When Better Work Reveals Less

1. Better Work, Worse Evidence

A three-month randomized field experiment in patent law presents a problem that ordinary productivity measures cannot resolve. Autor and colleagues gave 133 practicing patent lawyers across eleven U.S. intellectual-property firms access to a custom AI drafting assistant. AI access improved benchmark drafting quality after both ten and ninety days, and junior lawyers experienced particularly large assisted gains. At the end of the intervention, however, participants completed an unaided patent-redlining task requiring professional judgment. Treated lawyers performed better overall, but the improvement was concentrated among senior lawyers; junior lawyers showed no average advantage despite their larger gains while working with AI (Autor et al., 2026).

The result does not show that AI harmed junior lawyers, nor that unaided redlining exhausts what professional expertise means in an AI-mediated workplace. It establishes something narrower and more consequential: improvement in the work produced by a human–AI system need not identify a proportional improvement in every human capability embedded within that system.

That distinction matters because work performs two organizational functions. It produces value, and it produces information about workers.

A diagnostic interpretation, financial analysis, patent draft, engineering recommendation, software review, or client presentation is valuable because of what it accomplishes. It is also evidence. Managers use repeated performances to infer who understands difficult problems, who can recognize an exception, who can be trusted with less supervision, who merits scarce mentoring, and who is ready for more demanding work. No single artifact carries all of that information. Yet careers inside firms depend on repeated attempts to learn what workers can do from evidence generated while those workers perform.

The inference has never been clean. Performance reflects task difficulty, teams, tools, managerial support, luck, and prior assignment as well as individual capability. Organizations nevertheless have to make consequential allocations before uncertainty disappears. They cannot wait to know a worker perfectly before deciding who will lead a client matter, receive an exception case, enter a rotation, or supervise others.

Generative AI changes the process that generates this evidence.

Let J denote independently exercisable judgment: the human capability relevant to a particular organizational decision when the focal AI contribution is unavailable, unreliable, inappropriate, or deliberately held constant. Let H denote human–AI orchestration capability: the ability to elicit, evaluate, constrain, integrate, and appropriately rely on machine output. Let A denote the contribution of AI itself. The representation is schematic rather than ontological. In practice these components interact. Domain expertise can improve orchestration; orchestration can reveal judgment; and AI can change which aspects of human capability become valuable. The point is that equally strong observed outputs can arise from materially different human–machine configurations.

Organizations often have no reason to separate those configurations. If a role is durably hybrid, the ability to produce excellent outcomes with AI may be exactly the capability the organization should reward. Human–AI orchestration is not counterfeit competence.

Separation becomes consequential only when an organizational decision depends on a human capability that must sometimes survive independently of the focal AI contribution. A radiologist may need to recognize an anomaly when automated support fails. A lawyer supervising AI-mediated work may need to detect a plausible but invalid argument. An engineer may need to diagnose a novel condition outside a model’s reliable range. A manager deciding who should receive exception authority may need information that a polished hybrid artifact no longer supplies directly.

Recent work already establishes several adjacent claims: AI can remove developmental work, weaken the relation between observable output and human capability, alter attribution and status dynamics, and under some conditions support expertise through intensive review (Sarala, 2026; Ayeni et al., 2026; Williams, 2026; Gallacher, 2026; Call et al., 2026; Petropoulos, 2026). The present theory does not claim novelty for those mechanisms. Its question begins one organizational step later: what happens when altered capability evidence enters an internal allocation system whose decisions determine both future development and future evidence?

The unresolved organizational mechanism begins after the evidence changes. Firms do not merely hold beliefs about workers. They act on those beliefs by allocating jobs, clients, rotations, mentoring, autonomy, difficult cases, and responsibility. Those assignments can change what workers subsequently learn. They also change what workers subsequently have the opportunity to demonstrate. An inferential problem can therefore become a developmental problem even if AI does not eliminate a single developmental task from the organization.

This is the distinction between two pathways. In the direct pathway emphasized by apprenticeship theories, AI changes the tasks a worker performs, thereby changing what the worker practices and learns. In the indirect pathway developed here, AI changes what the organization can infer about a worker, which changes who receives developmental opportunities, which then changes later capability.

