Capability Compression: Generative AI and the Negotiable Distance to Functional Performance

Abstract

Research on generative artificial intelligence has documented productivity augmentation, performance-gap compression, cognitive offloading, occupational expertise transfer, deskilling, and changes in the value of human capital. These literatures illuminate important consequences of AI assisted work, but they do not consistently distinguish a reduction in the resources required to reach functional performance from a change in competence already possessed by the worker. This paper proposes capability compression as an integrative framework for making that distinction.

Capability compression is defined as a technology associated reduction in the resources a specified individual requires to reach a specified threshold of functional performance on a specified task, under specified conditions of assistance, accountability, and error tolerance. The framework distinguishes three analytically separable thresholds: assisted execution, in which a user can produce an acceptable outcome with technological assistance; supervisory competence, in which the user can evaluate, diagnose, constrain, and correct technologically produced output; and independent competence, in which the user can produce, adapt, or repair the relevant outcome when assistance is absent, degraded, or incorrect. Existing evidence suggests that generative AI can substantially compress distance to the first threshold without proportionately compressing distance to the second or third.

The paper further distinguishes capability compression from erosion of independent competence and proposes a conditional dynamic in which compression may enlarge an individual's reachable task set and, under appropriate conditions, create opportunities for recombination with existing expertise. Knowledge distance, task characteristics, evaluative ability, error tolerance, and access conditions are proposed as boundary conditions. The framework therefore rejects both strong equalization claims and simple deskilling accounts. Performance compression is not necessarily learning compression, and learning compression is not necessarily competence compression.

Technology has long altered the distance between intention and execution. The distinctive possibility posed by generative AI is that capability distance may become unusually negotiable through interactive, iterative assistance across a broad range of symbolic tasks.

Keywords: capability compression, generative AI, skill, deskilling, augmentation, expertise, human capital, labor

---

1. Introduction

Generative artificial intelligence has produced an unusual disagreement about skill.

One observer sees a novice completing work that previously required years of training and concludes that expertise has been democratized. Another sees the same event and concludes that expertise has been automated away. A third points to measurable productivity improvements and describes augmentation. A fourth observes declining demand for particular forms of labor and describes deskilling, substitution, or commoditization.

Each may be observing something real.

The difficulty is that these descriptions frequently refer to different phenomena.

Consider a person with little programming experience who uses a generative AI system to construct a functioning script. Several questions immediately arise.

Can the individual produce an acceptable outcome while the system is available?

Can the individual determine whether the generated code is incorrect, insecure, or inappropriate?

Can the individual modify or repair the program if the system fails?

Has the individual learned anything that persists after assistance is removed?

Has an experienced programmer lost any existing competence because the novice gained this new capacity?

Has the market value of programming changed because acceptable output has become easier to obtain?

These are not different ways of asking the same question.

This paper proposes capability compression as a conceptual framework for separating them.

Capability compression is:

a technology associated reduction in the resources a specified individual requires to reach a specified threshold of functional performance on a specified task, under specified conditions of assistance, accountability, and error tolerance.

The central analytical object is therefore not "skill" in the abstract. It is threshold distance: the multidimensional resource gap between a person's present capability state and a defined performance threshold under specified conditions.

Resources may include time, training, money, procedural knowledge, formal instruction, feedback, expert supervision, motor execution, cognitive effort, infrastructure, iteration cost, and the burden of detecting and correcting errors.

A foundational ontological distinction is necessary at this point:

Capability is relational and condition specific: it describes what a person can functionally accomplish under specified conditions, including the use of available technological assistance. Competence is human and retained: it describes what a person can reliably evaluate or perform independent of a particular assistance configuration.

This paper treats capability as the broader term, encompassing both assisted and unassisted functional reach. The three thresholds specified below describe different forms of capability, with different implications for what has been learned, retained, or made portable.

This concept overlaps with established scholarship on learning curves, augmentation, human capital acquisition, barriers to entry, cognitive offloading, deskilling, and technological substitution. The claim of this paper is consequently modest. Technology has always altered what people can do and how difficult it is to learn. Generative AI does not originate that phenomenon.

The proposed contribution is instead an analytical separation.

Existing research commonly studies how technology improves task performance, redistributes tasks, alters skill requirements, shifts labor demand, or changes learning. Capability compression asks a narrower question:

How has technology changed the resource distance between this person and this performance threshold?

That question becomes particularly important when assistance produces competent looking outputs without equivalent increases in the user's ability to evaluate or independently reproduce them.

The governing distinction of the paper is therefore:

Performance compression is not equal to learning compression, and learning compression is not equal to competence compression.

Generative AI can change one without changing the others by the same amount.

