The Cognitive Atrophy Paradox: What the Research Actually Says

Someone sent me a long report on "cognitive atrophy" from AI use and asked what I thought. I spent an afternoon checking the citations. Most of the big claims held up. A few did not. The interesting thing is that the story you get after throwing out the weak sources is sharper than the one the report was trying to tell.

Short version. There is real, peer-reviewed evidence that leaning on LLMs changes how your brain engages with work. This is not pop-psych panic. It is also not the calculator discourse on repeat. The specifics are different this time, and they matter.

What the evidence actually shows

Four studies do most of the work. I am going to name them individually because this field is drowning in Medium posts that cite each other in circles.

1. MIT Media Lab, "Your Brain on ChatGPT"

Kosmyna and colleagues ran an EEG study. Participants wrote essays under one of three conditions. LLM assistance, search engine, or brain-only. The LLM group showed the weakest neural coupling in the alpha band, which tracks internally driven processing like semantic retrieval and brainstorming. The brain-only group had the strongest and most widely distributed networks.

The follow-up finding is the one that stuck with me. LLM users struggled to accurately quote their own essays minutes after finishing them. The synthesis happened outside their heads, so it never got encoded. The authors named the phenomenon cognitive debt. Short-term ease, long-term cost.

One caveat. This was 54 participants across sessions one through three, and only 18 in session four. That is a small study. The effect is suggestive, not conclusive. I would not bet a policy on it yet, but the direction is worth taking seriously, and the EEG signal is the kind of thing that is hard to fake.

2. APA, April 2026, overreliance undermines confidence

The American Psychological Association published a study in Technology, Mind, and Behavior on April 16, 2026. 1,923 adults in the US and Canada ran ten simulated work tasks using commercial AI tools. After the tasks:

• 58% agreed that AI "did most of the thinking."
• Those same participants reported lower confidence in their independent reasoning and less perceived ownership of their ideas.
• Participants who actively challenged or modified AI output reported the opposite.

This is a much larger sample than the MIT work, and the effect is behavioral rather than neural. The lead researcher's framing is the one to hold onto. The risk is not that AI makes you dumber. The risk is that you stop doing the deeper cognitive work that produces novel thinking.

3. Gerlich (2025), the age gap is real

Published in Societies in January 2025. 666 UK participants across three age bands. The correlations are unusually clean for this kind of work.

A quick note on what the numbers mean. An r value is a correlation coefficient. It runs from -1 to +1. Zero means no relationship. +1 means the two things move together perfectly. -1 means they move in exact opposite directions. In social science, anything above 0.5 in either direction is considered a strong relationship. Gerlich is reporting values above 0.7, which is strong.

• Cognitive offloading vs. AI usage: r = +0.72. More AI use, more offloading. Tight coupling.
• Cognitive offloading vs. critical thinking: r = −0.75. More offloading, lower critical-thinking scores. Also tight coupling.
• The 17 to 25 cohort showed the highest AI reliance and the lowest critical-thinking scores. The 46 and up cohort showed the inverse.

Higher education softened the effect. That is the interesting knob to pay attention to.

4. MDPI, November 2025, the CAP model

"Cognitive Atrophy Paradox of AI to Human Interaction: From Cognitive Growth and Atrophy to Balance" in Information, Vol 16, Issue 11, Article 1009. It proposes a four-phase trajectory. Cognitive delegation, automation complacency, epistemic erosion, systemic dependency. It also introduces a Cognitive Sustainability Index (CSI) built from five behavioral parameters. Autonomy, reflection, creativity, delegation, reliance.

This is a theoretical model, not an empirical finding. Treat it as vocabulary, not proof. It is useful vocabulary though, and I will probably end up borrowing some of it.

5. DiNapoli op-ed, April 2026, the policy side

The NY State Comptroller's April 2026 op-ed is not about cognition directly. It belongs in this picture anyway. DiNapoli is pressing US companies to disclose the workforce impacts of AI honestly instead of only the productivity numbers. When you step back, the cognitive story and the labor story are the same story. Both ask whether the efficiency gain is being measured on the right ledger.

What I think is going on

The frame that keeps coming up is "this is just the calculator again." It is wrong in one specific way. A calculator automates a sub-step. You still have to know what to calculate and why. LLMs automate the choice of what to do, the structure of the answer, and the verification of the result. The loop closes without you in it.

