Cognitive Debt

Cognitive debt is the term MIT's Media Lab attached to the pattern they measured in the "Your Brain on ChatGPT" study: people who outsource the framing and drafting of a task to an LLM save mental effort in the moment and lose downstream capacity to think critically about the result. The name is deliberate — like technical debt, it accrues silently, can be productive for short bursts, and compounds destructively if never paid down.

The MIT measurements

The Media Lab compared three groups writing essays: LLM-assisted, search-engine-assisted, and unaided ("brain only"). EEG measurements showed brain-region connectivity scaling down with every layer of external support, with the LLM group showing the weakest coupling. The most concrete finding: after writing the essay, 83% of LLM users couldn't quote a single line of what they had just produced.

The structural reading: writing is an externalisation of an internal model. If the externalisation happens outside the writer's head, the internal model never updates. The output looks the same, but the writer's brain has not been through the work.

Convergent evidence

Two other 2026 studies — covered together in Osmani's piece — land in the same place:

Anthropic's randomised library-learning trial: AI-assisted engineers and manual engineers finished tasks at the same speed, but the AI group scored 50% on a follow-up comprehension quiz vs 67% for the manual group. The intra-group split was the more interesting finding: engineers who asked the model conceptual questions scored 65%+, copy-pasters scored under 40%. The tool didn't determine the outcome — posture did.
CHI 2026: when an LLM was available at the start of a task, it framed the whole problem, and downstream human decisions were measurably worse even when the human did most of the work alone.

Different instruments, same finding.

What makes the debt persistent

A few features distinguish cognitive debt from ordinary forgetting:

It's invisible at the work product. The artifact looks the same, sometimes better. There's no immediate signal that the writer doesn't understand what they wrote.
It's path-dependent. CHI's finding shows that LLM use early in a task shapes the whole task, not just the part it touches.
It compounds with retraining cost. The skill you didn't practise this month gets harder to use next month, which makes the AI loop the path of least resistance again.

How it relates to other framings

skill-atrophy-supervision-paradox is Lars Faye's heavier version aimed at senior engineers in the agent-supervisor role: the comprehension required to review agent output is the same comprehension agent use erodes. Cognitive debt is the underlying mechanism; Faye describes the role-level consequence.

agentic-coding-fatigue is the decision-fatigue companion: the quantity of micro-judgments grinds you down even when each is small. Cognitive debt is what's eroded; fatigue is what's depleted. Both compound.

fooling-yourself-with-ai is the same self-assessment failure measured on speed rather than comprehension. METR's experienced developers believed they had finished faster with AI and were measured at roughly 19% slower — the artifact looked fine, and their estimate of their own state was wrong in the same direction as the 83% who couldn't quote their own essay.

programming-as-theory-building gives the older frame for what's being lost. Naur's claim is that the team's shared theory of the program is the primary product; code is a projection. Cognitive debt is the degradation of that theory in any individual head.

ironies-of-automation is the version from human factors, forty years older: skills and system knowledge decay when a machine does the routine work, and the person is left to supervise it without them.

owning-ai-written-code explains why the debt is new: writing code used to force understanding it, so nobody counted understanding as a separate cost until agents took over the writing. writing-code-vs-building-software gives the team-level version: a 3,000-line generated pull request adds 3,000 lines to what the team must understand, and a passing test suite hides that the understanding never arrived.

The countermeasures

The Osmani prescription (see dont-outsource-learning) is posture-based: hypothesise before asking, ask for explanation before code, treat output like a junior's PR, re-derive by hand occasionally, ask the model to teach you what it just did. None of these are dramatic; they're small interruptions in the default loop. The point is that the defaults won't choose them for you.