This matters because debates about AI frequently move from one category to another without noticing the transition. A finding that novices complete work more quickly becomes a claim that novices have acquired expertise. A decline in the scarcity of an output becomes a claim that the incumbent's expertise has disappeared. Economic displacement becomes evidence of illegitimate participation. Successful assisted performance becomes evidence of learning.

The capability compression framework is designed to prevent those jumps.

A brief note on terminology: this paper uses "negotiable" to describe the distinctive interactivity of generative AI: users can iteratively refine, rephrase, constrain, and redirect the system toward a threshold through low cost natural language interaction, without needing to master a specialized formal interface. This does not imply that no prior technology was interactive. It suggests that the breadth, speed, and generality of interactive distance shortening may be historically unusual.

Technology has long altered the distance between intention and execution. The table below illustrates this pattern with several historical examples:

Technology Compressed Did Not Necessarily Compress

Calculator Arithmetic execution Mathematical judgment, estimation, error checking

GPS Route acquisition Spatial orientation, environmental judgment, navigation without signal

Spreadsheet Calculation and scenario manipulation Model design, assumptions, auditing, interpretation

CAD Drawing and visualization Engineering judgment, material behavior, safety validation

The point is not that generative AI is "just like" these technologies. The point is that capability compression predates generative AI, and historical technologies demonstrate that performance expansion, learning, deskilling, market disruption, and competence erosion can vary independently.

---

2. The Academic Neighborhood

Capability compression does not begin from a blank literature.

The task based tradition in labor economics already separates workers' skills from the tasks to which those skills are applied and examines how technologies change their allocation. Autor, Levy, and Murnane's (2003) analysis of computerization demonstrated that technology may substitute for some routine activities while complementing others. Acemoglu and Autor (2011) subsequently developed a more general task framework for understanding technological change and labor demand.

These approaches primarily ask how technologies alter the distribution of productive tasks and the demand for different forms of labor.

Deskilling research asks another important question: what happens to knowledge, discretion, control, and exercised competence when work is reorganized technologically? Braverman's (1974) classic labor process analysis emphasized the separation of conception from execution and the relocation of knowledge and control away from workers. Subsequent automation research has shown that technological assistance can also create complacency, reduce monitoring, and weaken readiness to intervene when systems fail (Parasuraman and Manzey, 2010).

Augmentation research takes yet another angle. It asks what workers can accomplish when technology assists them.

Recent generative AI evidence makes this last phenomenon particularly visible. Brynjolfsson, Li, and Raymond (2025) find that a generative AI conversational assistant increased productivity by approximately 14 percent among more than 5,000 customer support agents, with substantially larger gains among novice and lower skilled workers. Their evidence suggests that the system can transmit aspects of high performing workers' practices to less experienced employees.

Noy and Zhang (2023) likewise find that generative AI can reduce completion time and increase evaluated quality on professional writing tasks, with larger improvements among initially weaker performers.

Such findings establish important performance effects.

They do not by themselves establish equivalent changes in independent human competence.

That distinction places capability compression adjacent to augmentation rather than inside it.

Augmentation asks what performance improves while the tool is present.

Capability compression asks how much the resource distance to a specified threshold has changed, and which threshold has actually been crossed.

The distinction becomes especially important once multiple thresholds are recognized.

---

3. Threshold Distance and Functional Thresholds

3.1 Threshold Distance

This paper defines threshold distance as the multidimensional resources required for a specified individual to move from their current capability state to a specified functional performance threshold under specified conditions.

The concept is intentionally person specific.

Two individuals facing the same target task may have the same broad disciplinary background but radically different threshold distances because they differ in adjacent knowledge, technological fluency, access to tools, time, feedback, risk tolerance, physical ability, or institutional support.

Threshold distance is also task specific.

A person may have a short threshold distance to producing a simple visualization but a very long one to validating a statistical model.

Finally, threshold distance is threshold specific.

This is crucial because "can do the task" conceals different kinds of capability.

For analytical clarity, the framework distinguishes three threshold specific forms of distance:

· Assisted execution distance: Resources required for a person using a specified technological configuration to produce acceptable output.

· Supervisory acquisition distance: Resources required for the person to acquire reliable evaluation, diagnosis, constraint, correction, calibration, and escalation capability.

· Independent acquisition distance: Resources required for the person to acquire robust performance when technological assistance is unavailable, degraded, or misleading.

3.2 Three Thresholds

The framework distinguishes three functional thresholds. These are not necessarily a developmental ladder; they are different kinds of functional capability with different empirical requirements and different implications for learning, accountability, and system design.

Assisted Execution Threshold

The user can produce an outcome that satisfies a contextually appropriate minimum standard while technological assistance is available.

