Cognitive debt is the accumulated cost of repeatedly outsourcing mental effort: each delegation saves time immediately and leaves a deficit in learning and critical engagement that compounds across subsequent tasks.
The term comes from MIT Media Lab's 2025 study Your Brain on ChatGPT. It has since escaped into general use, usually stripped of the evidence that produced it. So here is what was actually measured, why the metaphor is a good one, and where it stops working.
What was measured
Fifty-four participants wrote essays across three conditions — using an LLM, using a search engine, or using nothing — while researchers recorded EEG. Each participant stayed in one condition across three sessions; eighteen returned for a fourth in which conditions were swapped.
The LLM group showed the weakest and least distributed neural connectivity of the three. The unaided group showed the strongest. The search engine group sat between them, which matters: it suggests the effect scales with how much of the thinking the tool absorbs, rather than being a property of "using a computer".
Across sessions, the LLM group scored worst at the neural, linguistic and scoring levels.
The finding that actually lands
The brain imaging gets the headlines. The result worth remembering is simpler.
Asked to quote from essays they had submitted minutes earlier, LLM-group participants frequently could not.
Not "struggled to recall the argument". Could not produce a line of text they had ostensibly just written. The work had passed through them without leaving a trace.
That is the clearest demonstration I know of that producing output and learning are separable, and that we have built a tool which separates them efficiently. The essay existed. The encoding did not happen.
Why "debt" is the right metaphor
Financial metaphors in technology are usually lazy. This one earns its keep on three counts.
The cost is deferred, not avoided. You get the time back now. The deficit shows up later, as a weaker foundation for the next task.
It compounds. A weaker foundation makes the next task harder, which makes delegating that task more attractive, which deepens the deficit. The loop is the mechanism, not a side effect.
It is invisible on the balance sheet you look at. Your output goes up. Your throughput goes up. Nothing in your daily experience reports the liability, which is exactly why technical debt gets its name too.
Where the metaphor breaks
Two places, and both matter.
Debt has a principal. You can state what you owe. Cognitive debt has no quantity — nobody can tell you how much you have accrued, and the MIT study offers no unit. It is a description of a direction, not a balance.
Debt is repayable on known terms. Skill-retention research says unused capabilities recover with deliberate practice, so something like repayment exists. But the learning that would have happened during the task you delegated is not recoverable by doing something else later. You cannot re-run the encoding you skipped. What you can do is take on comparable difficulty again, which is repayment in the loose sense that going to the gym repays a sedentary year.
What it does not mean
The study has been reported as evidence that ChatGPT damages your brain. It does not show that, and the authors do not claim it.
It measures cognitive engagement during assisted work. Finding reduced neural activity while a machine does part of the task is close to definitional — that is what delegation is. The genuinely novel result is the memory failure, and even that is 54 participants on one task with one model, in a preprint.
The larger claim — that this accumulates into lasting capability loss — is inferred from skill atrophy research that predates AI by decades. The inference is reasonable. It is not the same as having been tested.
What to do with the idea
Cognitive debt is most useful as a question to ask yourself in the moment, and the question is not "am I using AI too much".
It is: when I delegated that, did I skip the part where I would have learned something?
Answering a factual lookup: no debt. You were never going to encode it. Having an argument constructed for you on a topic you are meant to understand: debt, and the fluent result is what stops you noticing.
The distinction is beneficial versus detrimental offloading, and it is the only part of this conversation that translates into a decision you can actually make.
If you want the underlying studies with what each does and does not establish, they are indexed on the research page. If you want a fixed measurement of your own work to compare against later, that is what the Drift Score is — and how it is calculated, including its own considerable limits, is published in full.