Cognitive debt

Cognitive debt is the idea that relying on artificial intelligence (AI) tools to do mental work saves effort now but builds up costs later, in the form of weaker thinking skills, poorer memory, or a shallower understanding of the material. The term draws on the metaphor of financial debt: the borrowed effort has to be repaid, with interest. It entered wide use through a June 2025 preprint by Nataliya Kosmyna and colleagues at the MIT Media Lab, who described cognitive debt as a condition in which repeated reliance on large language models (LLMs) "replaces the effortful cognitive processes required for independent thinking."^[1:1]

In the study, 54 university students and staff wrote essays with ChatGPT, with a search engine, or with no tools, while their brain activity was recorded by electroencephalography (EEG). The ChatGPT group showed the weakest brain connectivity, had the most trouble quoting essays they had just written, and felt the least ownership of their work.^{[1:2][1:3][1:4]} The study has not been peer-reviewed, and other researchers have questioned its sample size, statistics, and interpretation of the EEG data.^[1:5][2:1]

From early 2026, software engineers adopted the term for a related problem at the level of teams: the gradual loss of a team's shared understanding of a codebase when much of the code is written by AI.^[3:1][4:1] Researchers have also linked cognitive debt to cognitive surrender, the tendency to accept AI output without checking it. Cognitive surrender describes what happens in a single decision; cognitive debt describes what builds up over many.^[4:2]

Origin in the MIT essay-writing study

The preprint, titled Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, was posted to arXiv on June 10, 2025, and revised on December 31, 2025.^[1] Its authors are Nataliya Kosmyna, Eugene Hauptmann, Ye Tong Yuan, Jessica Situ, Xian-Hao Liao, Ashly Vivian Beresnitzky, Iris Braunstein, and Pattie Maes.^[1] A later commentary in Frontiers in Developmental Psychology describes the term as the MIT group's coinage.^[5:1]

Design

The researchers analysed data from 54 participants aged 18 to 39, all recruited from five universities in the Boston area: MIT, Wellesley College, Harvard University, Tufts University, and Northeastern University.^[1:6] Participants were split into three groups of 18 and wrote an essay in each of three sessions. The LLM group could use only ChatGPT. The search engine group could use any website except AI chatbots. The "brain-only" group could use no tools at all. Each essay had a 20-minute limit.^[1:7][1:8] In a fourth session, attended by 18 participants, the groups swapped: former ChatGPT users wrote unaided, and former brain-only writers used ChatGPT.^[1:8] The sessions took place over four months.^[1]

Throughout, participants wore EEG headsets that recorded electrical activity at the scalp. The researchers used these recordings to estimate connectivity, the degree to which activity in different brain regions was coordinated. The essays were analysed with natural language processing and scored by human teachers and an AI judge, and participants were interviewed after each session.^[1:8]

Findings

Sources: Kosmyna et al.^{[1:2][1:3][1:4]}

In the fourth session, participants who moved from ChatGPT to writing unaided showed weaker connectivity in two EEG frequency bands, which the authors interpret as a sign that they were still under-engaged after losing the tool.^[1:4] In the discussion section, they write that many of these participants returned repeatedly to a narrower set of ideas, and they present this as evidence of accumulated cognitive debt. They caution that this finding is preliminary and needs a larger sample.^[1:1]

The authors' definition

Kosmyna and colleagues describe cognitive debt as something that "defers mental effort in the short term but results in long-term costs, such as diminished critical inquiry, increased vulnerability to manipulation, decreased creativity." They argue that when people reproduce AI suggestions without evaluating them, they "forfeit ownership of the ideas" and risk taking on "shallow or biased perspectives."^[1:1] On this basis they recommend an educational approach that "delays AI integration until learners have engaged in sufficient self-driven cognitive effort."^[1:9]

Criticism and limitations

The authors themselves list several limitations: a small group of participants from a few neighbouring institutions, the use of a single chatbot, findings specific to essay writing in an educational setting, and no long-term follow-up.^[1:5] The paper remains marked as a preprint under review.^[1:5]

In a commentary posted on December 29, 2025, psychologists Miloš Stanković, Ella Hirche, Sarah Kollatzsch, and Julia Nadine Doetsch, of the University of Vienna and TU Dresden, raised concerns about the study's sample size, reproducibility, EEG methods, reporting consistency, and transparency.^[2:1] Their main points were:

• Statistical power. They estimated that about 159 participants would be needed to reliably detect the kinds of effects reported, and noted that some analyses rested on as few as two to four essays per group.^[2:2]
• Reading the EEG data. The study counted "significant connections" that were stronger in one group than another. The commenters argue that such counts do not show overall brain activity was weaker, and that lower connectivity could reflect shifts of attention, different strategies, or lower motivation rather than less thinking.^[2:3]
• The search engine result. The search engine group used an external tool yet quoted its essays as well as the brain-only group. The commenters argue that this sits awkwardly with the idea that handing off more of a task builds more cognitive debt, and they question whether recalling a single sentence is a good measure of learning.^[2:4]
• Task conditions. Participants reported feeling pressed for time, which may have pushed them to lean on their tools. Reusing prompts in the fourth session may also have introduced practice effects.^[2:5]

