsoft and CMU workers who kept their edge described doing more verification and stewardship, not less work overall. Their critical thinking did not disappear. It moved from generating answers to scrutinizing them. That shift is the whole game.
| Mental task | Safe to offload to AI | Keep doing yourself |
|---|---|---|
| Remembering facts and context | Yes, recall is pure friction | No need |
| Summarizing long text | Yes, then verify the summary | Spot-check against the source |
| Drafting a first version | Yes, as raw material | Edit and decide what stays |
| Judging if a claim is true | No, verify it yourself | Yes, this is the reasoning |
| Choosing between options | No, ask it to lay out tradeoffs | Yes, make the call |
| Deciding what the goal is | No | Yes, always yours |
How to use AI without outsourcing judgment
The practical fix is a short set of habits that keep your reasoning in the loop while the tool still does the heavy lifting. None of them slow you down much. They move your effort from producing the answer to pressure-testing it, which is exactly the shift the Microsoft and CMU researchers saw in the workers who kept thinking. The point is not friction for its own sake. It is to stay the editor of your own judgment instead of the passive recipient of someone else's.
- Form a rough answer before you ask. Even a bad guess primes you to evaluate the AI's response instead of swallowing it whole.
- Ask for the reasoning, not just the conclusion. Make the model show its steps so you can find the weak ones.
- Verify any load-bearing fact against a primary source. Treat AI output as a confident draft, never as a citation.
- Make the final decision yourself. Have the tool lay out options and tradeoffs, then you choose.
- Notice when you stop disagreeing with it. The moment you accept everything is the moment offloading has crept into your reasoning.
A fast self-test: after an AI session, can you explain the answer to someone else and defend it? If not, you offloaded the thinking, not just the typing.
Why dependence builds so quietly
The reason this slips past you is that each individual shortcut is rational. Accepting one summary saves five minutes, and the answer was probably fine. The cost is invisible per use and only compounds across hundreds of uses. It has the shape of any habit. No single instance feels like the problem. By the time skipping the reasoning is automatic, you have stopped noticing you do it, which is precisely why these studies measured the effect across a population rather than in one dramatic moment.
There is also a confidence loop worth naming. When the tool is right over and over, your trust climbs, and the Microsoft and CMU work tied higher trust in the tool to lower critical thinking. The fix is not blanket distrust, which just wastes effort. The fix is calibrated checking: scrutinize hardest where being wrong is expensive, and let low-stakes recall flow through untouched. Spend your scrutiny like a budget, not like a reflex you have switched off.
Where memory fits in
Here is the connection that ties the research to daily tooling. A lot of the friction that pushes people to over-rely on AI is recall friction: re-explaining the same context to ChatGPT every session, re-pasting the same brief, re-establishing what the tool should already know. When that grind disappears, your attention is free to land where it should, on the reasoning rather than on re-remembering.
That is the slice MemX handles. It is a private memory layer that gives ChatGPT, Claude, and Gemini persistent recall across sessions, so you stop rebuilding context and the assistant stops forgetting. The honest framing matters here. This offloads recall, the friction half, and leaves the reasoning with you. It is not a claim that the tool makes you a sharper thinker. The study cannot prove that, and neither can a product. What it can do is keep the easy stuff off your plate so your effort goes to judgment.
On privacy, MemX is private by architecture: per-user isolation, encryption at rest with customer-managed keys, on-device handling where it applies, and your memory is not used to train models. Offloading recall should not mean donating your context to someone else's training set. None of this is legal, medical, or financial advice, and using MemX does not make you or your organization compliant with any regulation.
Frequently asked questions
01Does using AI make you less intelligent?
No study has shown that. The 2025 Gerlich study found a correlation between heavy AI use and lower critical thinking scores, mediated by cognitive offloading, but correlation does not prove AI causes the decline. Treat it as a reason to use AI deliberately, not as evidence of harm.
02What is cognitive offloading?
Cognitive offloading means handing a mental task to an external aid instead of doing it in your head. Writing a note, using a calculator, or asking an AI to summarize a report all count. It is usually helpful. The risk appears when you offload reasoning and judgment, not just memory.
03How many people were in the AI critical thinking study?
Michael Gerlich's study, published in the journal Societies in January 2025, surveyed 666 participants in the United Kingdom across a range of ages. A separate Microsoft and Carnegie Mellon study at CHI 2025 surveyed 319 knowledge workers about 936 real AI use cases.
04Are younger people more affected by AI dependence?
In Gerlich's data, younger participants used AI tools more heavily and scored lower on critical thinking than older participants. The study could not determine why, whether from age, habit, or another factor, so the finding is a flag for attention rather than a confirmed cause.
05How do I use AI without weakening my thinking?
Form your own rough answer before asking, request the reasoning rather than just the conclusion, verify any important fact against a primary source, and make final decisions yourself. Offload recall freely and keep the judgment. If you cannot defend the answer, you offloaded too much.
The takeaway
AI is not quietly erasing your ability to think, but heavy, unexamined use correlates with thinking less, and cognitive offloading is the mechanism behind it. The fix is not to use AI less. It is to be deliberate about the boundary: hand off recall, the part where forgetting is just friction, and keep reasoning, the part where the skill lives. Use the tool to do more thinking, not to skip it.
