Cognitive debt, capability debt, and the rest

Cognitive debt, capability debt, epistemic debt, culture debt, distributed de-skilling. Five terms, published between June 2025 and June 2026, for versions of the same worry: that using AI well in the short term costs something that only shows up later. This page maps them. The finding that came out of doing so was not the one expected. Almost nobody claims to have coined any of them.

Every source in this map was read at the primary source#

The terms below were checked against the paper or the publisher's own page, not against summaries. Where a source could not be reached, it is named at the foot of this page and left out of the map rather than described from secondary coverage. Three sources fell into that category, and one of them would have supported a point this page would otherwise have liked to make.

The debt family, and who actually claims what#

A coinage claim is a specific thing. It reads "we introduce", "we term", "we propose the term". Using a phrase, even defining it carefully, is weaker than claiming it. That distinction turns out to matter more than the terms themselves.

• Cognitive debt. Kosmyna and colleagues, MIT Media Lab, arXiv 2506.08872, June 2025 with a revision on 31 December 2025. The phrase appears four times in 216 pages. It is not in the abstract. There is no "we introduce" or "we coin" anywhere in the paper, and no citation attached to any of the four appearances. The paper never connects it to technical debt, the analogy most coverage assumes it is making. Defined once, on page 151, as "a condition in which repeated reliance on external systems like LLMs replaces the effortful cognitive processes required for independent thinking."
• Capability debt. Rohde, "Short-Term Gain, Long-Term Fragility", SSRN, written 20 April 2026, revised 27 April 2026. Defined in a paragraph that also defines institutional debt, both introduced by way of the technical debt analogy. Rohde makes no coinage claim. What he claims is a mechanism: his stated contribution is "to identify and formalize a mechanism of capability masking and capability erosion". The paper is, in his own words, "a conceptual synthesis rather than a new empirical study".
• Epistemic debt. Sankaranarayanan, arXiv 2602.20206, February 2026, revised March 2026. The author puts the term in quotation marks in his own title, which signals borrowed usage rather than ownership, and explicitly builds on Kirschner's distinction between cognitive offloading and outsourcing. What he does claim as novel is an instrument, the "Explanation Gate".
• Culture debt. Deloitte, press release of 4 March 2026, defined inline as "the negative consequences an organization accumulates by neglecting its culture". Introduced in quotation marks, with no authorship claim.
• Distributed de-skilling. BCG, 17 June 2026. The only firm ownership language in the whole set: "We call this distributed de-skilling, a collective erosion of human skills that undermines organizational intelligence and resilience over time."
• Intent debt. Storey, University of Victoria, in ACM Queue, with the preprint at arXiv 2603.22106, March 2026. Defined as "the absence or erosion of explicit rationale, goals, and constraints that guide how a system evolves". Storey does make a claim, and it is worth reading exactly: "In this article, I propose a triple debt model for reasoning about software health." The claim attaches to the model, not to any of the three terms inside it. She is also the only author in this whole set who builds the bridge to technical debt explicitly, citing Cunningham 1993, which is the connection most coverage of the MIT paper assumes that paper was making.
• Skill atrophy. Jarrahi, Professor at the University of North Carolina at Chapel Hill, writing at CognitiveWorld on 19 March 2026. Defined as "the hidden, accumulating loss of human skill, judgment, and capacity that happens when organizations and workers leverage automation in ways that reduce practice, learning, and ownership", and reasoned through the borrowing metaphor without ever naming it as such: "The organization 'borrows' capability now for speed or convenience, but it must 'pay it back' later." His phrasing on the term is "this is a form of skill atrophy", which is description rather than ownership. What he does claim is a mechanism: "what I call useful friction".

Two more entries, two more instances of the same pattern. Every firm claim in this corpus attaches to an instrument, a model or a mechanism. Not one attaches to the vocabulary.

The actual finding: convergent metaphor, no shared lineage#

Rohde, Sankaranarayanan and Deloitte each reach for a debt metaphor within four months of each other. None of them cites either of the others. Three separate authors, working in organisational economics, computing education and human capital consulting, independently arrived at the same accounting figure for the same phenomenon.

That convergence is more interesting than any individual coinage would have been. It suggests the metaphor is doing something the field genuinely needs, which is to describe a cost that is incurred now and paid later, invisible on any current measure. It also means the question "who said it first" is close to unanswerable and probably not worth answering.