The indirect pathway becomes recursive because assignments do two things simultaneously. They can form capability, and they can reveal capability. A worker entrusted with a difficult project encounters a developmental experience. The same project can reveal whether that worker recognizes hidden dependencies, handles ambiguity, escalates appropriately, recovers from error, and transfers knowledge to unfamiliar conditions. A worker denied the project loses both opportunities.

The theory’s distinctive claim is not that AI can obscure human capability, nor that developmental assignments build capability. It is that these mechanisms become recursively linked inside organizations. Capability beliefs shape developmental allocation; developmental allocation changes both the capability being estimated and the evidence subsequently available for estimating it. Consequently, an initially inaccurate capability inference can alter the conditions under which that inference could later be corrected. In the false-negative case, a worker can lose both the experience through which capability would develop and the experience through which existing capability would become visible.

How does generative AI alter the observation technology through which organizations learn human capability, and what happens when those beliefs govern assignments that both develop and reveal future capability?

This recursive mechanism yields two theoretical extensions. First, it extends employer-learning and internal-labor-market theory by making the information content of performance endogenous to production architecture. Second, it extends work-design theory by showing that jobs are also evidence-generating architectures: assignment changes not only what workers do and learn but what organizations can validly infer about them. Formation–revelation feedback connects the two extensions by showing how allocation changes both future capability and future evidence.

2. Internal Labor Markets as Learning Systems

The mechanism begins before AI. Organizations allocate people under incomplete information. Workers differ in capability, jobs differ in requirements, and many relevant capabilities become visible only through experience.

Employer-learning theory provides the first anchor. Farber and Gibbons develop a dynamic model in which employers learn about worker ability as performance observations accumulate, progressively changing the role of prior signals in labor-market inference. Their model concerns wage dynamics rather than internal developmental allocation, but its central premise is foundational here: observable work outcomes are information about latent worker ability, and the value of earlier signals changes as new evidence arrives (Farber & Gibbons, 1996).

Gibbons and Waldman deepen that logic by integrating learning about ability with job assignment and on-the-job human-capital acquisition. Careers emerge from their interaction. Workers are assigned partly because of what firms believe about them; assignment affects productivity and skill accumulation; later performance supplies new evidence. A worker’s career is therefore not a sequence in which firms first discover fixed ability and only then allocate work. Learning, assignment, and capability accumulation proceed together (Gibbons & Waldman, 1999).

This distinction matters because a static signal account would be insufficient for the present theory. If capability were fixed and assignments had no developmental consequence, weaker capability evidence would create sorting errors but not a mechanism of expertise reproduction. If assignments changed capability but generated no additional information, the system would remain dynamic but an initial inference error would not necessarily change the organization’s future ability to correct itself. Formation–revelation feedback requires both properties: jobs change workers and jobs generate observations through which firms revise beliefs about workers.

Organizational updating is also imperfect. Black and Vance show that managers’ initial assessments of employee ability continue to predict promotion decisions for years even as substantial performance evidence accumulates. Managers reduce the weight placed on those initial assessments as new signals arrive, but more slowly than the relative informativeness of the evidence would imply. For this theory, the implication is not simply that managers are biased. It is that prior assessments can retain allocative force even after the information environment changes (Black & Vance, 2021).

Organizations explicitly distinguish current performance from capability for future work. Grabner, Künneke, and Moers study a professional-services organization that introduced potential assessments because current performance offers incomplete information about suitability for a prospective role with different demands. The intervention did not automatically improve sorting: supervisors differed in how well they discriminated among employees, and promoted employees performed worse on average after introduction of the new system. Church and colleagues similarly distinguish performance ratings, leadership assessment data, high-potential designation, and promotion outcomes in a large global organization. Together these studies make an important point for the present argument: organizations already institutionalize the proposition that current performance and prospective capability are not identical objects (Grabner et al., 2025; Church et al., 2021).

Generative AI can therefore create a familiar organizational problem under unfamiliar production conditions. A work product can remain behaviorally influential in staffing or promotion deliberations even after the human contribution required to produce that work has changed.

