Output Is Easy to Count, But Capability Isn't - Open Magazine

TEXT START: A 2025 study of 319 knowledge workers, conducted jointly by Microsoft and Carnegie Mellon University, found that greater confidence in generative AI was associated with lower reported critical thinking effort, with critical thinking itself shifting away from generation and toward verification, response integration, and task stewardship.

The Dissection

The essay identifies a real secondary effect of automation: AI-assisted output can rise while unaided human capability decays. Its central move is to rename this risk “developmental safety” and convert it into an organisational design obligation. The proposed cure—capability withdrawal tests, protected cognitive struggle, apprenticeship, and supervised disagreement—treats deskilling as a governance failure that competent managers can correct.

That diagnosis is useful but strategically subordinate. The essay studies the condition of the human worker inside an organisation while barely confronting the system-level consequence: once AI performs economically necessary cognitive work more cheaply and reliably, the organisation’s incentive is to remove developmental labour, not preserve it. The dashboard is not merely blind. It is often measuring exactly what the institution wants: output after the human developmental substrate has been externalised.

The Core Fallacy

The core fallacy is treating preserved human capability as though it preserves human economic necessity.

Even if the proposed withdrawal protocol proves that workers have become worse at unaided reasoning, that does not reverse the Discontinuity Thesis. It only measures the decline more elegantly. A workforce can retain judgment, discernment, and apprenticeship while losing its position in production. Human capability and human bargaining power are not the same asset.

The essay also assumes that organisations facing superior cognitive automation will voluntarily retain inefficient developmental work because it benefits long-term capability. That is a weak assumption. Competitive systems reward immediate cost, speed, scale, and measurable output. “Protect the struggle” is structurally fragile when the struggle can be automated and the worker who needs it can be replaced by a system that does not need to develop.

The article correctly sees that AI may hollow out the people meant to supervise it. It misses the harsher implication: this creates a temporary management problem, not a durable human domain. Once supervision, verification, and correction themselves become automatable, the supposed residual value of independent human judgment narrows further.

Hidden Assumptions

• Human unaided capability will remain economically valuable enough to justify its preservation.
• Organisations will sacrifice near-term efficiency to maintain developmental experiences.
• Human-only or human-led cognitive domains can be stabilised at scale.
• Apprenticeship and deliberate practice can survive after the underlying work is economically displaced.
• Better measurement will produce corrective action rather than merely document declining human necessity.
• AI dependence is primarily a workflow-design problem rather than a competitive pressure.
• Human judgment remains the final authority when systems become more capable, cheaper, and more scalable.
• The transition can be managed by individual organisations without being dictated by capital competition.
• “Capability return” is a meaningful organisational objective even when the market rewards output return instead.
• The damage is reversible if institutions preserve enough struggle. That may be true for skill formation and irrelevant to the larger employment circuit.

Social Function

This is a partial truth functioning as transition management and elite self-exoneration.

It accurately warns that AI can produce capability atrophy, automation bias, and a thinner supervisory bench. But it relocates the crisis from political economy to organisational practice. The remedy is framed as better design, better diagnostics, and more conscientious management—not ownership of AI capital, control of productive infrastructure, or the collapse of mass labour’s bargaining position.

That makes the essay useful to institutions preparing for automation. It gives them a vocabulary for managing the degradation of human capability without questioning the displacement that causes it. The organisation can claim to protect judgment while continuing to remove the work through which judgment was formed. The result is not resistance to the transition. It is a more sophisticated operating manual for surviving it.

The Verdict

The essay sees the smoke but mistakes the fire for a faulty dashboard.

Its empirical warning is credible within the supplied text: AI can increase measured output while weakening independent capability. Its systemic conclusion is inadequate. Developmental safety may preserve some humans as better validators, maintainers, or transition intermediaries for a time, but it cannot restore the mass employment–wage–consumption circuit once AI severs it. The proposed protocol can reveal the corpse’s remaining reflexes. It cannot make the corpse economically necessary again.