What happened next: the term acquired a second meaning#

The map above was drawn in August 2026. Rechecking it, the pattern has changed in a way worth recording, because it is the opposite of what a contested category usually does.

The 2025 cohort converged on a metaphor without citing each other. The 2026 cohort does cite, and it converges on one source, the MIT paper. But it converges on the source while diverging on the meaning. Storey is explicit about the split in her own text: the term "has also been used to describe measurable reductions in individual neural engagement during AI-assisted tasks", whereas "our use of the term, however, focuses on the team-level and longitudinal dimension". She defines cognitive debt as a property of a team, an "erosion of shared understanding across a software system over time", which is a different object from anything measurable on an individual EEG.

So the most-cited term in the family now has two referents that do not reduce to each other. One is a state inside a single head. The other is a gap between several heads. They share a citation and not a definition.

There is a name for this, and the neatest part is where it comes from. Thoughtworks published volume 34 of its Technology Radar on 15 April 2026 under a headline about combating cognitive debt, and in the same document named the mechanism: "The industry is coining terms for emerging practices before their meanings have stabilized, leading to semantic diffusion." The Radar was one of the largest distribution events the term has had, and it arrived carrying a warning about exactly what distribution at that speed does to a word.

Six weeks later, on 28 May 2026, a Thoughtworks blog post on cognitive debt as an organisational risk described the MIT work in a single sentence: "The researchers called this phenomenon cognitive debt." Checked against the paper, the researchers did no such thing. They used the phrase four times in 216 pages, never in the abstract, never with a citation, and never with any language of introduction. This is not a serious error, and the post is a thoughtful piece that reports the underlying caution honestly. It is worth recording only because it is the mechanism working in real time: a firm names semantic diffusion in April and performs a small instance of it in May.

What the most-cited term actually rests on#

Cognitive debt is the term that reached the public. The study underneath it deserves reading rather than citing.

Fifty-four participants, aged 18 to 39, recruited from five universities in the Boston area, writing essays across three sessions. Three groups of eighteen: one using ChatGPT, one using a search engine, one unaided. EEG on 32 channels. The finding most quoted, that removing AI left the LLM group unable to quote their own work, comes from session four, which only 18 participants attended, nine per arm. The widely repeated 78 per cent and 11 per cent are seven of nine and one of nine.

The paper still carries "Preprint, under review" in the footer of all 216 pages of its December 2025 revision. It has not been peer reviewed. The authors state their own limits clearly, including that findings "are context-dependent and are focused on writing an essay in an educational setting and may not generalize across tasks", and they hedge the central passage themselves: "This next finding should be considered preliminary, as a larger participant sample is needed to confirm the claim."

On 29 December 2025 a formal Comment was posted by Stanković and colleagues at Vienna and TU Dresden, arXiv 2601.00856. Their power analysis puts the required sample at roughly 159. Their sharpest point is about the term itself: the search engine group used an external tool and showed no impairment, with the comparison against the unaided group returning p = 1. If offloading to a tool produced cognitive debt, that group should have shown it. The Comment is also a preprint, offered as a critique rather than a refutation.

None of this makes the study worthless. It makes it a small, unreviewed, contested pilot carrying a term that a great deal of subsequent commentary treats as settled.

What the most influential survey rests on#

BCG's June 2026 article will be quoted in boardrooms for the rest of the year, and its headline numbers are worth reading with their provenance attached. Seventy C-suite leaders and senior executives. Half already observing de-skilling. More than 60 per cent expecting it to be a material threat within three to five years. Judgement and decision-making named as carrying the highest de-skilling risk.

The article states no countries, no fieldwork dates, no sampling method, no response rate, no industry breakdown and no question wording. It has no limitations section, no endnotes and no reference list. A "de-skilling risk score" is reported as though it were a defined metric and is never defined. By the grading used across this estate, that places it as an institutional survey of unknown representativeness, useful as a signal of what senior people now believe and not usable as a measurement of what is happening.

The same applies to Deloitte's March 2026 release, which reports that 85 per cent of leaders call adaptability critical while 7 per cent say they are leading on it, and gives no sample size, no countries and no fieldwork dates in the release itself.

The pattern across all three is the same. The vocabulary is ahead of the evidence, and the evidence is ahead of its own methodology sections.

Where this research sits, stated against itself#

This estate uses capability debt. It makes no claim of first use, and the check above is the reason it never will: Rohde defines the term in print and does not claim it either, so the honest description is that two people arrived at an obvious metaphor for the same problem at roughly the same time, along with several others who reached for adjacent versions.