Managers are not passive conduits in this process. Let manager m observe an evidence set S and form an estimate of the focal capability:

The term M captures evaluator characteristics such as prior beliefs, attention, domain expertise, incentives, experience with the technology, assessment quality, and susceptibility to status or attribution effects. This manager term matters because organizational learning is heterogeneous. Grabner and colleagues document variation in supervisor assessment quality; Black and Vance show persistence in managerial beliefs; and Minni demonstrates that managers materially shape internal worker-job matching. Using personnel records covering roughly 200,000 white-collar workers and 30,000 managers across 100 countries, Minni finds that higher-quality managers cause workers to reallocate through lateral and vertical moves, with persistent gains in career progression and productivity. Managers help discover workers’ skills and change where those skills are deployed (Minni, 2026).

That evidence suggests two informational roles for managers. Managers act on existing beliefs, and they can create situations that improve those beliefs. A manager uncertain whether an employee can handle ambiguous client problems can deny access to such work, or can allocate a bounded, supervised case that limits downside while generating evidence. The first strategy treats knowledge as a prerequisite for assignment. The second uses assignment as an instrument of organizational learning.

Internal mobility and developmental-assignment research supports that second possibility. Bidwell and Keller argue that lateral moves can advance careers by allowing employees to develop and demonstrate new skills. Li, Krackhardt, and Niezink exploit random early-career project assignment and find that newcomers assigned to more complex projects obtain more learning-related credentials and more organizational visibility, with downstream career consequences. DeRue and Wellman find that developmental challenge is associated with leadership-skill development subject to diminishing returns and feedback conditions, while Dragoni and colleagues link access to developmental assignments with later managerial competence. Assignment therefore cannot be treated as merely the use of already-known human capital; it can change human capital and change the evidence available about it (Bidwell & Keller, 2024; Li et al., 2025; DeRue & Wellman, 2009; Dragoni et al., 2009).

Work-design theory provides the sociotechnical bridge. Parker, Morgeson, and Johns synthesize a century of work-design research organized around how job and role characteristics shape experience, motivation, behavior, learning, and performance. Parker, Van den Broeck, and Holman show that work design itself is shaped by higher-level organizational, local, managerial, and individual forces. Parker and Grote extend work design directly into the digital era, arguing that the consequences of automation and algorithms depend on autonomy, skill use, feedback, monitoring, relational features, and other design choices rather than on technology alone (Parker et al., 2017a; Parker et al., 2017b; Parker & Grote, 2022).

The present extension is informational. A workflow determines not only what cognitive work a person performs and what output the system produces, but what aspects of the worker become observable to organizational decision makers. Work design is therefore also evidence design—but the slogan matters only because it extends an established theoretical object. The informational properties of a job can affect the next allocation made from evidence generated by that job.

This makes work design endogenous to the learning system in a second sense. Managers and organizational structures partly determine which tasks are automated, which stages require review, where exception authority sits, and which workers receive low- versus high-scaffolding assignments. They therefore partly choose the environments that will later generate evidence about workers. The loop is not merely worker → performance → manager. It can also run manager → work architecture → evidence → managerial belief. The same organizational actors who consume capability evidence can influence how that evidence is produced.

The paper therefore makes an integration-and-extension claim rather than a priority claim over every component. It should fail as a contribution if existing work already explains the complete AI-mediated sequence from altered capability evidence through developmental allocation to later capability and subsequent evidence. It should also lose importance wherever AI-era measurement architectures recover the relevant human capability accurately before allocation occurs.

3. Capability Observability

The construct needed to analyze the first link in this mechanism is capability observability: the degree to which available organizational evidence supports accurate discrimination and calibration of a specified human capability for a specified decision, given the provenance of the observed performance.

Capability observability is therefore a property of an evidence–capability–decision relation, not a property of a worker, an artifact, or a measurement system in isolation. The definition is deliberately decision-relative. It does not ask whether the worker is globally “observable.” Organizations rarely need, and should not seek, total worker legibility. They need evidence about a particular capability because a particular allocation depends on it: independent error detection, judgment under ambiguity, client handling, transfer to novel problems, supervisory oversight, or another capability whose absence would materially change the decision.

Capability observability contains two properties familiar from assessment—discrimination and calibration—and one property that becomes especially consequential in AI-mediated work: provenance.

