Output Is Easy to Count, But Capability Isn’t

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. A 2025 multicentre observational study of experienced endoscopists found that adenoma detection during colonoscopy without AI assistance fell by six percentage points after the introduction of routine AI assistance. The International AI Safety Report 2026 cites the finding as evidence relevant to how AI may affect the maintenance of human skills, while noting that long term evidence remains limited. McKinsey’s 2026 global survey of 1,719 respondents across 97 countries found that while eight in ten reported individual productivity gains, only 37% reported that AI had contributed positively to their organisations’ EBIT. 47% of mid level managers and individual contributors reported at least one negative effect from AI use. 20% of mid level managers said AI was hurting their ability to think critically.

None of these findings proves that AI makes people less capable. They point to a more interesting problem — we are getting better at measuring what people can produce with AI than at measuring what they can still do without it.

AI productivity gains are increasingly visible. Their developmental consequences are comparatively invisible.

AI creates the possibility of a capability paradox. AI can improve performance without improving capacity, and current organisational metrics are calibrated to detect only the first. This is not a claim that AI inevitably makes people less capable. It is an observation that the metrics, organisations currently rely on cannot detect the difference between AI assisted performance and underlying human capability, and that the two can diverge.

To understand what is at stake, we need to separate two things that organisational thinking habitually conflates. Productivity asks how much work gets done. Capability asks what kind of human being is emerging through the work. Work produces both outputs and people, and AI can alter both simultaneously, while organisations remain much better equipped to measure the first than the second.

Philosopher Avigail Ferdman argues in a 2025 paper in AI and Society that AI can create what she calls ‘capacity-hostile environments’, structural conditions that systematically impede the habituation and practice through which human capabilities are cultivated. Deskilling, on this account, is not a failure of individual discipline. It is a property of the organisational environment in which work takes place.

I argue that AI should be understood not merely as a productivity technology but as a developmental perturbation — a force capable of reshaping the conditions under which human judgment, expertise, and discernment emerge.

A productivity technology can be evaluated on its outputs alone. A developmental perturbation cannot. Its full consequences are visible only across time, in what people become or fail to become as a result of working alongside it. A 2025 MIT Media Lab preprint found weaker functional brain connectivity patterns during LLM assisted essay writing than in search engine and brain only conditions, a small, task specific finding that calls for replication, but suggestive of the direction in which AI engagement may alter cognitive work. Today, AI adoption produces measurable productivity gains. In the medium term, as deliberate practice and independent judgment formation become less frequent, expertise may develop differently, shallower, more dependent and less transferable. A 2026 qualitative study of 38 knowledge workers across ten sectors identified four ways knowledge workers perceived their skills changing — development, maintenance, erosion, and revaluation — depending on how they engaged with AI. The outcome is not determined by the technology alone. It is determined by the conditions of use.

Developmental Safety and Capacity Formation

Amy Edmondson’s research established psychological safety as a construct concerning interpersonal risk taking in teams — whether people can ask questions, admit mistakes, challenge assumptions, and disagree without fear of humiliation or reprisal. The AI era makes psychological safety more consequential. A worker may need psychological safety not only to challenge a colleague but to say, plainly, that they think the AI is wrong.

Yet psychological safety answers only part of the problem the AI era poses. A person may feel entirely safe disagreeing with an AI recommendation and still lack the underlying capability to evaluate it independently. That creates a second, distinct question that psychological safety was not designed to address.

I call this second condition developmental safety. It concerns the conditions under which capability continues to form — through practice, cognitive struggle, reflection, and progressively independent judgment. Psychological safety addresses one set of conditions for learning — whether people can take interpersonal risks. Developmental safety addresses another — whether the work itself continues to provide the practice, struggle, and progressive independence through which capability forms. Ferdman’s concept of a ‘capacity-hostile environment’ provides an important structural foundation for this argument. Her analysis draws attention to environments in which the conditions for human capacity cultivation and exercise are restricted. Building on this insight, the more specific question developmental safety asks is, when AI performs part of the work, what conditions must organisations create or preserve so that the work continues to support the development of particular human capabilities?