The relevant observable is successful assisted performance.

This threshold says nothing by itself about whether the individual understands the underlying process.

Supervisory Threshold

The user can sufficiently evaluate, diagnose, constrain, and correct the technological system's output for the relevant context.

This includes the ability to:

· recognize incorrect suggestions;

· identify significant omissions;

· distinguish plausible output from valid output;

· constrain the system appropriately;

· know when escalation to deeper expertise is necessary;

· maintain acceptable quality control.

Supervisory competence does not require complete independent production.

It requires sufficient understanding to judge the machine.

Crucially, supervisory competence includes competent uncertainty: the ability not only to evaluate and correct output, but to recognize when one cannot reliably determine whether that output is trustworthy and escalation is required.

The supervisory threshold is domain relative. In some domains, such as simple design, routine documentation, and low stakes prototyping, it may be substantially less demanding than independent competence. In others, particularly safety critical systems, adversarial environments, high stakes judgment, and deeply contextual domains, reliable supervision may require expertise approaching independent professional competence. Recognizing when one cannot reliably determine whether output is trustworthy (competent uncertainty) is therefore not a lower bar; it is a qualitatively different kind of competence, one that may be especially important in precisely those domains where the cost of error is highest.

Independent Competence Threshold

The individual can produce, adapt, diagnose, or repair the relevant outcome when technological assistance is absent, degraded, misleading, or wrong.

This threshold is closest to conventional accounts of acquired human competence.

The thresholds are analytically distinct and may overlap, diverge, or be reached in different sequences. A professional may possess independent competence before ever using AI. A novice may reach assisted execution without reaching supervisory competence. Another user may develop sufficient supervisory competence without being capable of reproducing the full output independently. This middle state is central to AI mediated work.

---

4. The Supervisory Threshold

The supervisory threshold may be the most consequential distinction introduced by the framework.

Conventional discussions often contrast two states:

the human can perform the task

versus

the machine performs the task.

AI mediated work introduces a third:

the machine performs substantial portions of the task, but the human remains responsible for knowing whether the result should be trusted.

That distinction has empirical support.

Research on novice programmers using generative AI found that successful task completion could coexist with weak metacognitive monitoring. Prather et al. (2024) conducted 21 lab sessions using observation, interviews, and eye tracking. Twenty of 21 participants completed the programming problem, but the successful looking outputs concealed an important divide: more capable participants used suggestions selectively and rejected bad suggestions; struggling participants often retained earlier metacognitive problems, accepted poor suggestions, overestimated their performance, and finished with what the authors describe as an "illusion of competence."

The point is not that the weaker participants accomplished nothing.

They did.

They crossed an assisted execution threshold.

The important observation is that this did not necessarily mean they could reliably determine when the system was wrong.

Thus:

Successful assisted execution is not sufficient evidence of supervisory competence.

This distinction becomes increasingly important as stakes rise.

A user creating a disposable visual prototype may reasonably tolerate substantial uncertainty.

A user deploying software affecting financial transactions cannot.

A clinician, engineer, attorney, pilot, or security professional may need substantially greater supervisory capacity because the cost of plausible but incorrect output is high.

The framework therefore embeds accountability and error tolerance inside the definition of capability compression.

The same apparent output can represent radically different levels of meaningful capability depending on whether failure is detectable, reversible, and consequential.

The supervisory threshold also has practical implications for training, accountability, and system design. Training may increasingly need to emphasize evaluation and diagnosis rather than independent production. Accountability regimes will need to clarify who bears responsibility for output that a human supervisor could not reasonably have detected as erroneous. System design may need to prioritize explainability and uncertainty signaling over raw generation capability.

However, the framework does not assume that assignment of supervisory responsibility implies possession of supervisory competence. In some AI mediated work arrangements, workers may increasingly be assigned directing, inspection, diagnosis, constraint, correction, and escalation functions. Whether this constitutes enriched supervisory competence, nominal oversight, or responsibility without effective control is an empirical and institutional question. The phrase "responsibility without effective control" captures the risk that workers may be held accountable for outputs they lack the authority, information, time, knowledge, or institutional power to meaningfully supervise.

---

5. Assisted Performance and Independent Competence

Cruces et al. (2026) provide unusually useful evidence for distinguishing assisted performance from persistent capability.

Their randomized study involved 1,174 adults performing workplace style business problem solving tasks with or without a generative AI assistant, followed by a module without AI. In the unassisted control condition, the education based performance gap was 0.548 standard deviations. Under AI assistance, it fell to 0.139 standard deviations, closing roughly three quarters of the initial difference.

If analysis stopped there, one might conclude that AI largely erased the underlying productivity difference.

The follow up complicates that interpretation.