Discrimination

Discrimination is the extent to which available evidence differentiates workers who possess meaningfully different levels or configurations of the target capability. Suppose two analysts produce similarly strong AI-assisted recommendations. One can identify a subtly false assumption, reconstruct the causal logic, and generalize the analysis to an unfamiliar case; the other cannot. Their final outputs may discriminate poorly with respect to independent diagnostic judgment even if the same outputs discriminate meaningfully with respect to hybrid productivity.

Calibration

Calibration concerns whether inferred capability corresponds to actual capability at the level relevant to the allocation decision. A manager can rank two workers correctly while systematically overestimating both. Alternatively, average estimates can be accurate while individual ranking is poor. Overestimation can produce premature autonomy or responsibility; underestimation can withhold assignments through which already-present capability could become consequential and more visible.

Provenance

Provenance concerns whether the organization understands the generating process well enough to interpret what observed performance says about the human capability at issue. This is the AI-specific hinge of the construct.

The issue is not simply that AI adds statistical noise. AI can reduce noise and make performance more consistent. The problem is that the mapping from human capability to observed output becomes increasingly composite. A strong result may reflect independent human judgment, orchestration capability, model capability, retrieval quality, task scaffolding, or their interaction. The relative contributions can also change across model versions, tasks, workers, and time.

Provenance is therefore not a third generic psychometric property. It is the organizational link between the production process and the interpretation of its evidence. A disclosure that “AI assisted this task” can be accurate and still provide little information about the human contribution relevant to the next decision. Conversely, a full interaction trace can contain thousands of observations yet fail to reveal the focal capability if evaluators do not know which behaviors are diagnostic. One well-designed transfer task can sometimes provide more valid provenance than complete behavioral telemetry.

This asymmetry matters for the construct. Discrimination and calibration describe properties of the organization’s inference. Provenance conditions whether observed work can validly support that inference in the first place.

Capability observability is relational rather than intrinsic: the same artifact can be highly informative for one capability and nearly useless for another, and the same evidence can be sufficient for one allocation decision but inadequate for another.

Nothing in the theory implies that AI must reduce observability. AI can improve it. Interaction traces can show whether a worker challenged a model, noticed an inconsistency, sought corroboration, recognized an out-of-distribution problem, escalated uncertainty, or revised an attractive but faulty answer. Structured AI interaction can therefore generate more diagnostic evidence than a polished unaided artifact ever did.

This possibility is especially important for orchestration capability. If the future role requires workers to supervise, constrain, and appropriately rely on AI, then observable human–AI interaction may be the relevant performance rather than contamination of an otherwise “pure” human measure. A worker who cannot reproduce an AI-assisted artifact unaided can still possess high capability for the role if they reliably detect model failure, seek missing evidence, calibrate reliance, and know when to escalate. The theory therefore does not rank independent and hybrid capability in advance; it requires organizations to identify which capability a decision actually concerns.

This is also why capability observability is not equivalent to predictive validity, attribution, potential assessment, or developmental task value. Predictive validity asks whether a measure forecasts another outcome. Attribution asks who deserves credit for a hybrid result. Potential assessment is an organizational estimate of prospective capability. Developmental task value concerns whether performing a task contributes to future capability. Capability observability concerns the information available for making a specified capability inference from evidence generated by real work architecture.

The construct is measurable without being tautological. Discrimination can be assessed through correlations, rank-order accuracy, or classification accuracy between managerial estimates and an independently measured criterion capability. Calibration can be assessed through absolute or probabilistic error. Provenance accuracy can be tested by asking whether evaluators correctly identify which parts of performance depended on the human capability relevant to the decision rather than on AI contribution or orchestration. Criterion tasks can include unfamiliar transfer, deliberately misleading AI, AI degradation, error detection, appropriate abstention, causal diagnosis, or justified disagreement when those behaviors validly sample the capability of interest.

Measurement should also be consequentially aligned. The same evidence can have high observability for one decision and low observability for another. A rich active-review trace may reveal supervisory judgment but provide weak information about independent first-principles generation; an unaided transfer task may reveal independent reasoning but understate the worker’s ability to coordinate a hybrid production system. Observability is therefore indexed to a capability and a decision, not possessed by an artifact in the abstract.

The baseline comparison is not an imaginary pre-AI world in which human capability was perfectly visible. Human work has always been team-based, tool-mediated, and context-dependent. The theoretical claim is comparative: a change in production technology can change the uncertainty surrounding a specified human capability because it changes the evidence-generating process.