In this sense, psychological safety enables the courage to question the answer. Developmental safety preserves the capacity to work through the question. Where psychological safety says it is safe to speak, developmental safety says it is safe to remain unresolved. Not to remain incompetent, but as permission to stay cognitively engaged with a problem long enough for genuine understanding to develop. That is the condition AI most readily removes. When AI performs the cognitive work through which people would otherwise have developed — the drafting, the analysis, the synthesis, the difficult recommendation — it does not merely automate a task. It removes a developmental experience. The organisation gets the output. The person does not get the growth.

Research published this year in the Early Childhood Education Journal offers a useful developmental analogy. In a small study of four children aged three to four, open-ended questioning during loose parts play supported problem solving and persistent thinking, but only after teacher modelling and more directive questioning. The study does not establish an organisational parallel, but it suggests a principle relevant to AI-supported work that developmental independence may require structured scaffolding before responsibility can be progressively transferred to the learner.

Capability Return and the Organisational Obligation

AI as a developmental perturbation does not determine its own outcome. A perturbation disrupts existing conditions, but what emerges depends on the choices organisations make in response. Organisations facing cognitive automation therefore face a choice. Either they can allow the displacement of developmental work to quietly erode human capability, or deliberately redesign the conditions under which capability forms. I call this second response capability return — making the human capability that remains, reemerges, or has changed when AI assistance is withdrawn, relative to capability before and during AI use.

Building on the empirical evidence that AI produces heterogeneous skill outcomes, Augmentative Psychology proposes a five part conceptual framework for thinking about possible developmental trajectories of AI engagement in organisations. Augmentation, in which capacity deepens through AI engagement. Compensation, in which AI provides temporary scaffolding without lasting developmental change. Substitution, in which AI progressively performs work in place of the person’s own developing capacity. Atrophy, the possibility that repeated substitution of human cognitive work gradually reduces the person’s unaided capacity to perform it. And incorporation, in which sustained AI engagement restructures the underlying cognitive processes through which the person perceives and judges, a change that can work in either direction, internalising the AI’s patterns and biases as the person’s own, or, under richer conditions, reorganising capacities in ways not previously available. These five trajectories are possibilities rather than predictions, which is precisely why organisations need a framework for detecting which one is actually occurring.

This essay proposes a capability withdrawal protocol for examining whether, and in what form, human capability persists, reemerges, or changes when AI assistance is removed. The protocol would compare unaided performance, reasoning, error detection and performance on novel cases against baseline measures, where possible. The 2026 International AI Safety Report documents automation bias, the tendency to accept incorrect AI recommendations, especially when correction requires additional effort. Capability is not merely the ability to produce an answer. It includes the ability to evaluate the answer. The withdrawal diagnostic can help distinguish substitution from possible atrophy by comparing unaided performance against a baseline, and this is distinct from ordinary tool dependence. The question capability return asks is whether repeated delegation has reduced independent reasoning below its previous baseline, and whether the organisation has any means of detecting that change. For the incorporation trajectory, the relevant test is whether the person’s judgments in situations AI never touched show patterns that were not present before, new errors in the negative case, new capacities in the positive.

Another design nuance is that the same task may be tedious for an experienced practitioner and genuinely developmental for a novice. What is merely efficient to automate at one career stage may be irreplaceable at another. Capability return therefore demands not just workflow design but developmental stage awareness. In practice, this means protecting cognitive struggle, maintaining apprenticeship relationships where expertise is transmitted through practice, and creating conditions for supervised disagreement with AI recommendations so that practitioners develop the epistemic agency to recognise when a confident system is confidently wrong.

None of this is a case against AI. It recognises that augmentation and development require different design logics. AI should remove work that is merely tedious. It should not remove developmental work. And what counts as developmental depends on who is doing it.

Organisations that systematically hollow out the developmental experiences of their employees may eventually discover, over time, that they have a thinner bench of independent thinkers capable of supervising, correcting, or improving the systems they depend on. The productivity gain was real. The capability cost arrives later, and compounds.

What Cannot Be Purchased

The AI era will not ultimately be judged by how efficiently organisations automated work. It will be judged by whether they preserved the conditions under which human judgment continued to emerge. Organisations need to recognise that the central question of AI adoption is no longer only whether AI improves what people can do. It is whether the way people work with AI preserves the conditions through which they remain capable of doing it without AI, and of judging when AI should not be trusted.

Developmental safety names what must be protected. Capability return names the organisational obligation that follows. Both must be built into how work is designed, not merely how it is described.

The dashboard will not tell you when the capability has gone quiet. That is the whole problem.