Cognitive Debt — What Happens to the Brain and Society When We Delegate Thinking to AI
Brain connectivity among ChatGPT users dropped 55%, with 83% unable to cite their own writing. MIT Media Lab research reveals the structure of cognitive debt.
TL;DR
- MIT Media Lab experiments found ChatGPT users' brain network connectivity dropped up to 55%, with 83% unable to accurately cite their own essays.
- Cognitive debt compounds like financial debt — following the 'use it or lose it' principle, a self-reinforcing loop of declining cognitive capacity begins.
- The risk depends not on AI use itself but on the 'attitude' during use, requiring systemic approaches to protect cognition as a socially shared resource.
What's Happening
MIT research shows ChatGPT users experience significant drops in brain connectivity and cognitive activity
Having AI write text is convenient. Fast. It returns reasonably well-structured output. But what is happening to our brains behind this "convenience"?
In 2025, a research team at MIT Media Lab published an experimental result. They conducted a study measuring brain waves (EEG) of 54 participants divided into three groups over four months while writing essays. The groups consisted of those using ChatGPT, those using search engines (like Google), and those using only their own minds without any tools—18 participants each.
The results were clear.
The ChatGPT group showed brain network connectivity up to ~55% lower than the Brain-only group. Moreover, connectivity did not immediately recover when AI was removed in the 4th session.
Brain network connectivity strength—the degree to which different brain regions work cooperatively—was highest in the no-tools group, intermediate in the search engine group, and lowest in the ChatGPT group. Particularly, the activation of alpha and theta waves, which are indicators of cognitive activity, was approximately 47% lower in the ChatGPT group.
Numbers alone might be difficult to grasp. More symbolic were the results of memory tests conducted after essay writing. For the task "Please accurately quote the text you wrote," nearly all participants in the no-tools group could quote accurately. Meanwhile, in the ChatGPT group, approximately 83% could not accurately quote their own text. In other words, despite nominally having "written" the text themselves, most participants had not retained its content in memory.
Furthermore, natural language processing (NLP) analysis showed that essays from the ChatGPT group had significantly higher similarity in vocabulary, structure, and arguments compared to other groups. Eighteen essays written by eighteen people became increasingly similar to each other. This was a phenomenon of lost textual diversity.
In the latter half of the study, participants in the ChatGPT group were moved to AI-free conditions in the fourth session. They were asked to write without AI "training wheels." Brain wave data showed that cognitive network activity did not immediately recover. The brain could not quickly reclaim thinking processes once delegated externally.
The research team named this phenomenon "Cognitive Debt." This concept suggests that while delegating thinking to AI may increase work efficiency in the short term, it may weaken human cognitive activity itself in the long term.
Something should be noted. This paper is a preprint submitted to arXiv and has not yet undergone peer review. The sample size of 54 is small-scale, and results are limited to the specific task of essay writing. Whether four months of observation can be called "long-term effects" is also debatable. However, this research raises issues that cannot be ignored in that it quantitatively demonstrates the relationship between AI use and cognitive activity using objective physiological indicators like brain waves.
Background and Context
Explores the broader implications and research background of AI-assisted thinking on human cognition
The Relationship Between Technology and Cognition—From "Google Effect" to "Cognitive Debt"
The idea that technology affects human cognition is not new.
In 2011, Betsy Sparrow and colleagues at Columbia University reported a phenomenon called the "Google Effect." Information easily accessible through search engines is no longer actively memorized by the human brain. Instead, the brain remembers meta-information about "where to search to find it." A cognitive redistribution occurred, where people remember not the information itself but the access routes to information.
This could be seen as a rational adaptation. If there's no need to memorize everything, brain resources could be directed to other cognitive activities. However, the emergence of generative AI qualitatively changed this structure. While search engines were tools for "finding information," generative AI is a tool that "substitutes thinking." What we delegate expanded from memorizing information to the thinking process itself.