The theory also moves across levels without treating them as interchangeable. Capability J is an individual-level property. Ĵ is a manager- or organizational-level belief about that individual. Developmental allocation is an organizational decision that changes the individual’s subsequent capability and evidence. Outcomes such as succession robustness and allocation efficiency emerge from the repeated distribution of those individual-level processes across an organization.

Necessary conditions for provenance-contingent observability

Proposition 1 requires four conditions. First, the organization must care about a specified human capability rather than only combined-system output. Second, AI contribution must vary enough across workers, tasks, or episodes to alter the generating process of performance. Third, that variation must be insufficiently separable from the human contribution in the evidence ordinarily used for the allocation decision. Fourth, no alternative assessment can already recover the focal capability with comparable validity at negligible cost. If any condition fails, AI assistance need not reduce observability.

The proposition is therefore not a monotonic claim that more AI produces less information. Increasing AI contribution can improve observability when it creates diagnostic interaction behavior or standardizes irrelevant task variance. A model that exposes workers to the same calibrated failure cases, for example, can make differences in verification judgment easier to observe. The relevant comparison is between the information about J contained in alternative work architectures, not between “human” and “AI” work as categories.

Provenance also has to be capability-specific. Knowing that a model generated seventy percent of a draft does not establish whether the worker can recognize a defective legal argument; observing the worker reject one consequential hallucination may be far more informative. What matters is not a complete causal decomposition of output but sufficient evidence linking observed behavior to the human capability the decision requires.

4. Developmental Allocation: Assignment as Intervention and Measurement

Changes in observability become organizationally consequential only if capability beliefs affect what happens next. The central claim of this section is therefore simple: assignments are endogenous measurement environments.

Organizations allocate workers to heterogeneous experiences. Some are routine and highly scaffolded. Others contain novelty, responsibility, feedback, discretion, customer exposure, cross-functional interaction, exception handling, or leadership. Research on developmental assignments establishes that these differences can matter for capability formation. DeRue and Wellman find that developmental challenge is associated with leadership-skill development subject to diminishing returns and feedback conditions; Dragoni and colleagues link access to developmental assignments with later managerial competence. Experience is therefore not uniformly developmental, but its design and allocation can alter capability (DeRue & Wellman, 2009; Dragoni et al., 2009).

Kahneman and Klein provide an important boundary on what counts as developmental experience. Expertise requires environments containing sufficiently learnable regularities and opportunities for feedback. Preserving human exposure to work that is effectively unpredictable does not automatically produce judgment. A developmental assignment therefore earns that label through challenge, responsibility, feedback, and the possibility of learning relevant structure—not simply because a human performs it without AI (Kahneman & Klein, 2009).

The same assignments also generate evidence. Li, Krackhardt, and Niezink’s random project assignment shows that more complex early projects increase both learning-related outcomes and organizational visibility. Bidwell and Keller theorize lateral moves as opportunities through which employees can develop and demonstrate new skills. An assignment can therefore be valuable simultaneously because it produces output, changes the worker, and reveals something about the worker (Li et al., 2025; Bidwell & Keller, 2024).

Production is immediate organizational value. Formation is capability change caused by the experience. Revelation is the new evidence the experience generates about capability. Status is the visibility or standing associated with receiving or succeeding in the assignment. The four need not covary. A prestigious project can create status while teaching little. A simulation can reveal capability while producing no commercial output. A heavily scaffolded AI workflow can create substantial production value while revealing little about independent exception judgment.

The theory centers on formation and revelation. Let

where developmental allocation D depends partly on inferred capability Ĵ, along with staffing constraints, preferences, project requirements, and opportunity scarcity represented by Z and C. Assignment then changes capability and the next evidence set:

Assignment is therefore simultaneously an intervention on the capability the organization is trying to estimate and an intervention on the future data from which that estimate will be made.

A worker who never receives an exception case cannot demonstrate exception-handling judgment in that case. A worker denied customer exposure cannot produce evidence about client judgment. An employee routed repeatedly toward highly scaffolded production may continue generating valuable work while producing few observations about the capabilities required for less scaffolded responsibility.

This creates informational path dependence. The consequences become clearer when inference accuracy is separated from allocation.