Once AI was unavailable, a substantial education gap reappeared. Yet the assisted group did not simply collapse back to baseline. Lower education participants retained part of their gain, particularly where AI use had been accompanied by sustained human effort.

This produces exactly the kind of distinction capability compression is intended to describe.

AI may create:

1. substantial assisted performance compression;

2. some learning compression;

3. less complete independent competence compression.

These effects need not be identical.

This also prevents an overly pessimistic interpretation.

AI assistance is not necessarily mere delegation.

Under some conditions it may scaffold learning.

The empirical question is therefore not simply whether AI "teaches" or "does the work."

It is:

Under what conditions does compression of assisted performance transfer into supervisory and independent competence?

That question creates a tractable research program.

---

6. Capability Compression and Competence Erosion

Earlier formulations of this framework contrasted capability compression directly with deskilling.

That distinction remains useful, but "deskilling" is too broad to serve cleanly as a single measurement axis.

Deskilling can refer to:

· loss of retained ability;

· reduced exercise of ability;

· reduced skill requirements within a job;

· reduced discretion;

· transfer of control to management or machines;

· occupational restructuring;

· declining market rewards.

These phenomena may interact, but they are not interchangeable.

For the core model, this paper therefore uses a narrower comparison:

assisted performance compression

versus

erosion of independent competence.

They are analytically orthogonal even if they are causally related over time.

 Low Erosion of Independent Competence High Erosion of Independent Competence

Low Assisted Performance Compression Traditional practice / limited tool effect Automation dependency without substantial access expansion

High Assisted Performance Compression Assisted capability expansion Performance outsourcing with competence fragility

The theoretically interesting condition is high assisted performance compression with low erosion of independent competence. This is the distinctive case in which technological assistance expands what a person can functionally reach without necessarily diminishing what they could already do independently.

However, the word "immediate" matters.

A technology that initially expands assisted capability may later reduce practice, weaken diagnostic readiness, or change occupational training structures. Capability compression and competence erosion are separable dimensions, not temporally independent processes.

A proper empirical analysis should therefore specify time horizon.

---

7. Knowledge Distance and the Limits of Compressibility

Capability compression is not unlimited.

Vendraminelli et al. (2025) provide an important boundary condition through their concept of knowledge distance.

Their study distinguishes occupational insiders, adjacent outsiders, and distant outsiders asked to perform tasks associated with the insider occupation. The authors find a "GenAI wall": generative AI is more successful at reducing expertise gaps for nearer occupations than for distant occupations, and its effects differ between conceptualization and execution tasks.

Knowledge distance and threshold distance should therefore be kept separate.

Knowledge distance concerns the mismatch between a person's prior knowledge configuration and the knowledge required by the target task.

Threshold distance concerns the resources required to move sufficiently far across that gap to reach a specified threshold.

Knowledge distance is consequently one potential antecedent of compressibility.

Two individuals with similar knowledge distance may nevertheless have different threshold distances because they differ in:

· access to AI systems;

· digital fluency;

· feedback quality;

· available time;

· financial security;

· organizational support;

· regulatory constraints;

· ability to detect errors;

· physical execution requirements;

· availability of knowledgeable collaborators.

The amount of achievable compression therefore depends on both the target and the traveler.

---

8. The Jagged Frontier

Even within a domain, apparently similar tasks may differ sharply in compressibility.

Dell'Acqua et al. (2026) demonstrate this through the "jagged technological frontier." Their field experiment with consultants found strong benefits from GPT-4 on tasks inside the system's effective frontier. On a complex task outside that frontier, however, control participants achieved 84.5 percent correctness while AI assisted groups achieved approximately 60 percent and 70.6 percent.

This is not merely a limitation of current models.

Conceptually, it demonstrates why assisted output should not be treated as a smooth proxy for capability.

A tool can:

· reduce execution effort;

· increase confidence;

· generate convincing output;

while simultaneously:

· reduce correctness;

· obscure failure;

· make supervision more difficult.

Thus compression may be jagged.

A useful future theory of compressibility will likely need to incorporate:

· codifiability;

· tacit knowledge;

· embodiment;

· reversibility of errors;

· availability of objective feedback;

· task novelty;

· adversariality;

· stakes;

· contextual dependence;

· model reliability;

· supervisory competence.

The framework deliberately does not propose a single "compression coefficient" at this stage.

The evidence is too discontinuous.

---

9. Compression, Expansion, and Conditional Recombination

Capability compression does not necessarily end with substitution.

When a previously inaccessible task becomes reachable, the individual's practical task set may expand.

This produces a conditional dynamic:

Compression leading to possible Expansion leading to possible Recombination

Compression reduces threshold distance.

