A polished AI-generated answer can feel like proof that learning happened. Yet a 2025 MIT study raises a harder question: if a tool supplies the words, reasoning, and structure, how much of that work stays with you?
The concern isn't that using ChatGPT makes someone less intelligent overnight. It is that repeated AI dependence may reduce the practice that builds memory, writing ability, and independent judgment.
The MIT Essay Experiment Put AI Use Under an EEG Cap
MIT researchers studied what happens in the brain while people write essays with different levels of technological help. Their preliminary research, Your Brain on ChatGPT, followed 54 participants who wore 32-electrode EEG caps while completing writing tasks.
Three ways to write an essay
Participants worked under three conditions:
- One group wrote using only their own knowledge and ideas.
- A second group could use a search engine.
- A third group used ChatGPT.
EEG does not read thoughts or measure intelligence. Instead, it tracks electrical activity at the scalp, which researchers can use to study patterns associated with attention, cognitive effort, and coordination across brain regions.
The AI group showed weaker connectivity
The brain-only group showed dense, distributed brain connectivity during writing. The search group showed less connectivity. Participants who used ChatGPT showed the weakest connectivity of the three groups.
That result does not prove that ChatGPT causes permanent harm or that every AI-assisted assignment produces the same effect. However, it does point to lower cognitive engagement during this particular writing task.
The more of the writing process the tool handled, the less connected the participants' brain activity appeared to be.
Writing is more than placing sentences on a page. It requires recalling information, choosing evidence, testing ideas, arranging an argument, and deciding what words mean. When a chatbot produces much of that work, the finished essay may still look strong, but the user's mental role can become much smaller.
What Happened When ChatGPT Was Taken Away
The MIT study did not stop after measuring assisted writing. Researchers also asked participants to write without AI after they had used ChatGPT during earlier sessions.
The users struggled to recall their own work
According to the study, the ChatGPT group had trouble quoting or accurately recalling essays they had previously submitted. Their EEG patterns also remained less engaged when the AI tool was no longer available.
That finding matters because it separates output from understanding. A person can turn in a coherent essay without being able to reconstruct its argument, explain its evidence, or remember its wording later.
A completed answer does not always mean the brain completed the learning.
Cognitive debt is a useful warning label
The researchers used the term cognitive debt for the possible cost of repeatedly outsourcing mentally demanding work. Like financial debt, the benefit appears first. You save effort now, while the cost may show up later when you need to perform without the tool.
For a student, that moment might be an exam, a class discussion, or a blank page with no chatbot available. For a professional, it could be a client meeting, an emergency decision, or a task where the AI output needs to be checked from first principles.
Assistance and replacement are different. AI can help someone spot gaps in an argument or suggest revisions. It becomes more concerning when it replaces the act of forming the argument in the first place.
The Google Effect Shows How Memory Adapts to Tools
Long before ChatGPT, researchers had already found that the mind changes its strategy when information is available elsewhere. In 2011, Betsy Sparrow, Jenny Liu, and Daniel Wegner published the Science study on the Google effect.
People may remember where information lives
Across several experiments, the researchers found that when people expected information to be saved on a computer, they were less likely to remember the information itself. Instead, they were more likely to remember where they could retrieve it.
That does not mean external memory is bad. Calendars, books, maps, libraries, and search engines all extend what people can do. Still, the study shows that memory responds to the environment. When the mind expects easy retrieval, it may spend less effort storing the fact.
Chatbots can take over more than retrieval
A search engine typically gives you a list of pages. You still have to choose a source, read it, assess its credibility, compare claims, and put the answer into your own words.
A chatbot can skip many of those steps. It can summarize material, frame an argument, write transitions, and produce an entire draft in seconds. That convenience can be useful, but it also reduces the pauses where people normally sort information into memory.
With a search engine, you may remember:
- The question you asked and the site where you found an answer.
- A few facts you selected after reading competing sources.
With a chatbot, you may remember:
- That an answer appeared quickly and sounded plausible.
- Less of the evidence, reasoning, and wording behind the response.
The second outcome is a risk, not a rule for every user. It depends on whether the person reads actively, checks sources, revises the material, and can explain it afterward without looking.
GPS Offers a Familiar Example of Mental Offloading
Navigation offers another way to understand the trade. GPS is helpful when you are lost, rushed, or driving somewhere unfamiliar. Yet when an app gives every turn, it also removes much of the route-planning work your brain could otherwise practice.
Practice changes navigation skills
Finding your own way requires attention to landmarks, distance, direction, and spatial relationships. A GPS device handles those decisions on demand. That reduces effort in the moment, which is often exactly why people use it.
However, less effort also means less practice. If you always follow turn-by-turn instructions, you may reach the destination without forming a usable mental map of the route.
London taxi drivers show that practice leaves a mark
A well-known study of London taxi drivers found differences in the hippocampus, a brain region involved in memory and navigation. The PNAS research on taxi drivers' hippocampi reported greater volume in the posterior hippocampus among taxi drivers than in control participants.
