When AI Does the Work, What Is Learning For? Post-Instrumental Learning and the Risk of Capacity Dissolution

Synthesis: Argues that as AI systems become capable of producing the artifacts through which institutions recognize competence, existing ethical frameworks centered on AI failures become insufficient. Develops the concept of "post-instrumental learning" and warns that each technical improvement appears to weaken the case for human learning itself, risking "capacity dissolution."

Key Findings

• If the case for learning rests only on current AI failures — bias, opacity, hallucination, labor extraction, privacy risk, and weak accountability — then each technical improvement appears to weaken it.
• The article develops an idealization of AI that executes specified tasks flawlessly while lacking authority over purposes, legitimacy, and responsibility, and argues for post-instrumental learning that preserves the capacities people and institutions need when many useful outputs can be delegated.
• Five capacities are analyzed — end-setting, reason-giving, contestability, refusal/revision, and participation — and their erosion is named capacity dissolution.
• The central case is Assessment under generative AI: when a polished artifact no longer reliably evidences understanding, institutions must assess the learner's accountable relation to AI-mediated work rather than the artifact alone.
• The takeaway is practical: AI governance should evaluate not only whether systems perform well, but also whether their deployment leaves people able to understand, challenge, revise, and share responsibility for the practices those systems mediate.

The Central Case: Assessment under Generative AI

The argument opens with a recognizable scene: a student submits a fluent essay and cannot explain its central claim; a manager approves an elegant market analysis without knowing which assumptions make it plausible. In each case the output looks successful, but the relation between the person, the institution, and the reasons behind the work has become thin. Existing criticisms of large-scale AI systems — that they reproduce inequality, hide responsibility, depend on invisible labor, and deepen relational injustice — remain indispensable, but they do not fully answer the educational question. Treating learning merely as preparation for producing outputs makes AI improvement a reason to produce those outputs with less human effort, a conclusion the author argues is too quick.

Implications for AI Governance

The article translates the argument into a deployment-review question: what will people stop learning if this system works? Losing obsolete routines may be harmless, but losing the ability to set goals, give reasons, contest, and participate is not. The practical right of contestability is not a demand that everyone master every technical detail: a student need not reproduce every step of a model's generation to defend a claim made in a paper. Rather, each person needs a path from a problem they experience to reasons they can understand and, where appropriate, to a process that can revise the decision — a path that depends on trained intermediaries, accessible records, meaningful appeal, and public settings where reasons can be tested. The article warns that when privileged users can contest AI while marginalized users receive automated completion or automated suspicion, the institution has changed the distribution of expertise while calling the result access. For AI Governance and Over-Reliance debates, this reframes the goal of schooling: preserving capacities such as Ethics-inflected judgment and accountable participation rather than maximizing efficient output generation.

What this means for practice

• Instructors. Ask of every assessment, "what will this student stop learning if the AI does this work?" and require students to defend source selection, explain revisions, and answer for the claims in their artifacts.
• Instructors. Build at least one task per course around the four limits that perfect execution cannot remove — end-setting, reason-giving, contestability, and participation — rather than adding AI-detection rules.
• Researchers. Treat validity as the first question of an AI policy rather than cheating: measure whether an assignment still evidences the capacity it claims to assess.
• Administrators. Guarantee a contestability path — records, trained intermediaries, meaningful appeal, and public settings where reasons can be tested — so that AI-mediated decisions remain answerable.
• Researchers. Test the dissolution claim directly by tracking which capacities decline once a delegated task is automated, instead of only whether outputs improve.

Limitations

• This is a conceptual article: it reports no sample, intervention, or outcome measure, so the capacity-dissolution claim rests on argument rather than data.
• The analysis is built on an idealization — a "perfect AI" that executes specified tasks flawlessly once goals, constraints, and roles are given — which is a deliberate analytic device, not an observed system.
• The five capacities (end-setting, reason-giving, contestability, refusal and revision, participation) are analytic categories rather than validated instruments, so the argument cannot estimate how fast or how far dissolution proceeds in a given program.
• Illustrative scenes, such as a student who cannot explain the essay they submitted, carry the empirical weight; no comparison of institutions with and without AI delegation is offered.

Citation

Kai Yao (2026). When AI Does the Work, What Is Learning For? Post-Instrumental Learning and the Risk of Capacity Dissolution.