Expansion occurs when the user becomes capable of undertaking tasks outside their previous practical range.

Recombination may occur when those newly reachable capabilities interact with what the person already knows.

The recombination mechanism itself is not new.

Kogut and Zander (1992) argue that organizations develop new knowledge and skills by recombining existing capabilities, while emphasizing that knowledge is often tacit, socially embedded, and path dependent. Weitzman (1998) likewise treats new ideas as combinations of existing ideas.

Capability compression therefore does not contribute the proposition that recombination exists.

The narrower hypothesis is:

Technological compression may alter which capability components become practically available for recombination, to whom, and at what acquisition cost.

An illustrator gaining functional programming access does not necessarily become a software engineer.

But the illustrator may become able to produce interactive visual work that previously required a collaboration, additional employee, or entirely separate education.

Likewise, a historian using AI mediated statistical tools need not become a statistician to ask quantitatively sophisticated historical questions that were previously outside their practical reach.

The interesting comparison is therefore not always:

newcomer versus incumbent specialist.

It may be:

specialist before compression versus specialist after access to an adjacent capability.

This leads to a consequential unanswered question:

Does generative AI broaden the population capable of productive cross domain recombination, or does it primarily amplify individuals who already possess the evaluative competence, autonomy, resources, and networks necessary to exploit it?

The present paper proposes that question. It does not claim to have answered it.

---

10. Scarcity, Market Value, and Legitimacy

Capability compression can also alter economic scarcity.

If many more people can cross an assisted execution threshold, the market supply of certain classes of acceptable output may increase.

That does not imply that established expertise has vanished.

It implies that the relationship between expertise and scarcity may have changed.

This distinction can be expressed simply:

A skill can remain real while its scarcity premium changes.

Economic consequences should nevertheless be modeled downstream rather than incorporated into the definition of compression.

Lower threshold distance does not necessarily imply lower wages.

Outcomes depend on:

· demand expansion;

· complementarity;

· reputation;

· quality differentiation;

· regulation;

· credentialing;

· liability;

· ownership;

· bargaining power;

· market structure.

Nor do market consequences settle normative questions.

An incumbent may correctly identify a loss of income without thereby establishing that a technologically assisted entrant is illegitimate.

Questions about copyright, training data consent, authorship, creative status, accountability, or professional standards require separate arguments.

Capability compression is intended to clarify that separation, not resolve every controversy surrounding generative AI.

---

11. Adaptive Possibility and Unequal Opportunity

The framework also produces a tempting but invalid inference.

If AI reduces threshold distance, one may conclude that workers threatened by technological change should simply use the same technology to acquire something else.

This paper rejects that move.

Technical possibility does not create an individual obligation to adapt.

Nor does it imply equal practical opportunity.

The ILO and World Bank (Gmyrek et al., 2024) estimate that as many as 17 million jobs in Latin America and the Caribbean that could potentially receive productivity enhancing benefits from generative AI are constrained by gaps in digital access and infrastructure. Benefits are also more likely to reach formal, urban, better educated, and higher income workers.

OECD research similarly identifies skills, training, data maturity, organizational capabilities, and institutional support as important constraints on AI adoption (OECD/BCG/INSEAD, 2025). OECD's 2026 skills review reports that lack of skills remains a major adoption barrier and emphasizes shared responsibility for training among workers, firms, and governments (OECD, 2026).

Therefore:

From the fact that technology makes a transition technically possible, it does not follow that individuals possess equal opportunity to undertake it, that they are obligated to do so, or that firms and governments are absolved of responsibility for transition costs.

Capability compression raises a distributional question.

It does not answer it.

---

12. A Literary Lens: Flowers for Algernon

The relational consequences of rapidly changing capability are difficult to capture in economic terminology alone.

Daniel Keyes's (1966) Flowers for Algernon provides a useful conceptual lens.

The analogy is not Charlie Gordon's disability.

The relevant structure is a rapid change in the distribution of capability within an established social hierarchy.

Charlie does not merely acquire new capacities. Everyone around him must reconsider who he is relative to them.

Relationships that appeared to rest on affection, expertise, seniority, or normal social roles are revealed to have depended partly on an assumed capability distance.

That insight maps onto capability compression without requiring Charlie's transformation to stand for AI use literally.

The useful recognition is:

A change in another person's capability does not automatically constitute a reduction in one's own competence.

Yet the economic qualification remains:

A capability can remain intact while the scarcity or market value associated with it changes.

This is why the incumbent should not be caricatured.

The worker threatened by capability compression may be experiencing a genuine material disruption.

The symmetry of the framework nevertheless matters.

A reduction in threshold distance need not be available only to newcomers entering an incumbent's field.