London cab drivers train extensively to learn the city's streets and routes. The study does not show that using GPS occasionally shrinks anyone's brain. It does show that sustained, demanding navigation practice is associated with measurable structural differences.
The lesson applies to writing and recall. Skills improve through use. If a tool repeatedly takes over planning, explanation, or memory retrieval, those mental operations get fewer repetitions.
Convenience can preserve time while reducing the practice that makes a skill available when convenience disappears.
A Medical Study Raised Similar Concerns About AI Reliance
The concern becomes more serious when it appears in expert work. A 2025 study by K. Budzyn and colleagues examined whether exposure to AI-assisted colonoscopy affected clinicians' independent performance after the AI system was removed.
Detection rates fell after AI exposure
The study tracked experienced endoscopists using AI support to detect adenomas, which are precancerous growths. After months of access to the tool, researchers compared standard, non-AI-assisted colonoscopies before and after that exposure.
The adenoma detection rate dropped from 28.4% before AI exposure to 22.4% afterward, a decline of six percentage points. The PubMed record for the endoscopist deskilling study provides the reported figures and study details.
The result needs careful interpretation
These were not novices learning their first procedure. They were experienced clinicians. That makes the finding uncomfortable because it suggests that a helpful system may improve results while it is available, yet leave users less effective when they must work alone.
The tool may improve short-term performance while weakening independent performance underneath.
Still, this was an observational study. It cannot prove that AI use caused the drop, and its findings concern one clinical task, one setting, and a particular kind of AI support. It would be wrong to treat it as proof that every doctor, worker, or student becomes worse after using AI.
The study does justify a more careful question: how should professionals train with decision-support tools while preserving the ability to detect errors and perform without them?
Convenience and Capability Are Not the Same Thing
AI systems can save time, reduce frustration, and help people work through difficult tasks. In medicine, they may support detection. In writing, they can generate an outline, clarify awkward sentences, or point out missing counterarguments. Those are real benefits.
Yet an answer that arrives with little effort can hide an important difference between getting a result and building a capability.
| Short-term gain | Possible long-term cost |
|---|---|
| A draft appears faster. | You may remember less of the subject. |
| The language may sound polished. | You may get less practice explaining ideas. |
| A difficult task feels easier. | You may struggle when the tool is unavailable. |
| AI can flag details you might miss. | You may become less attentive to the details yourself. |
The table does not mean that every use produces every cost. It shows why the quality of a final answer is an incomplete measure of learning. A strong paragraph on a screen may say little about whether its user can defend the point, revise it, or write a new version tomorrow.
Each time a machine supplies the answer, it may also remove some of the practice needed to find the answer independently.
Use AI Without Handing It the Whole Job
The best safeguard is to keep the core thinking step in your own hands. Start with what you know before opening a chatbot. Write a rough thesis, identify the question you cannot answer, or explain the issue in plain language from memory.
Let AI challenge your work
AI is often more useful after you have created a first attempt. At that point, it can help you find weak spots rather than becoming the source of every thought.
You might ask it to:
- Identify a missing objection in an argument you already wrote.
- Explain why a paragraph is unclear without rewriting it for you.
- Suggest questions that would improve your research.
- Compare two approaches after you have described both in your own words.
This approach leaves you responsible for the claim, the evidence, and the final judgment. It also makes it easier to catch errors, because you have a mental model of the subject before the answer appears.
Test whether the skill is still yours
Occasional unaided practice is a simple reality check. Write a short summary of an AI-assisted article without looking at it. Explain a topic aloud before you ask for help. Find your way around a familiar neighborhood without GPS.
These small tests reveal whether a tool is supporting your ability or substituting for it. If you cannot reconstruct the reasoning without the screen, the result may belong more to the system than to you.
What the Research Still Cannot Answer
The available evidence raises valid concerns, but it does not settle the full question of AI and the brain. The MIT EEG work is a preprint, which means it has not yet completed peer review. Its essay-writing results also cannot predict every use of every chatbot.
The Google effect shows that memory adapts when information is stored elsewhere. The taxi-driver research links intense navigation practice with brain differences. The colonoscopy study found a decline in unassisted detection after AI exposure, but it cannot establish cause and effect on its own.
Important: The concern is not that every machine makes people less capable. The concern is that repeated dependence can reduce practice in the skills a machine replaces.
AI may be a useful aid for research, editing, accessibility, and technical work. The unanswered question is how much independent effort people need to retain before the convenience begins to cost them more than it saves.
The Skill Behind the Screen Still Matters
The MIT experiment offers a useful reminder: a task can be completed without being fully learned. Search engines, GPS, and AI assistants all make difficult work easier, but they can also change what the mind chooses to remember and practice.
The point is not to reject AI. It is to use it with enough intention that your writing, recall, and judgment remain your own abilities, not features that only appear when a tool is open.
0 facebook:
Post a Comment