# The Effort Principle **Why durable reasoning is built by retrieval and problem-solving, and why ambient assistance is the mechanism of cognitive debt** Neuro ASI / Phroneme · Draft for review · 2026 --- ## Abstract Cognitive debt is usually described as an effect, a decline observed after people offload thinking to an AI assistant. This paper states the mechanism instead. Durable reasoning capacity is built by effortful retrieval and by working problems through, and the effort is not overhead attached to learning but the event that produces it. It follows that assistance which removes the effort removes the thing that was building the capacity, whatever it does to the quality of the output. That gives cognitive debt a specific and falsifiable shape: not "AI is bad for you" but "the operation that was doing the work is no longer being performed." We set out the evidence for the effort principle, distinguish the procedural skill of using a tool from the cognitive capacity the tool stands in for, name the limits of the argument honestly, and draw the consequence for measurement. If effort is the mechanism, then the only informative measurement is performance without assistance, tracked over time. ## 1. The claim The claim is narrow on purpose. Reasoning capacity, meaning the ability to carry an unfamiliar problem from premises to a defensible conclusion, is trained by performing the operation, not by being present while a competent output is produced. Reading a good argument is not the same event as constructing one. Watching a solution appear is not the same event as reaching it. Where the two diverge is where the capacity is or is not built. This is not a claim about whether AI assistance is good. Assistance is frequently the correct choice, and a person who uses a tool well to produce better work has done something worth doing. The claim is about what that transaction does and does not deposit in the person, and therefore about what happens when the tool is withdrawn. ## 2. Effort is the event, not the obstacle The clearest demonstration is also one of the more durable results in the learning literature. When people study a prose passage and are then asked to recall it, they retain substantially more of it a week later than people who spend the same time rereading the passage. The reversal is the important part: measured minutes after study, rereading looks better, and only at a delay does retrieval practice pull ahead (Roediger and Karpicke, 2006). Two things follow. The first is that the effort of generating an answer, which feels like difficulty and is often experienced as inefficiency, is the operation that produces retention. The second is that immediate performance is an unreliable guide to what was learned, because the condition that felt easier and looked better in the moment was the one that had deposited less. The scope deserves stating. This is a finding about the retention of studied material under controlled conditions. It is not a measurement of general intelligence, it does not license a claim about how any individual should study, and the size of the advantage depends on the material, the delay, and whether feedback was given. What it establishes is the direction: effortful retrieval and passive exposure are different events with different consequences, and the difference shows up later rather than immediately. There is a structural version of the same point, formalized long before the current tools and in a different discipline entirely. In the standard treatment of learning from consequences, an agent improves by acting, observing the result, and updating on the gap between what it expected and what occurred (Sutton and Barto, 2018). An agent that never acts generates no such gap and has nothing to update on. That framework is computational and is not a claim about human neural mechanism, so we cite it for the structure and not as evidence about people. The structure is nonetheless the same one the retention data show: no attempt, no error signal, no learning. ## 3. Two kinds of learning, and only one of them transfers Using a capable assistant well is a real skill, and it is learned the way skills are learned, by practice. Prompting improves. Judgment about when to delegate improves. Recognizing a bad output improves, at least within familiar territory. None of this is illusory and none of it is worth disparaging. But it is a procedural skill whose object is the tool. Practicing it exercises the operation of the tool. It does not require retrieving the relevant knowledge from memory, because the assistant supplies it. It does not require constructing the intermediate representation of the problem, the partial model a person builds while working through something, because the assistant returns a finished artifact and the intermediate steps are never occupied. That intermediate representation is what a person carries to the next problem, particularly to an unfamiliar one, and it is the specific thing that goes unbuilt. The practical shape of this is that fluency and capacity can move in opposite directions without either being visible to the person. Someone can get measurably better at producing good work with assistance while getting worse at producing it without, and the first improvement is conspicuous while the second decline is not. Nothing in the person's own experience separates these, for the reason Section 2 gives: in-the-moment performance is exactly the signal that fails to distinguish them. ## 4. What the evidence currently supports Two findings bear directly on this and neither is strong enough to carry more than it does. Fifty-four participants wrote essays across three conditions, unaided, with a search engine, and with a large language model, with EEG recorded throughout. The group using the language model showed the weakest measured connectivity across networks associated with planning, integration, and recall, and a majority could not quote a sentence from the essay they had just produced. When participants who had relied on the model later wrote unaided, the reduced-engagement pattern did not immediately reverse. The authors named the residue cognitive debt (Kosmyna et al., 2025). This is a preprint with a small, task-specific sample and a smaller number completing the final session. It establishes that a signal exists and is worth measuring properly. It does not establish long-term harm, a population-level effect, or a causal path, and no one should describe it as though it did. Separately, survey and behavioral work reports an inverse association between frequent AI tool use and critical-thinking performance, statistically mediated by cognitive offloading (Gerlich, 2025). This is correlational and self-report-weighted. It is consistent with the mechanism argued here and it is not evidence of direction. Read honestly, these two results do not prove the effort principle. The effort principle is what makes them expected rather than surprising, and what makes them testable: if effort is the mechanism, the deficit should appear specifically in unaided performance on unfamiliar problems, should not appear in assisted output quality, and should scale with how completely the assistance removed the intermediate work. ## 5. What this argument does not claim We hold these limits in view rather than in a footnote. It does not claim AI use lowers intelligence. The construct at issue is a trainable capacity under standardized conditions, not general ability, and nothing here speaks to the latter. It does not claim a measured brain signal reveals what a person is thinking or what they are capable of. The EEG result above is a group-level correlate under one task. Treating a neural measurement as a read-out of cognition is the error this project's companion paper is devoted to refusing, and the effort principle does not need it. The argument stands on behavioral evidence and would survive the neural finding being withdrawn. It does not claim any particular individual is in debt. Population findings are weak guides to a single person, which is the reason the instrument reports an individual trajectory with its uncertainty rather than a verdict. It does not claim that assistance should be avoided. It claims that if reasoning capacity is something a person wants to retain, the operation that builds it has to be performed somewhere, deliberately, because it is no longer being performed incidentally. ## 6. What follows for measurement If the effort is the mechanism, the measurement follows without much argument. Measure the unaided case, because that is the only condition in which the capacity in question is the thing being exercised. Assisted output measures the pair, and the pair is not what is at risk. Measure at a delay and repeatedly, because Section 2's central result is that the informative difference is invisible immediately and appears later. A single session cannot see it by construction. Measure with unfamiliar problems, because retrieval and transfer are the operations at issue, and a familiar problem can be passed by recognition. Do not measure it from a typed sentence, a questionnaire, an uploaded document, or the output of a model. None of those is the unaided performance of a novel reasoning task, and none becomes one by being processed more cleverly. This is the entire design rationale for measuring behavior longitudinally and treating any neural signal as convergent evidence only. It is not caution bolted onto a product. It is what the mechanism implies. --- ## References 1. Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. *Societies, 15*(1), 6. 2. Kosmyna, N., Hauptmann, E., Yuan, Y. T., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab (preprint, arXiv:2506.08872). 3. Roediger, H. L., and Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. *Psychological Science, 17*(3), 249-255. 4. Sutton, R. S., and Barto, A. G. (2018). *Reinforcement Learning: An Introduction* (2nd ed.). MIT Press.