Capability as the Dependent Variable | HumanitAI

Under what conditions does an increase in machine capability produce a durable increase in human or institutional capability?

This is the question the Mirandola Institute exists to answer. It is not a new worry dressed in new language. Three established literatures already study pieces of it — cognitive offloading, automation complacency, and the labor economics of human-AI complementarity. Each has real findings. None of them, on its own terms, asks what we are asking.

This essay states that argument as five theses and invites anyone, without credential or introduction, to challenge them. This is, deliberately, an old form. The Institute’s namesake opened his own century with nine hundred theses and an offer to defend every one of them, in public, against any scholar in Europe who wished to try. We are not claiming his ambition. We are borrowing his method.

Three literatures, three dependent variables

Cognitive offloading asks what happens to memory when retrieval is externalized. The founding result is Sparrow, Liu, and Wegner’s 2011 study in Science, which found that people who expect information to remain accessible online encode the information itself less effectively, while encoding where to find it more effectively — the “Google effect.” A 2025 MIT Media Lab study (Kosmyna et al., released as a preprint and not yet through full peer review) extends this to generative AI directly: writers using an LLM showed weaker neural connectivity than those using a search engine or working unaided, and struggled to recall or quote from essays they had just written — “cognitive debt.” The dependent variable throughout is narrow and specific: what a person retains in memory after using a tool.

Automation complacency asks what happens to vigilance when a system can be trusted to act on its own. Parasuraman and Riley’s 1997 framework — use, misuse, disuse, and abuse of automation — has organized decades of aviation and clinical research on why skilled operators stop monitoring systems that usually work, and what happens when those systems fail. The dependent variable is behavioral: does the human keep watching. The remedies proposed tend to live at the level of interface design and individual training. The literature diagnoses an individual failure mode inside an otherwise unexamined institutional structure.

Complementarity economics asks what happens to output and wages when workers gain access to AI. Noy and Zhang (2023), in a randomized experiment with 453 professionals, found that ChatGPT raised average output quality on writing tasks and reduced the gap between stronger and weaker performers. But the mechanism, in the authors’ own words, was that the tool “mostly substitutes for worker effort rather than complementing worker skills.” Dell’Acqua and colleagues’ 2023 field experiment with 758 BCG consultants found large gains on tasks inside GPT-4’s capability range and a 19-percentage-point degradation on tasks outside it, driven partly by consultants misjudging where that boundary sat. Both studies measure output. Neither was designed to determine whether the humans involved became more or less capable, independent of the tool, by the end of the study.

Five theses

Thesis 1. Cognitive offloading research measures what a person retains in memory after using a tool. It does not measure, and is not designed to measure, what happens to that person's judgment, expertise, or institutional capacity.

Thesis 2. Automation complacency research diagnoses a failure of individual vigilance and proposes remedies at the level of the individual and the interface. It does not, as a rule, ask whether institutional structure — mandate, bounded authority, evidence, escalation — could make the outcome independent of any one person's sustained attention.

Thesis 3. Complementarity economics measures output, quality, and wages, and treats skill largely as a fixed variable that determines who benefits. It does not, as a rule, measure whether the underlying skill itself grew, shrank, or was substituted for.

Thesis 4. No literature currently in wide circulation tracks the same person or institution across individual, institutional, epistemic, and governance levels at once, longitudinally, to ask a developmental question: not did performance improve, but what remains, and under what conditions would more remain.

Thesis 5. A research program that makes capability formation itself the central, cross-level, longitudinal dependent variable — rather than an incidental finding inside a study built to measure something else — is therefore a distinct contribution, not a restatement of existing work.

None of these theses claims the adjacent literatures are wrong. Thesis 4 in particular is a claim about absence, which is the easiest kind of claim to falsify: a single existing study that does track capability formation longitudinally, across levels, as its primary design, would defeat it outright. That is by design. A thesis that cannot be defeated by evidence is not worth defending.

A narrower gap than it may first appear

The three literatures above are not the only ones that study how capability forms. Expertise-acquisition research, following Ericsson, Krampe, and Tesch-Römer’s (1993) account of deliberate practice, studies exactly the developmental question Thesis 4 asks — what conditions convert sustained effort into durable skill — in detail this essay does not attempt to reproduce. Organizational-learning research, following Argyris and Schön (1978), studies how institutions convert experience into altered practice, which is a close cousin of what this Institute calls Institutional Integrity. Skill-formation economics, following Cunha and Heckman’s (2007) model of dynamic complementarity — the finding that skills acquired early make later skill acquisition more productive, not merely additive — studies capability formation as a rigorous, longitudinal, quantitative object in its own right.

Thesis 4 would be false if any of these literatures already did what this essay claims none of them does: track a person or institution across individual, institutional, epistemic, and governance levels at once, in relation to AI-mediated delegation specifically, as a primary research design. None of them does this, because none of them was built to. Deliberate practice studies skill formation through sustained, structured effort, not through the specific dynamics of delegating a task to a system capable of performing it unassisted. Organizational learning studies institutions adapting to experience, not to a form of capability that can be borrowed from outside the institution entirely, on demand, at near-zero marginal cost. Dynamic complementarity models how a person’s own skills build on one another; it does not model what happens to that sequence when a machine can supply, at any stage, the skill the sequence assumes the person is building.

The gap this essay identifies is accordingly narrower than “capability formation is unstudied.” It is not. The gap is that no existing framework asks how AI-mediated delegation interacts with the capability-formation processes these adjacent literatures have already established — whether delegation accelerates the sequence Cunha and Heckman describe, substitutes for the practice Ericsson’s tradition treats as necessary, or bypasses the institutional learning Argyris and Schön describe entirely. That is a real and, we think, still-open question. It is also a considerably more precise one than the essay’s first draft posed, and the precision is owed to those who pointed out the omission.

What this is not

This is not a claim that AI Humanism supersedes offloading, complacency, or complementarity research. The Capability-Conversion Problem is built directly on what those literatures have already established; it could not be asked without them. The claim is narrower: that none of them, individually or together, currently treats capability formation as the thing to be measured, cultivated, and designed for, across the levels at which people actually act.

An invitation to prove these wrong

The Institute takes its name, and this format, from Giovanni Pico della Mirandola, who at twenty-three proposed to defend nine hundred theses in Rome against any doctor of theology who would come argue them, regardless of that doctor’s rank or his own. Any of the five theses above may be challenged in writing, by anyone, without credential.

This essay’s claims are subject to challenge under the terms of the Institute’s Disputation Protocol. Editorial disclosure, including this publication’s AI review process, appears in the closing section of this edition.

References

Argyris, C., & Schön, D. A. (1978). Organizational Learning: A Theory of Action Perspective. Addison-Wesley.

Cunha, F., & Heckman, J. J. (2007). The technology of skill formation. American Economic Review, 97(2), 31–47.

Dell’Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Technology & Operations Management Unit Working Paper No. 24-013. Published in Organization Science (2025).

Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406.

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv preprint arXiv:2506.08872.

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.

Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253.

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.