Mirandola Research

The Capability-Conversion Problem

When does greater machine capability become durable human or institutional capability?

The analytical distinction

Performance and capability are different outcomes.

Artificial intelligence can improve what a person or institution accomplishes immediately. That improvement does not by itself establish that the person or institution has become more capable.

Machine capability What an intelligent system can do.

AI-assisted performance What a person or institution can accomplish while using it.

Durable capability What remains as a reliable capacity for future understanding or action.

The arrows are empirical relationships, not assumptions. The three constructs may rise together. They may also diverge.

Formal definition

Capability conversion is a change in durable capacity.

Capability conversion is the process through which machine capability or AI-assisted performance contributes to durable human or institutional capability.

Durable capability is not synonymous with performing the same task without artificial intelligence. A capability is durable when a reliable capacity persists beyond a transient machine output and contributes to future understanding or action. At the institutional level, that capacity may be embedded in people, knowledge, practices, relationships, or governable human-machine systems.

A causal estimand, where the design permits one

CC_u,d(Δ) = C^AI_u,d,t₁+Δ − C^cf_u,d,t₁+Δ

For unit u, capability domain d, and a meaningful horizon Δ, capability conversion is the difference between durable capability after AI-enabled work and the capability expected under the relevant counterfactual.

Positive conversion means durable capability increased. Zero conversion means an assisted performance gain left no detectable durable change. Negative conversion means capability deteriorated relative to the counterfactual. Where causal identification is not credible, research should report the observed capability trajectory and the evidence supporting attribution rather than imply causal precision it does not possess.

There is no general reason to define a universal “conversion rate” as capability gained per unit of machine capability. The constructs need not share a common scale. Measurement should follow the capability domain under study.

Two levels of conversion

People and institutions can develop differently.

Individual capability conversion

Individual conversion concerns what changes durably in the person: knowledge, judgment, problem formulation, expertise, creative range, calibrated delegation, authorship, or the ability to perform and adapt on later tasks.

Useful evidence can include delayed retention, transfer to new problems, explanation, error recognition, independent performance, improved question formation, or better direction and evaluation of intelligent systems.

The strongest test depends on the capability. Unaided performance is often informative, but it is not the only valid test of development.

Institutional capability conversion

Institutional conversion concerns what changes durably in the organization: its ability to pursue purposes, combine knowledge, allocate authority, preserve understanding, learn from experience, adapt, and execute reliably.

Evidence can include altered routines, durable knowledge capture, repeatable execution, stronger exception handling, changed decision rights, learning across cases, resilience to personnel or model substitution, and improved ability to direct machine-enabled work.

Institutional capability is not simply the sum of individual learning. An institution may become more capable even when execution remains machine-dependent, provided the resulting capability is reliable, intelligible, governable, and available for future action.

Why assisted performance is insufficient evidence

The same immediate gain can produce different developmental outcomes.

A performance measure taken while the tool is available identifies what the human-machine combination can accomplish at that moment. It does not identify what changed in the person or institution after the episode.

Performance gain

Noy & Zhang

In a randomized experiment with 453 college-educated professionals, ChatGPT reduced time on professional writing tasks by 40% and raised output quality by 18%.

Clear evidence of assisted performance. The experiment was not designed to establish durable capability formation.

Boundary dependence

Dell’Acqua et al.

A field experiment with 758 knowledge workers found large gains on tasks inside the model’s capability frontier and worse performance on a task outside it.

Immediate benefit depends on the relationship between task and system capability; performance alone does not reveal what users learned.

Negative conversion

Bastani et al.

In a field experiment with nearly 1,000 high-school mathematics students, access to a general-purpose GPT interface improved practice performance, but students later performed worse when access was removed. A tutor design with learning safeguards largely mitigated the loss.

Better assisted performance can coexist with weaker later performance; interaction design can change the developmental result.

Positive conversion

Kestin et al.

A randomized controlled trial in undergraduate physics found that a research-based AI tutor produced greater measured learning gains in less time than an in-class active-learning condition.

AI assistance can support capability formation when the system is designed around learning rather than answer production.

The empirical question is not whether AI helps. It is what kind of help leaves what kind of capability behind, for whom, under which conditions, and for how long.

Adjacent literatures

A narrower gap than “capability formation is unstudied.”

Capability formation is already studied across several mature traditions. The Capability-Conversion Problem builds on them. Its narrower contribution is to make the relationship among machine capability, AI-assisted performance, and durable human or institutional capability the primary object of inquiry, especially under AI-mediated delegation.

Cognitive offloading

Primary question: What happens when memory or cognition is externalized?
What it contributes: Shows that access to an external resource can change what people retain and what they remember how to retrieve.
Conversion question: When does AI-mediated offloading support higher-order capability, and when does it weaken knowledge needed for later judgment?

Automation complacency

Primary question: What happens to monitoring and vigilance when automation usually works?
What it contributes: Shows that automation can create overreliance, under-monitoring, and new human-system failure modes.
Conversion question: Which institutional designs make reliable execution less dependent on continuous individual vigilance while preserving meaningful oversight?

Human–AI complementarity

Primary question: How does AI change output, quality, time, wages, and task allocation?
What it contributes: Measures the immediate productive effects of machine assistance and where those effects vary across tasks and workers.
Conversion question: Do performance gains leave greater underlying skill, judgment, and adaptive capacity behind after the assisted episode ends?