The same standard applies to the rest of the vocabulary here. Usage theatre and the verifier's discount carry no first-use claim. Synthetic seniority and the missing rungs do, and are dated. Applying to your own vocabulary the test you apply to everyone else's is the cheapest credibility available, and most of the sources on this page do not do it.

One further observation was offered here as an impression rather than a measurement: that the academic terms describe what happens inside an individual head, cognitive debt and epistemic debt and the verification bottleneck, while the consulting terms describe what happens to an organisation, distributed de-skilling and culture debt.

That impression did not survive the next two sources, and it is left standing here with the correction attached rather than quietly deleted. Storey is an academic and her cognitive debt is explicitly a team-level property. Jarrahi is a professor and his skill atrophy is explicitly organisational. The split was never between academics and consultants. It was between people writing in June 2025, when the individual measurement was the only evidence anyone had, and people writing nine months later with organisations in front of them.

The gap the original observation pointed at is real even though the reason given for it was wrong. It is where the mechanism lives, at human capability in the age of AI, and it stays thinly occupied because working there needs both literatures at once.

Adjacent terms worth knowing#

• Verification bottleneck and verification paradox. Huemmer and colleagues, arXiv 2601.17055, January 2026. A three-wave longitudinal pilot in an academic setting reporting that participants leaned hardest on AI for difficult tasks, 73.9 per cent, while accuracy on complex tasks fell to 47.8 per cent and belief-performance gaps widened to 34.6 percentage points. The authors put their own limitations in the abstract, including convenience sampling and no control condition. Neither phrase is claimed as a coinage.
• Fragile experts. Sankaranarayanan again, describing "developers whose high functional utility masks critically low corrective competence". His experiment, 78 participants across three conditions, found unrestricted AI users failed a subsequent AI-blackout maintenance task at 77 per cent against 39 per cent for a scaffolded group.
• Skill Automation Feasibility Index and AI Impact Matrix. Jadhav and Danve, arXiv 2604.06906, April 2026. These are claimed explicitly, "we present" and "we propose". Note what that means: in this whole corpus, the firm claims attach to instruments and frameworks, never to the debt vocabulary.
• The human advantage and changefulness. Deloitte, March 2026. Positioning language rather than defined concepts.
• Knowledge collapse. Acemoglu, Kong and Ozdaglar, NBER working paper 34910, issued February 2026. The most formally modelled thing in this entire corpus and the only one carrying a DOI from an established institution rather than a preprint server or a press release. Their result is conditional and they state the condition: "when human effort is sufficiently elastic and agentic recommendations exceed an accuracy threshold, the economy can tip into a knowledge-collapse steady state in which general knowledge vanishes ultimately, despite high-quality personalized advice". Note the shape of that. It is a model producing a possible steady state under stated assumptions, not a measurement of anything currently happening, and the authors are careful about the difference in a way that most of the vocabulary above is not.
• Cognitive surrender. Shaw and Nave, SSRN 6097646, 2026, described by Storey as "adopting AI outputs with minimal scrutiny, bypassing both intuition and deliberate reasoning". Storey draws the distinction that matters: this is not cognitive offloading, which is the deliberate delegation of a discrete task to a tool and a rational choice. Cited here at one remove, through Storey, and not read at the primary source.
• Comprehension debt and context debt. The first attributed by Storey to Alakmeh and colleagues, 2026, as the gap between what developers can produce with AI and what they actually understand. The second she attributes to practitioners rather than to any paper, meaning the information an AI agent needs and does not have. Both cited here at one remove.

What could not be verified, and is therefore absent#

Three sources were pursued and left out.

• "The AI Deskilling Paradox", Communications of the ACM. The page returns an empty body to every fetch. The title is confirmed; nothing else is. Secondary coverage attributes a byline and a date, and neither is used here.
• "AI Debris", arXiv 2606.12432. PDF-only with no extractable text. Secondary summaries suggest it contains a term close to "capability debris". That phrase could not be confirmed to exist in the paper, so it is not on the map.
• "The apprenticeship void", VILAKSHAN XIMB Journal of Management. The DOI resolves and the article exists. The publisher's bot wall prevented reading it, so its authors, method and any coinage claim are unknown.

Naming these is the point rather than an apology. A map of who said what, built partly from search summaries, would be a map of what search engines believe.