When a capable worker is accurately recognized and receives difficult work, assignment can both deepen capability and generate confirming evidence. When a worker is accurately judged not yet ready, restraint need not end development: the organization can allocate lower-risk learning or diagnostic experiences that preserve a route to later reassessment. These are ordinary success cases.

False positives and false negatives are more theoretically revealing. A false positive occurs when the organization overestimates capability and grants an opportunity for which the worker is not yet ready. The assignment may expose the error. But it can also develop the worker enough that the initial belief becomes less false over time. Some assessment errors are therefore performative: allocation can change the underlying capability that the original assessment incorrectly presumed.

A false negative creates the sharper dynamic. The organization underestimates a capable worker and withholds developmental or revealing work. The worker loses both an opportunity to build additional capability and an opportunity to produce evidence that could overturn the original underestimate:

The worker becomes less evidenced as capable partly because the organization withheld the situations in which capability could become evident. I call this observability foreclosure: an adverse capability inference reduces access to the situations capable of generating evidence that could overturn that inference. The absence of confirming evidence can then be mistaken for evidence of absence. This is the strongest form of formation–revelation feedback.

The mechanism also distinguishes developmental scarcity from diagnostic scarcity. Developmental scarcity exists when only some workers can access experiences that build a capability. Diagnostic scarcity exists when only some situations reveal that capability clearly enough to update organizational beliefs. The two can coincide, but they need not. A bounded failure simulation can be highly diagnostic while only modestly developmental; a difficult internal project can be developmentally rich yet weakly visible to future evaluators. When both types of scarcity are allocated from the same uncertain beliefs, the recursive mechanism is strongest.

Scarcity also makes managerial conservatism dynamically consequential. When a manager routes high-consequence work to already trusted people, the decision can be locally rational even if the manager is unbiased and well intentioned. But repeated conservative routing increases the experienced workers’ stock of both capability and evidence while reducing the occasions on which less-established workers can become legible. The feedback loop therefore does not require explicit favoritism. It can emerge from risk-sensitive allocation under uncertainty.

A derived implication follows from reduced observability. When current AI-mediated work becomes less diagnostic of the focal capability, prior evaluations, credentials, reputation, and status become relatively more informative for developmental allocation. This does not require irrational managers. It follows from relative information. Black and Vance show that prior assessments can retain allocative force even as new performance evidence accumulates; Call, Jiang, and Idso supply a complementary AI-specific mechanism through which star status can shape value capture under ambiguous hybrid performance (Black & Vance, 2021; Call et al., 2026). Valid direct capability evidence should attenuate this history dependence.

Why allocation mediates rather than merely follows inference

For formation–revelation feedback to matter, managerial beliefs must influence at least some allocation margins. The theory does not require capability estimates to determine every assignment. Client demand, staffing availability, worker preference, seniority, and formal rules can dominate many decisions. The mechanism operates where managers retain discretion over access to work that differs materially in developmental or diagnostic value.

This condition also identifies a clean empirical boundary. In organizations that rotate workers randomly through equivalent developmental experiences, capability-inference error should have little effect on formation even if observability is poor. In highly discretionary systems, the same inference error can be consequential because it changes exposure. Thus the causal quantity of interest is not the correlation between belief and assignment alone, but the change in developmental or revealing exposure attributable to belief after other assignment determinants are held constant.

Prior substitution concerns relative rather than absolute reliance on historical signals. Credentials or reputation can influence allocation even when current work is perfectly informative. The prediction is that their marginal weight increases as current capability evidence becomes less diagnostic and decreases when valid direct evidence is supplied. This moderation distinguishes signal substitution from a simpler theory of stable status advantage.

5. Dynamic Predictions: Diagnostic Scarcity

The recursive model produces two counterintuitive predictions that can be understood as forms of diagnostic scarcity: a shortage of naturally occurring situations in which consequential human capability becomes observable.

Reliability and exception-state capability

Consider a capability expressed mainly when ordinary automation is wrong or insufficient: recovery, verification, escalation, justified disagreement, or diagnosis under unusual conditions. When a system fails frequently, workers encounter many occasions on which these capabilities can be exercised and observed. Those failures are undesirable from a production standpoint, but they generate information.

