Chatbot use erodes the user's underlying skills

The argument

Skill in any complex domain is built and maintained through deliberate, often uncomfortable practice — writing the bad first draft, hitting the dead end, looking the answer up properly. Chatbots short-circuit that loop. A 2025 study from Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers and found that those who reported higher confidence in generative-AI tools reported correspondingly lower critical-thinking engagement on the tasks they delegated to those tools; their cognitive effort shifted from problem-solving to checking the model's output. The pattern mirrors earlier findings on GPS use, where heavy reliance on turn-by-turn navigation was associated with measurably worse independent spatial memory, and on calculator use in basic arithmetic. Each generation of automation has degraded the underlying competence in those who delegate the work; chatbots are uniquely positioned to do so because they cover almost the entire surface of literate knowledge work. Over a decade of widespread adoption, the predictable outcome is a workforce able to produce average-quality output instantly when the tool works, and largely unable to recover when it does not. The user appears more capable than ever, until the moment the model is wrong, unavailable, or applied to a problem outside its frontier — and at that moment the user is materially worse off than they were before.

Premises

Counter-arguments

Proponents reply that the skill-erosion worry conflates correlation with causation and over-generalises from early findings: the Microsoft/CMU study measured self-reported reliance and in-the-moment critical-thinking engagement, not long-run competence loss, and confident users delegating routine tasks may simply be reallocating effort to higher-value work. Every prior automation — calculators, spellcheckers, compilers — provoked the same deskilling fear, yet freed practitioners to operate at a higher level of abstraction rather than leaving them helpless. Whether a tool erodes or extends skill depends on how it is used; used as a checkable aid to learn from, chatbots can build competence, so "net harm" does not follow from the mere possibility of misuse.

Rejecting the premises

[Rejecting P1] That competence requires effortful practice is true, but it does not follow that every task must be done unaided; offloading routine sub-tasks can free effort for the harder practice that builds higher-level skill. [Rejecting P2] The cited study measures self-reported reliance and momentary engagement, not durable skill loss; reallocating cognitive effort is not the same as eroding underlying competence. [Rejecting C] Since the effect depends on how the tool is used — as a crutch or as a checkable aid to learn from — sustained chatbot use need not erode skill, so "net harm" is not established.