The incumbent may also encounter newly negotiable distances to adjacent tasks.

The same mechanism that permits one actor to cross an old boundary may create new routes for others.

But those routes are opportunities, not obligations.

The literary lens therefore illuminates the social dimension of the theory rather than serving as evidence for it. It demonstrates that crossing a threshold can alter social relationships even when it does not erase differences in deeper competence.

---

13. Research Agenda

The framework generates several testable questions.

13.1 Threshold Divergence

How frequently do users cross assisted execution thresholds without crossing supervisory thresholds? Which domains produce the largest gaps?

13.2 Supervisory Competence

What predicts the ability to detect, reject, and repair erroneous AI output? Does supervisory competence require substantial underlying domain knowledge, or can it itself be compressed? Does competent uncertainty correlate with domain expertise or with metacognitive skill? Under what conditions does supervisory responsibility become responsibility without effective control?

13.3 Learning Transfer

Under what conditions does repeated AI assisted performance produce independent learning? Cruces et al. (2026) suggest that effort while using AI may matter for post assistance retention.

13.4 Knowledge Distance

How does prior knowledge distance interact with task structure to determine achievable compression?

13.5 Competence Erosion

Does long term AI assistance reduce independent competence through disuse, or can assisted practice strengthen learning? Under what conditions does each occur?

13.6 Recombination

Does compressed access to adjacent domains actually broaden the population producing valuable hybrid work? Or does it disproportionately advantage individuals who already possess high agency, strong networks, and evaluative expertise?

13.7 Market Consequences

When does assisted performance compression reduce scarcity premiums? When does increased demand or quality differentiation offset that effect?

13.8 Operationalization

Future studies might measure threshold distance profiles using:

· time to threshold;

· direct cost;

· training hours;

· feedback required;

· iteration burden;

· error detection performance;

· unaided follow up;

· transfer performance;

· expert supervision required;

· knowledge distance measures.

The immediate goal should not be a universal scalar index. It should be reliable task specific profiles.

---

14. Limitations

Several limitations should constrain the interpretation of the framework.

First, "functional performance" is context dependent. A threshold suitable for informal prototyping may be unacceptable in a safety critical domain.

Second, assisted execution is not equivalent to professional authorization. Capability compression cannot override licensing, accountability, legal, or ethical standards.

Third, the framework is currently conceptual. Although several empirical studies exhibit patterns consistent with its distinctions, threshold distance has not yet been independently validated as a measurement construct.

Fourth, the three thresholds may not capture every form of capability. Collaborative, organizational, and socially distributed competence may require additional levels of analysis.

Fifth, recombination remains a hypothesis about potential consequences rather than an established general effect of generative AI.

Sixth, capability compression occurs at the person task level, whereas occupational deskilling, labor market restructuring, and institutional change occur at other levels. These levels should not be conflated.

Seventh, the framework distinguishes capability from competence but does not resolve all questions about where human knowledge, organizational knowledge, and system capacity begin and end. The relational ontology of capability is proposed as a conceptual starting point, not a complete theory of distributed cognition.

Finally, generative AI should not be portrayed as historically unprecedented simply because it reduces acquisition barriers. Technologies have altered learning and execution costs throughout history.

The comparative claim is narrower:

Generative AI may make capability distance unusually negotiable because assistance is interactive, iterative, individualized, and applicable across a comparatively broad set of symbolic tasks.

---

15. Conclusion

The central problem in current discussions of generative AI and skill may not be disagreement about whether technology changes capability.

It may be disagreement about which capability has changed.

A user who can produce an acceptable result with AI has accomplished something real.

That achievement does not necessarily establish that the user can evaluate the result.

The ability to evaluate it does not necessarily establish the ability to reproduce it independently.

And independent competence does not imply that markets will continue to reward that competence as they once did.

These distinctions can coexist.

Capability compression provides a vocabulary for describing them.

It treats technological change as a change in threshold distance: the resources separating a particular person from a particular performance threshold under particular conditions.

It distinguishes this from erosion of previously possessed independent competence.

It distinguishes assisted performance from supervisory competence and independent competence.

It treats knowledge distance and the jagged technological frontier as boundary conditions.

And it proposes that compressed access may sometimes expand the user's reachable task set sufficiently to support new forms of recombination with existing expertise, without claiming that such recombination is automatic or universally beneficial.

The framework therefore rejects two symmetrical exaggerations.

Generative AI does not make everyone an expert.

Nor does technologically mediated performance become analytically meaningless simply because some of the competence resides in the system.

The more useful question is:

What threshold has become reachable, for whom, at what cost, under what conditions, and with what capability to recognize failure?

Technology has always altered the distance between intention and execution.

Generative AI may make that distance unusually negotiable.