The MIT paper and the APA paper are measuring two sides of the same thing. MIT shows the brain disengaging during the task. APA shows the user disengaging after it, with less confidence and less ownership. Gerlich points at the cohort where this matters most, which is the people who are still building the circuits in the first place.

That is the part I actually worry about. An adult who uses AI to summarize a paper has the skill already. They are spending it, not failing to build it. A 19-year-old who has never had to sit with an ambiguous prompt for an hour is in a genuinely different situation. The MDPI paper calls this developmental foreclosure and the name fits.

What to do about it, honestly

I do not have a program. I have habits I am trying to keep.

• Architect before prompting. If I cannot sketch the structure of the answer on paper, I am not ready to ask the model. The prompt is a symptom of the thinking, not a substitute for it.
• Treat output as a draft from a junior. Never accept, always edit. The APA result is basically an instruction. The people who challenged the model kept their confidence. The people who rubber-stamped it lost it.
• Keep no-AI zones. Specific projects where I work from scratch, not because it is more efficient, but because the muscle needs the load.
• Write the summary yourself. If I want to remember something, I summarize it without the model. The MIT finding about people not recognizing their own essays is the one that scared me into this habit.

Where the original report got sloppy

Since I went to the trouble of checking, the corrections are worth noting for anyone else writing about this.

• The widely quoted stat of "64% of employees report increased workloads, 5% feel they are maximizing AI" is from EY's 2025 Work Reimagined Survey, not Thomson Reuters. It gets misattributed constantly.
• "Cognitive debt" is the MIT paper's term, not a generic synonym for "unlearned concepts."
• The Gerlich age-gap finding gets cited to blogs more often than to Gerlich. The primary source is the Societies paper.
• The "Google Effect" (Sparrow et al., 2011) keeps getting cited via business blogs instead of the original Science paper.

None of these corrections change the conclusion. They sharpen it.

The real question

Nicholas Carr asked a version of this in 2008 about Google. The answer then turned out to be "yes, sort of, at the margins." The honest read on the current evidence is that LLMs are a bigger version of the same question. They hit a deeper layer of cognition, at a speed that is not giving the workforce or the school system time to adapt.

The useful frame is not will AI make us dumber. It is what do I do with the hour of thinking I just got back. If the answer is another hour of output, the trade was bad. If it is an hour of harder thought the model could not do, the trade was good. That is a choice, not a drift. The drift is what the CAP model is describing.

The uncomfortable thing about all of this is that the effect is quiet. Nobody wakes up one morning having lost the ability to think. What they notice, if they notice at all, is that their first drafts stopped feeling like theirs. That they reach for the tool before the thought has finished forming. That the confidence they used to have in their own reasoning got a little softer. The APA data says this out loud. The MIT data shows the same thing happening at the level of the EEG. The Gerlich data says the people most at risk are the ones who will not get a chance to compare the before and the after, because for them there will not be a before.

So the move is not to reject the tools. It is to decide, deliberately, where the thinking happens. To keep a few places in your week that stay yours. To notice when a prompt is a shortcut around an idea you have not had yet, and to write the idea first. The tools are good. They will get better. The question is whether the people using them stay sharp enough to tell the difference between an answer and a good answer. I think that is worth working at.

References

• Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. media.mit.edu/publications/your-brain-on-chatgpt
• Baldeo, S. et al. (2026). Overreliance on AI Programs May Undermine Confidence at Work. American Psychological Association, Technology, Mind, and Behavior, April 16, 2026. apa.org press release
• Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. mdpi.com/2075-4698/15/1/6
• MDPI (2025). Cognitive Atrophy Paradox of AI to Human Interaction: From Cognitive Growth and Atrophy to Balance. Information, 16(11), 1009. mdpi.com/2078-2489/16/11/1009
• DiNapoli, T. P. (2026). Corporate America Needs to Come Clean on AI's Impact on Jobs. Office of the New York State Comptroller, April 2026. osc.ny.gov
• EY (2025). Work Reimagined Survey 2025. ey.com
• Sparrow, B., Liu, J., Wegner, D. M. (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science, 333(6043), 776–778.
• Carr, N. (2008). Is Google Making Us Stupid?. The Atlantic, July/August 2008.