Expertise formation

Primary question: How do structured practice, feedback, and repeated performance develop expertise?
What it contributes: Provides theories and methods for studying durable skill formation rather than one-time task success.
Conversion question: When does AI accelerate useful practice and feedback, and when does delegation remove the practice through which expertise would otherwise form?

Skill formation across time

Primary question: How do capabilities accumulate, interact, and alter later learning?
What it contributes: Treats capability as dynamic and path-dependent rather than fixed.
Conversion question: How does machine-supplied capability change the sequence through which human capabilities become self-reinforcing or fail to develop?

Organizational learning

Primary question: How do institutions convert experience into altered knowledge and practice?
What it contributes: Makes learning an organizational property rather than only an individual one.
Conversion question: When AI performs the work, what causes the experience to become institutional knowledge rather than remaining externalized in machine performance?

The research program therefore does not claim to supersede offloading, automation, complementarity, expertise, skill-formation, or organizational-learning research. It asks how AI-mediated delegation changes the developmental processes those literatures already describe.

Open empirical questions

What would a cumulative research program need to learn?

Individual conversion

Which forms of AI use improve retention, transfer, judgment, problem formulation, and independent performance over time?
How do effects differ between novices, intermediates, and experts?
When does delegation accelerate expertise formation, and when does it remove the practice through which expertise develops?
Which interaction designs preserve human authorship while still exploiting machine capability?

Institutional conversion

Which combinations of authority, evidence, workflow design, knowledge capture, and review convert machine performance into institutional capability?
When does an institution become more capable rather than merely more dependent on a particular model, vendor, or expert operator?
How should institutions measure capability preserved in routines, relationships, knowledge systems, and governable human-machine systems?
How do AI-enabled changes alter apprenticeship, professional formation, institutional memory, and future decision rights?

Across levels and time

When can an individual gain capability while the institution loses it, or the reverse?
How are gains and losses distributed across roles, cohorts, career stages, and professional groups?
Which short-run performance gains predict long-run capability, and which conceal later deterioration?
What evidence is sufficient to attribute a capability trajectory reasonably to AI-enabled work without requiring impossible causal certainty?

Research design

Measure development across time and levels.

Separate the constructs.

Measure machine capability, assisted performance, and durable capability independently rather than inferring one from another.

Use delayed and transfer measures.

Capability claims should survive beyond the assisted episode. Retention, transfer, adaptation, and later judgment often matter more than immediate output.

Specify the counterfactual.

A capability trajectory is meaningful only relative to a credible alternative: prior practice, another tool, another workflow, another learning design, or an appropriate comparison group.

Study multiple levels.

Individual, team, organizational, professional, and institutional effects may differ. Research should identify where capability resides and where it moves.

Study the path, not only the output.

Delegation, feedback, review, error correction, evidence, and knowledge capture are candidate mechanisms through which conversion occurs.

Make distribution visible.

Aggregate improvement can conceal deterioration in novices, experts, career pathways, organizational functions, or other groups whose capabilities matter.

From research object to institutional program

One problem, several forms of inquiry and application.

Human capability

Studia Humanitatis

The educational framework asks how AI-assisted work can become durable human formation.

Explore human formation →

Empirical and theoretical argument

HumanitAI

“Capability as the Dependent Variable” develops the argument, maps adjacent literatures, and states falsifiable theses.

Read the HumanitAI essay →

Research program

The Intellectual Agenda

The Agenda places capability conversion within the four pillars of AI Humanism and a broader cumulative research program.

Read the Intellectual Agenda →

Institutional application

AI Humanist Institution

The institutional framework asks what organizations become more capable of as machine capability expands.

Explore the institutional framework →

Normative implementation

Provisional Standard v1.0

The Standard turns capability trajectory into a prospective institutional process: Identify → Commit → Observe → Respond.

Read the Provisional Standard →

Selected foundations and evidence

References

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. DOI ↗
Parasuraman, R., & Riley, V. (1997). “Humans and automation: Use, misuse, disuse, abuse.” Human Factors, 39(2), 230–253.
Bainbridge, L. (1983). “Ironies of automation.” Automatica, 19(6), 775–779. DOI ↗
Noy, S., & Zhang, W. (2023). “Experimental evidence on the productivity effects of generative artificial intelligence.” Science, 381(6654), 187–192. DOI ↗
Dell’Acqua, F., et al. (2026). “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality.” Organization Science, 37(2), 403–423. DOI ↗
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences, 122(26), e2422633122. DOI ↗
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports, 15, 17458. DOI ↗
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. DOI ↗
Cunha, F., & Heckman, J. J. (2007). “The technology of skill formation.” American Economic Review, 97(2), 31–47. DOI ↗
Argyris, C., & Schön, D. A. (1978). Organizational Learning: A Theory of Action Perspective. Addison-Wesley.

Deliberate-practice research is included as an important expertise-formation tradition, not as a settled claim that practice alone explains expert performance. Later research has debated the magnitude and interpretation of the original findings.

The research question

What reliable capability exists because machine capability became available?

That question turns AI adoption from a point-in-time performance question into a developmental one.