The academic task is to determine exactly which distances are shrinking, which are not, and what happens when people begin crossing them.

---

References

Acemoglu, D., and Autor, D. (2011). Skills, tasks and technologies: Implications for employment and earnings. In Handbook of Labor Economics (Vol. 4, pp. 1043–1171). Elsevier. https://doi.org/10.1016/S0169-7218(11)02410-5

Autor, D. H., Levy, F., and Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. The Quarterly Journal of Economics, 118(4), 1279–1333. https://doi.org/10.1162/003355303322552801

Braverman, H. (1974). Labor and Monopoly Capital: The Degradation of Work in the Twentieth Century. Monthly Review Press.

Brynjolfsson, E., Li, D., and Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae045

Cruces, G., Fernández Meijide, D., Galiani, S., Gálvez, R. H., and Lombardi, M. (2026). Does Generative AI Narrow Education-Based Productivity Gaps? Evidence from a Randomized Experiment. NBER Working Paper No. 34851. https://doi.org/10.3386/w34851

Dell'Acqua, F., McFowland, E. III, Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (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. https://doi.org/10.1287/orsc.2025.21838

Gmyrek, P., Winkler, H., and Garganta, S. (2024). Buffer or Bottleneck? Employment Exposure to Generative AI and the Digital Divide in Latin America. ILO Working Paper 121 / World Bank. https://doi.org/10.54394/TFZY7681

Keyes, D. (1966). Flowers for Algernon. Harcourt, Brace and World.

Kogut, B., and Zander, U. (1992). Knowledge of the firm, combinative capabilities, and the replication of technology. Organization Science, 3(3), 383–397. https://doi.org/10.1287/orsc.3.3.383

Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586

OECD/BCG/INSEAD. (2025). The Adoption of Artificial Intelligence in Firms: New Evidence for Policymaking. OECD Publishing. https://doi.org/10.1787/f9ef33c3-en

OECD. (2026). AI and Skills: What We Know So Far. OECD Publishing. https://doi.org/10.1787/f843b352-en

Parasuraman, R., and Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055

Prather, J., Reeves, B. N., Leinonen, J., MacNeil, S., Randrianasolo, A. S., Becker, B. A., Kimmel, B., Wright, J., and Briggs, B. (2024). The widening gap: The benefits and harms of generative AI for novice programmers. In Proceedings of the ACM Conference on International Computing Education Research (ICER '24), Vol. 1, 469–486. https://doi.org/10.1145/3632620.3671116

Vendraminelli, L., DosSantos DiSorbo, M., Hildebrandt, A., McFowland, E. III, Karunakaran, A., and Bojinov, I. (2025). The GenAI Wall Effect: Examining the Limits to Horizontal Expertise Transfer Between Occupational Insiders and Outsiders. Harvard Business School Working Paper No. 26-011. https://doi.org/10.2139/ssrn.5462694

Weitzman, M. L. (1998). Recombinant growth. The Quarterly Journal of Economics, 113(2), 331–360. https://doi.org/10.1162/003355398555595

---

Appendix A: The Supervisory Threshold in Practice

The supervisory threshold is the framework's most distinctive contribution. This appendix provides additional detail on its practical implications.

A.1 What Supervisory Competence Entails

Supervisory competence is not a single skill. It is a cluster of capacities that may be partially separable:

1. Recognition: The ability to identify when output is incorrect, inappropriate, or misleading.

2. Diagnosis: The ability to understand why an error occurred and what kind of correction is needed.

3. Constraint: The ability to set boundaries on the system's operation to prevent certain classes of error.

4. Correction: The ability to repair or improve output, either directly or through refined prompting.

5. Escalation: The ability to recognize when the problem exceeds one's own competence and seek appropriate assistance.

6. Uncertainty calibration: The ability to distinguish between confidence that is warranted and confidence that is not.

The last capacity, uncertainty calibration, is especially important. It is what the paper refers to as "competent uncertainty."

A.2 Why the Supervisory Threshold Matters for Training

Traditional training often assumes that competence means independent production. The supervisory threshold suggests an alternative: workers may need to become competent evaluators and correctors even if they never become fully independent producers.

This has implications for curriculum design:

· Emphasis on critical evaluation of AI generated output.

· Training in error identification and diagnosis.

· Practice with constraint setting and prompt refinement.

· Development of metacognitive skills for recognizing the limits of one's own judgment.

A.3 Why the Supervisory Threshold Matters for Accountability

If a worker supervises AI output and fails to catch an error, who bears responsibility?

The supervisory threshold makes this question salient. It suggests that:

· Accountability should reflect the level of supervisory competence that was reasonably expected.