In cognitive science, delegating cognitive processing to external tools or resources is called "Cognitive Offloading." Letting calculators handle mental arithmetic. Letting GPS handle route selection. These are typical examples of cognitive offloading. What generative AI substitutes for are the core activities of thinking: considering structure, building logic, and choosing words.
A 2025 study by Gerlich et al. (666 subjects) quantitatively supported this point. AI usage frequency was the most important predictor of critical thinking scores, with particularly pronounced effects among younger populations. The finding that longer educational years mitigated negative effects suggests that existing knowledge foundations may function as resistance to cognitive offloading.
Fairly Evaluating Counterarguments—The "Cognitive Efficiency" Hypothesis
We should pause here. Should MIT's results be interpreted as "AI destroys the brain"? It's not that simple.
- Low-risk, routine tasks
- High trust in AI
- Time pressure
- High-risk, complex tasks
- High confidence in one's own abilities
- Habit of verifying AI output
A 2025 survey by Microsoft Research of 319 knowledge workers (presented at CHI 2025) revealed a more nuanced picture. While AI use reduced critical thinking in low-risk routine tasks, there were cases where AI actually stimulated critical thinking in high-risk, complex tasks—conditional results.
More interesting is how trust in AI and trust in one's own abilities work in completely opposite directions. People who unconditionally trust AI show decreased critical thinking, while those confident in their professional abilities maintain critical thinking even when using AI.
We should be cautious about reading all "decreased activity" shown in brain wave data as "deterioration." The possibility of "cognitive efficiency"—delegating lower-order cognitive tasks like information retrieval and text formatting to AI while directing brain resources to higher-order strategic thinking—cannot be theoretically ruled out. However, the MIT paper's observed result of "inability to quote one's own text" goes beyond the scope of efficiency. Loss of ownership over written content cannot be explained by reallocation to higher-order thinking.
The Meaning Behind the Name "Cognitive Debt"
The research team's choice of the financial term "debt" for this phenomenon was intentional. Debt has several characteristics.
First, it's hard to see. While borrowing, it's convenient and comfortable. Next, it accrues interest. The longer repayment is postponed, the larger the total amount to be paid. And it compounds. Interest is charged on interest.
Cognitive debt has a similar structure.
"It's faster to let AI do it"
PrincipalReduced brain network connectivity
InterestCannot even cite your own writing
Compound"I can no longer write on my own"
InsolvencyPerpetual vicious cycle
A single instance of AI use doesn't "break" the brain. However, as habits of delegating thinking externally become established, opportunities for independent thinking decrease, and the frequency of brain cognitive network activation declines. According to the basic principle of neuroscience—use it or lose it—cognitive abilities gradually deteriorate. Decreased abilities induce further external delegation. Like financial debt, even though the initial principal is small, there's a structure that compounds over time.
Software engineering has a similar concept called "Technical Debt." Low-quality code written to prioritize development speed later rebounds as maintenance costs. In a 2026 commentary, Margaret Storey of the University of British Columbia pointed out that in the AI era, debt accumulates not "in the code" but "in developers' heads." Understanding of why code was written that way, why systems were designed that way, becomes fragmented. Technical debt lives in code. Cognitive debt lives in people's heads. The tradeoff between speed and understanding occurs with the same structure in individual brains and organizational collective intelligence.
Reading the Structure
Analyzes the framework and methodology used to understand cognitive debt patterns
Viewing Cognition as "Socially Shared Resources"
The discussion so far has stayed within the framework of "how individuals should interact with AI," like many existing articles. However, the scope of cognitive debt extends beyond that.
One person's cognitive abilities are not just that person's problem. The ability to think critically, consider multiple perspectives, and express judgments in one's own words—these are foundational capabilities for a democratic society to function.
Evaluating policy merits. Examining candidates' claims. Understanding regional challenges structurally. Such civic judgment capabilities are built on the accumulation of individual cognitive abilities. If cognitive debt accumulates on a societal scale, what's lost is not individual "smartness" but society's overall judgment capacity.