As AI reliability increases, average performance improves and naturally occurring failure episodes become rarer. The worker has fewer occasions to exercise exception-state capability, and the organization receives fewer observations from which to distinguish who possesses it:

Classical automation research provides the relevant genealogy but not the same outcome. Bainbridge’s “ironies of automation,” Endsley and Kiris’s work on out-of-the-loop performance, and Parasuraman and Riley’s analysis of automation use and misuse examine whether operators remain capable of effective intervention under automation. The present theory asks a second-order organizational question: even if some workers retain recovery capability, how will the organization know which workers possess it when reliable automation rarely creates a diagnostic episode? (Bainbridge, 1983; Endsley & Kiris, 1995; Parasuraman & Riley, 1997).

The paradox can become organizationally more important precisely as technical reliability improves. If a firm reorganizes work around the assumption that routine AI assistance will be continuously available, fewer workers may encounter naturally revealing failures while the organization becomes more dependent on competent human intervention in the rare states where ordinary automation is insufficient. The theory does not assume that dependence always rises with reliability; redundancy, deliberate practice, and direct assessment can break the relationship. It predicts that routine success can cease to generate enough information about rare but consequential capability.

The effect is bounded to capabilities expressed in exception states. Higher reliability can improve observability of orchestration skill, ordinary task knowledge, or capabilities expressed continuously in the workflow. The prediction concerns precisely those human capabilities that routine success renders latent.

Performance compression and latent divergence

Diagnostic scarcity can also arise when AI compresses visible performance differences. Suppose assistance raises weaker current performers toward a common acceptable threshold so that variance in routine output declines:

No developmental conclusion follows from compression alone. Latent divergence requires three additional conditions: developmental opportunities remain heterogeneous; managers allocate those opportunities selectively; and those opportunities materially affect later capability. Under those conditions, a worker receiving repeated complex or exception-rich assignments accumulates formation and evidence while another worker continues to produce similarly acceptable routine AI-assisted output without equivalent exposure.

The distinctive causal sequence is:

Observed equality and latent equality can therefore move in opposite directions. Current output can cease to reveal a future capability distribution that allocation is actively producing. The empirical signature is declining cross-sectional variance in routine output alongside increasing variance on direct or exception-state capability measures.

This prediction also distinguishes production equalization from developmental equalization. AI may be valuable precisely because it allows more workers to meet an acceptable performance standard. The theory does not treat that compression as undesirable. It predicts a different risk: organizations may infer capability convergence from output convergence even when selective developmental exposure is producing increasing differences underneath. The appropriate response is not to restore visible inequality in output but to measure the capability distribution directly where it remains consequential.

Both parts should weaken when organizations deliberately substitute valid evidence for what routine production no longer reveals. Direct assessments, simulations, transfer tasks, structured critique, degraded-AI conditions, and diagnostic assignments can all reduce dependence on ambiguous routine output. This is not a separate proposition; it is a boundary condition of the theory. The mechanism should be weakest exactly where measurement architecture adapts successfully to production architecture.

What would make diagnostic scarcity disappear?

The diagnostic-scarcity prediction requires more than reliable or compressive AI. It requires the focal capability to be expressed intermittently rather than continuously, and it requires ordinary work to remain an important source of organizational inference. If recovery capability is directly tested every quarter, for example, the disappearance of natural failure episodes need not reduce observability. If managers observe rich, validated process evidence during routine work, reliability can improve output without hiding the human capability needed for exceptional states.

The prediction also differs from simple skill atrophy. A worker can retain exception capability while the organization becomes less able to identify that worker. Conversely, the organization can maintain excellent information about exception capability through simulation even if routine work provides no practice. The human-factors and organizational-learning mechanisms can therefore move independently. This distinction is empirically useful: a study that measures only later human performance cannot establish diagnostic scarcity, while a study that measures only managerial confidence cannot establish capability erosion.

Part (b) of Proposition 3 has an equally important limiting case. Output compression should not generate latent divergence if developmental exposure is broadly distributed, if AI supplies equally effective learning across workers, or if independent assessments guide allocation accurately despite compressed routine output. The proposition therefore predicts a three-way interaction among performance compression, selective allocation, and developmental returns to assignment rather than a universal relationship between AI use and capability inequality.