· Systems should be designed to support supervisory competence, for example by surfacing uncertainty and explaining reasoning.

· Liability frameworks may need to distinguish between cases where a supervisor should have caught an error and cases where the error was not reasonably detectable.

A.4 Why the Supervisory Threshold Matters for System Design

Systems that merely generate output may be insufficient for high stakes contexts. Systems that support supervision, such as by providing explanations, uncertainty estimates, error flags, and interactive correction, may enable users to cross the supervisory threshold even when they cannot cross the independent competence threshold.

Design priorities might include:

· Explanatory interfaces that reveal why the system produced a particular output.

· Uncertainty indicators that signal when output is less reliable.

· Error detection tools that highlight potential problems.

· Interactive refinement workflows that allow users to correct and constrain output.

---

Appendix B: Operationalizing Threshold Distance

Threshold distance is proposed as a multidimensional latent construct. This appendix outlines potential approaches to operationalization.

B.1 Observable Components

Component Operationalization

Time to threshold Hours, days, or weeks from first attempt to consistent acceptable performance

Direct monetary cost Training costs, tool costs, lost income during learning

Training hours Hours of formal instruction, practice, or feedback

Procedural knowledge required Number of steps, complexity, codifiability

Feedback required Number of corrective interventions needed

Error burden Frequency and severity of errors during threshold attainment

Cognitive load Mental effort required, such as self reported or physiological measures

Equipment access Hardware and software requirements and availability

Motor skill demand Physical dexterity required

Risk exposure Cost of mistakes during learning

B.2 Measurement Approaches

1. Self report: Surveys asking individuals to estimate the resources required to reach competence.

2. Performance tracking: Logs of time, errors, and assistance used during learning.

3. Experiment: Randomized interventions that vary assistance, feedback, or instructional support.

4. Expert assessment: Subject matter experts evaluate acquisition paths for different tasks and individuals.

B.3 A Tentative Assessment Framework

For a given task, domain, and assistance condition, threshold distance might be assessed as:

Threshold Distance = f(Time, Cost, Effort, Feedback, Risk, Access)

Where each dimension is measured relative to a reference condition, such as traditional training without AI.

This is not proposed as a precise formula. It is proposed as a way of organizing empirical investigation.

---

Appendix C: Glossary of Key Terms

Term Definition

Assisted execution distance Resources required for a person using a specified technological configuration to produce acceptable output.

Assisted execution threshold The point at which a user can produce an acceptable outcome with technological assistance.

Capability Relational and condition specific: what a person can functionally accomplish under specified conditions, including the use of available technological assistance.

Capability compression A technology associated reduction in the resources a specified individual requires to reach a specified threshold of functional performance on a specified task, under specified conditions of assistance, accountability, and error tolerance.

Competence Human and retained: what a person can reliably evaluate or perform independent of a particular assistance configuration.

Competent uncertainty The ability to recognize when one cannot reliably determine whether an output is trustworthy and escalation is required. A component of supervisory competence.

Deskilling The erosion, bypassing, fragmentation, or atrophy of competence previously possessed, exercised, or required.

Erosion of independent competence A reduction in the ability to perform a task without technological assistance. The paper's operational Y axis, narrower than "deskilling."

Functional competence The ability to produce an outcome meeting a minimum threshold of acceptability within a specified context.

Independent acquisition distance Resources required for the person to acquire robust performance when technological assistance is unavailable, degraded, or misleading.

Independent competence threshold The point at which an individual can produce, adapt, or repair an outcome when technological assistance is absent, degraded, or incorrect.

Jagged technological frontier The phenomenon, identified by Dell'Acqua et al. (2026), in which AI improves performance on some tasks while degrading it on apparently similar tasks outside the system's effective frontier.

Knowledge distance The separation between a person's existing expertise and the knowledge requirements of a target task. An antecedent of compressibility.

Negotiable distance The claim that generative AI makes capability distance interactive, iterative, and adjustable rather than fixed.

Recombination The integration of newly accessible capabilities with existing expertise, judgment, relationships, and domain knowledge.

Responsibility without effective control A condition in which workers are held accountable for outputs they lack the authority, information, time, knowledge, or institutional power to meaningfully supervise.

Supervisory acquisition distance Resources required for the person to acquire reliable evaluation, diagnosis, constraint, correction, calibration, and escalation capability.

Supervisory threshold The point at which a user can sufficiently evaluate, diagnose, constrain, and correct a technological system's output for a relevant context.

Threshold distance The umbrella concept: resources separating a specified person from a specified functional threshold under specified conditions.

---

Acknowledgements

[To be added based on venue and co-author information]

---

Declaration of Conflicting Interests

The author(s) declare no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

---

Paper first. Bakery later.