Technical note 05
Offloading without Atrophy: The Question the Moral Framing Was Standing In Front Of
JDM Inc
Strip the moral vocabulary from the objection and a real empirical question is left standing. Cognitive science has been working on it for thirty years, and it is not the question of whether offloading is permissible.
The source paper makes a move at the end of Section 6 that is easy to read past. Having argued that the refusal is identity work and the conversion is not an update, it declines to conclude that the concern was empty. It says the opposite: the worry survives the social explanation, and it has independent support.
This note is about what the worry becomes once the moral vocabulary is removed from it — which turns out to be a well-posed and partially answered research question rather than an open one.
Offloading is not a pathology
The first correction is definitional. Using external resources to reduce internal cognitive demand is a normal and often adaptive feature of human cognition, not a failure of it (Risko & Gilbert, 2016). Writing is offloading. So is a checklist, a calculator, a colleague, and a filing system.
The extended-mind position goes further: under the right coupling conditions, external resources constitute part of a cognitive system rather than merely serving it (Clark & Chalmers, 1998). On that view the boundary the objection is trying to police — between thinking you did and thinking the machine did — is not where intuition places it.
Anyone who has ever written something down in order to think about it has already conceded the general case. What remains is the specific one.
Offloading changes what is retained
The second finding is the one that keeps the worry alive.
Sparrow and colleagues found that when information is reliably accessible, people remember where to find it rather than the content itself (Sparrow et al., 2011). This is the result usually cited in support of the degradation thesis, and it does show a real change. But notice what kind of change it is. It is a reallocation, not a loss: memory for location displaced memory for content, in a context where location was sufficient.
Whether a reallocation of that shape is a harm depends entirely on whether the retained capacity is the one that was needed. For a fact you will always be able to look up, it is not. For a judgment you have to make under time pressure without the system, it might be.
This is what turns the question from a moral one into a design one. The paper states it exactly: the open question is not whether offloading occurs but which capacities it preserves and which it lets atrophy, and under what conditions each occurs.
The seam
The most useful result available is the one that contains both outcomes in the same study.
Bastani and colleagues gave roughly a thousand secondary students access to a plain GPT-4 chat interface. Practice performance rose substantially. Subsequent unassisted exam scores fell relative to a control group. A second version, scaffolded with pedagogical safeguards, produced the practice gains without the exam decrement (Bastani et al., 2024).
The three things this establishes are worth separating.
First, the harm is real. The plain interface produced a measurable capability decrement on unassisted work. Anyone dismissing the concern as pure moral panic has to account for that arm of the study.
Second, the harm is not intrinsic to the tool. The same underlying model, differently configured, did not produce it.
Third, and most usefully, the difference between the two arms was a design difference. Not a policy difference, not a difference in user virtue, not a difference in how much effort the students were required to feel. The variable that moved the outcome was how the system was built.
The caveats belong here rather than in a footnote: one subject, one school system, secondary students rather than professionals. It does not generalize on its own, and the paper is careful not to let it. What it does is convert a moral dispute into an engineering specification with at least one known-good configuration.
What the moral framing cost
The paper’s sharpest practical claim is that framing this question morally has actively impeded answering it.
The evidence is institutional. A large body of educational AI policy was written on the assumption of a cheating epidemic, while survey data from three US high schools found self-reported cheating behaviors remained relatively stable after the release of ChatGPT (V. R. Lee et al., 2024).
So the resources that could have gone toward the atrophy question — which is real, and which has at least one experimental result pointing at a mitigation — went instead toward detection, prohibition and enforcement aimed at a behavior that does not appear to have moved much. That is not a small opportunity cost, and it is the specific damage the identity framework predicts: a framework that mistakes an identity threat for an integrity crisis will generate rules aimed at the wrong behavior.
Where this lands for us
Two things follow for anyone building in this space, and we hold them both.
The first is that the concern deserves better than dismissal. A vendor whose answer to “will this erode my team’s judgment” is that the question is a moral panic has not read the Bastani result. The decrement was measured. It is available to any system built without regard for it, and building one is the default rather than the exception.
The second is that the mitigation is a build problem, and the direction it points is the same one the legibility note arrives at from the identity side. A system in which the user selects, weighs, rejects and overrides — and in which every assertion carries the source it rests on — keeps the user performing the judgment rather than receiving it. Whether that specific configuration preserves the capacities that matter is an open question we have not tested, and we would rather name it as open than claim it.
What we will say is that it is the right question, that it is answerable, and that it is a different question from whether using the tool makes you a worse person.
References
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2024). Generative AI can harm learning. SSRN Working Paper No. 4895486.
Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7-19.
Lee, V. R., Pope, D., Miles, S., & Zárate, R. C. (2024). Cheating in the age of generative AI. Computers and Education: Artificial Intelligence, 7, 100253.
Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676-688.
Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776-778.