The citation that would have moved the date#

One check on this page mattered more than the others, because it would have rewritten the timeline for the whole family.

Storey's reference list dates the MIT work to 2024 and places it at a workshop: "Kosmyna, N., Beh, J., Kellogg, R., Sra, M., and Maes, P. 2024. Cognitive Debt in the Era of Generative AI: Evidence from Writing Assistance Using Large Language Models. CHI '24 Workshop on Human-Centred Evaluation of LLMs." If that paper exists, cognitive debt is a year older than this page says, the phrase sat in a title rather than four times in a body, and the ordering of the entire debt family changes.

It was checked against the MIT Media Lab's own publications list for Nataliya Kos'myna, which is the record the authors maintain themselves. There is no CHI '24 workshop paper of that name on it. The only entry carrying the phrase is the one already on this page: "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task", arXiv 2506.08872, 2025, and its co-authors are Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes. Beh, Kellogg and Sra are not among them. The year, the venue and three of the five named authors do not match anything in the record.

This is one slip in a long reference list, in a paper that is otherwise the most carefully argued thing in this corpus, and it is recorded here for one reason only. It is a citation for the origin of the field's most-quoted term, sitting in the most authoritative venue any of this vocabulary has reached. Citations of that kind get copied. If it propagates, a term whose actual first appearance is a June 2025 preprint will acquire a 2024 conference provenance it never had, and it will be almost impossible to unpick afterwards.

Which is the argument for this page in a single example. Vocabulary moves faster than the checking, and the checking is not difficult. It took one look at a list the authors publish themselves.

Key research and primary sources

• Kosmyna, N., Hauptmann, E., Yuan, Y.T., Situ, J., Liao, X.-H., Beresnitzky, A.V., Braunstein, I. and Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv 2506.08872, v2 31 December 2025. Preprint, under review.
• Stanković, M., Hirche, E., Kollatzsch, S. and Doetsch, J.N. (2025). Comment on: Your Brain on ChatGPT. arXiv 2601.00856, 29 December 2025. Preprint.
• Rohde, W. (2026). Short-Term Gain, Long-Term Fragility: AI Labor Substitution and the Erosion of Sustainable Capability. SSRN, written 20 April 2026, revised 27 April 2026. Preprint.
• Goel, S., Martin, D. and Kaffe, C. (2026). When Everyone Uses AI, Companies Risk Losing Critical Skills. Boston Consulting Group, 17 June 2026.
• Deloitte (2026). In a New Era of Work, Winning Organizations Will Build the Human Advantage. 4 March 2026.
• Sankaranarayanan, S. (2026). Mitigating "Epistemic Debt" in Generative AI-Scaffolded Novice Programming using Metacognitive Scripts. arXiv 2602.20206. Preprint.
• Huemmer, M., Durner, F., Shyiramunda, T. and Cummings-Koether, M.J. (2026). AI, Metacognition, and the Verification Bottleneck. arXiv 2601.17055. Preprint.
• Jadhav, R. and Danve, J. (2026). The AI Skills Shift. arXiv 2604.06906. Preprint.
• Storey, M.-A. (2026). From Technical Debt to Cognitive and Intent Debt: Rethinking Software Health in the Age of AI. ACM Queue, with preprint at arXiv 2603.22106, March 2026.
• Jarrahi, M.H. (2026). Skill Atrophy: Frictionless AI and Cognitive Debt. CognitiveWorld, 19 March 2026.
• Acemoglu, D., Kong, D. and Ozdaglar, A. (2026). AI, Human Cognition and Knowledge Collapse. NBER Working Paper 34910, February 2026. DOI 10.3386/w34910.
• Thoughtworks (2026). Technology Radar volume 34, 15 April 2026, and Kamelman, M. Cognitive debt is a real organizational risk, 28 May 2026.
• Publications list for Nataliya Kos'myna, MIT Media Lab. Used to check a disputed citation, as described above.

Related SuperSkills research#

For the category these terms are circling, human capability in the age of AI. For the mechanism, capability debt, the missing rungs and synthetic seniority. For how sources are graded here, the evidence base and what we actually know. For the oversight argument, human in the loop is not a safeguard.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Every source on this page was read at the primary source on 28 August 2026, with the three exceptions named above. Coinage claims were tested by searching each paper for explicit claiming language rather than by inference from usage. This page describes other people's work and has an obvious interest in one of the terms on it, so the standard applied to everyone else has been applied here first.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.