Other responses have been more accepting. In an August 2025 commentary in the British Journal of General Practice, Richard Armitage called the study "the first neurophysiological evidence that LLM assistance fundamentally alters how our brains process and retain information," and suggested that cognitive debt may also affect medical training.^[6:1] Jorge Pereira Campos and Tatiana Koff, writing in Frontiers in Developmental Psychology in July 2026, cautioned that "cognitive debt", "reduced neural engagement", and "reduced neural connectivity" are often used as though they mean the same thing, although the first is a theoretical idea and the other two are distinct measurements.^[5:1]

Whether cognitive debt is a durable effect, and not a short-term response to a timed task, remains an open question. The study's authors, their critics, and Pereira Campos and Koff all call for larger, longer-term, and more naturalistic studies.^{[1:5][2:5][5:2]}

Extension to adolescents

All participants in the MIT study were adults. Pereira Campos and Koff argue that the question matters more for adolescents. Between ages 10 and 20, the executive functions that support planning, weighing evidence, and building arguments are still maturing, and these are the same tasks students increasingly hand to AI. They propose that cognitive debt that may be temporary in adults "could be developmental" in children and adolescents, while stressing that "none of this has been demonstrated."^[5:2]

Use in software engineering

In February 2026, software engineering researcher Margaret-Anne Storey of the University of Victoria applied the term to AI-assisted programming. Linking to the MIT study, she described cognitive debt as "a term gaining traction recently" and contrasted it with technical debt: where technical debt lives in the code, cognitive debt "lives in the brains of the developers." Even when AI writes code that is easy to read, she wrote, the people responsible for it "may have simply lost the plot."^[3:1] Programmer and writer Simon Willison described the same experience in his own work. After prompting whole features into existence without reviewing the code, he wrote, "I no longer have a firm mental model of what they can do and how they work."^[7:1]

Storey developed the idea in a March 2026 paper proposing a "triple debt model" of software health.^[4:3] She opens with a student team in her entrepreneurship course whose project stalled by week eight. The team first blamed messy code, but the real problem was that no one could explain why design decisions had been made or how the parts of the system fit together.^[4:4]

Source: Storey.^[4:3]

Storey's use of the term differs from the MIT group's. She treats cognitive debt as a property of a team and a project, not of an individual brain, and says explicitly that her focus is shared understanding eroding over time, not reduced neural engagement.^[4:1] She argues that the problem is not new, since programmers have always worked with partial knowledge of large systems. What AI changes, in her view, is how fast the gap grows and how hard it is to notice. Writing code by hand forces a developer to build at least a rough mental model, while accepting AI-generated code does not.^[4:5]

Storey identifies cognitive surrender as the psychological mechanism behind cognitive debt. Drawing on Steven Shaw and Gideon Nave's finding that surrender raises people's confidence even when the AI is wrong, she argues that teams can feel they understand a system better than they do, so the debt stays hidden until it causes failures.^[4:2] The warning signs she lists include reluctance to change the code, changes that produce unexpected results, slow onboarding of new members, and a low "bus factor", meaning only one or two people truly understand the system. Her proposed remedies include human code review, pair programming, walkthroughs in which developers explain code they did not write, retrospectives, and having an AI agent reimplement a feature so the team can rebuild its understanding.^[4:6]

Related concepts

In an August 2026 position paper on human oversight of AI agents, Margaret Mitchell, Avijit Ghosh, and Samir Passi grouped cognitive debt with "cognitive dependence" and "cognitive surrender" as recent terms for the erosion of the human abilities that oversight depends on.^[9:1]

Analysis: cognitive debt and human agency

This section offers an evaluation grounded in the evidence above.

Cognitive debt bears on human agency, meaning people's lasting capacity to understand their situation, form aims, and act on them. What the debt metaphor adds is attention to time. A single choice to let an AI draft an essay or a function may be reasonable. The concern is that repeated choices of this kind may wear down the very abilities a person would need to judge the AI's work, or to do without it. In the MIT study, ChatGPT users could seldom recall what "they" had written and felt little ownership of it.^[1:3][1:4] In Storey's account, teams lose the ability to change their own software safely.^[4:6] In both cases, the people nominally in charge are less able to steer.

The evidence for the individual, neurological version of the idea is still thin. It rests mainly on one small, unreviewed study whose central measures are disputed.^[2:1] The search engine result also suggests that using external tools is not in itself the problem: what matters may be whether the tool does the thinking or supports it.^[2:4] This matches the line Shaw and Nave draw between offloading and surrender.^[8:1] If so, the most useful responses are those that keep people's own reasoning engaged, not blanket avoidance of AI. Examples include the timing of AI use in education that the MIT authors propose, and the review and explanation practices Storey recommends for software teams.^[1:9][4:6] Larger studies that follow people over months or years, in real settings, could show whether cognitive debt builds up as the metaphor implies, and whether it can be repaid.