The "decreased essay diversity" shown in the MIT paper takes on different meaning when reread in this context. Discussions become rich when 18 people bring 18 different perspectives. However, when everyone thinks through AI, outputs become similar and perspective diversity shrinks. Microsoft Research's survey also reported that groups with access to generative AI showed decreased diversity in outcomes for identical tasks.
The shrinking of diverse perspectives extends beyond innovation stagnation. Superficial policy discussions, fragile consensus-building, and a society that uncritically follows "optimal solutions presented by AI"—the social consequences of cognitive debt reach much deeper layers than individual-level productivity issues.
Estimating the Costs of "Convenience"
The cognitive debt concept is useful because it makes visible structures that have left costs invisible.
For organizations considering generative AI implementation—government agencies, NPOs, educational institutions—the benefits of efficiency are obvious. Report creation becomes faster. Initial stages of research can be streamlined. Limited human resources can be directed to more important tasks. These certainly have value.
However, when organizational members have fewer opportunities to research, think, and write themselves behind this efficiency, what is lost? While having AI write a grant application may save resources short-term, if the process of structuring and articulating an organization's activities through the application itself has value, skipping that accumulates as long-term debt.
Estimating what costs arise beyond "using it because it's convenient"—this itself is the first step in cognitive debt-conscious practice.
Beyond Binary Opposition—Toward Active Co-creation
Cognitive debt discussions easily fall into the binary opposition of "should we use AI or not." However, both the MIT paper and Microsoft Research's survey commonly show one important condition: The risk of cognitive debt depends not on AI use itself, but on the "attitude" during use.
A commentary by neuroscientists published in Nature in 2025 proposed a "3R Principles" framework:
- Results (Result Verification): Don't unconditionally accept AI output. Verify results yourself
- Responses (Response Agency): Respond to AI suggestions with your own judgment. Don't become a passive consumer
- Responsibility (Maintaining Responsibility): Humans maintain final judgment and responsibility for results
This principle suggests that the essence of the problem lies not in "tool selection" but in "relationship with tools." In actively engaging—thinking for oneself and using AI as a tool for verification and expansion in that process—cognitive debt is less likely to accumulate, and cognitive abilities may even be strengthened.
However, there's a structural challenge here too. For active co-creation to strengthen cognition, "the ability to think for oneself" is required as a prerequisite. Recovering active attitudes from a state where cognitive debt has already accumulated is not easy. The MIT paper's result that cognitive networks didn't immediately recover after AI removal suggests this difficulty.
In education, having sufficient experience thinking independently before encountering generative AI. In organizations, institutionally securing "thinking time" alongside AI implementation. These are not approaches relying on individual willpower but systematically protecting thinking.
Convenient tools, when misused, change their masters. Cognitive debt is still at a repayable stage. But given debt's characteristics, situations where "before we knew it, it had become unrepayable" can occur similarly in both individual life planning and social institutional design.
The question of how to use AI ultimately comes down to the question of "how humans continue thinking." The very capacity to engage with that question is subject to cognitive debt—this nested structure is where the core of this problem lies.
For organizational decision-making approaches and evidence utilization, see the practical guide Introduction to EBPM (Evidence-Based Policy Making), and for frameworks for measuring and evaluating social change, see Introduction to Social Impact Evaluation.
Related Columns
- Anthropic vs Department of Defense — AI Companies' Ethical Judgment and National Security
- AI Regulation, US Federal vs State Battles
References
Cognitive Debt: Assessing the Impact of AI-Assisted Writing on Learning — MIT Media Lab. arXiv (preprint)
Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips — Betsy Sparrow, Jenny Liu, Daniel M. Wegner. Science, 333(6043), 776-778
Is AI Making Us Less Critical Thinkers? — Teresa Gerlich et al.. Societies, 15(3)
Who Does the Thinking? A Push for Human-Centered AI at Work — Microsoft Research. CHI 2025 Conference Proceedings