6. Boundary Conditions: When Does Capability Observability Matter?

The strongest objection is not that AI corrupts assessment. It is that the attempt to recover “independent” human capability may itself be a category error. Expertise in a sociotechnical system is partly constituted by competent tool use. Separating the worker from the tool can measure a capability the organization neither uses nor should value.

This objection is substantially correct. Independently exercisable human capability has no universal organizational priority. If AI is continuously available, reliable across relevant states, institutionally supported, and superior at the focal activity, requiring unaided replication can duplicate effort, discourage useful complementarity, and preserve an obsolete form of competence.

The theory therefore concerns only human capabilities whose independent exercise remains causally necessary in identifiable future states. Those states can include model failure, distribution shift, adversarial input, regulatory requirements, supervisory responsibility, conflicting recommendations, high-consequence escalation, system redesign, or another condition in which a human must decide whether and how the machine should be trusted.

The expression is not a universal formula; it disciplines the claim. A low-consequence process supported by robust automation may require little independent-capability measurement. A high-consequence role in which humans remain accountable for unusual failures may require substantially more.

The full mechanism is therefore strongest when four conditions coincide: AI contributes materially to observed performance; capability-relevant provenance is limited; developmental or revealing opportunities are selectively allocated; and failure of the focal human capability remains consequential. Remove any one condition and the mechanism weakens.

The theory also predicts its own attenuation. As organizations introduce valid direct evidence of the target capability and use it in allocation, ambiguity in routine AI-mediated performance should matter less. AI can therefore solve part of the measurement problem AI creates. Process evidence, simulations, structured critique, transfer tests, and diagnostic assignments can all improve observability when they validly sample the focal capability.

The counterfeit solution is measurement maximalism. A firm could log every prompt, revision, pause, rejection, and model interaction in an attempt to make workers fully legible. Trace abundance does not guarantee construct validity. Workers can adapt to visible metrics; surveillance can alter behavior; and detailed telemetry can tempt evaluators to infer more than the data support.

The positive design principle is sufficient decision-relevant observability: enough valid information about a human capability that materially matters to a consequential allocation decision, under the conditions in which that capability will be required. The objective is not maximal worker legibility.

That principle also creates a contestability requirement. Repeated allocation from low-provenance evidence can become self-reinforcing, especially in false-negative cases. An under-observed worker should have some route to generate new valid evidence—through a diagnostic assignment, direct assessment, supervised exception task, or another mechanism appropriate to the role. The theory does not imply a right to promotion or to every assignment. It implies that an evidentiary system that never permits a negative inference to be challenged can reproduce its own mistakes.

7. Capability-Revealing Work Design

Organizations can recover lost capability information in two broad ways: assess capability outside ordinary production or redesign production so that useful evidence survives within it. The appropriate response depends on the capability being assessed and the future work architecture.

Judgment sequence

Yan’s randomized experiment with 240 preservice teachers compared an independent-judgment-first procedure with a recommendation-first procedure. Under misleading AI recommendations, participants who recorded their own judgment before seeing the recommendation were more accurate and less likely to reproduce the erroneous advice completely; under expert-keyed recommendations, the difference was not statistically significant. The study is scenario-based and does not establish long-run expertise or organizational allocation effects, but it shows that workflow sequence can causally change responses to misleading AI advice (Yan, 2026).

Sequence also has an informational consequence. A human-first architecture produces a pre-AI observation that can later be compared with the final hybrid judgment. An AI-first architecture does not. That does not make human-first work universally superior. A rushed pre-AI answer can be noisy, and novices can benefit from scaffolding. The information is valuable only when the pre-AI task validly samples a capability the organization still needs to know.

Active review

AI-first work can itself be developmentally and informationally rich. Petropoulos models review intensity as a determinant of long-run expertise formation: sufficiently intensive review can make the same AI technology expertise-building rather than expertise-eroding. The present theory adds an informational implication. Review can also generate capability-relevant evidence when workers must diagnose weaknesses, reconstruct reasoning, identify uncertainty, generate alternatives, justify rejection, or specify failure conditions (Petropoulos, 2026).

For some future roles, this evidence may be more valid than unaided reproduction. If the expert’s job is to supervise AI systems rather than recreate routine output manually, the focal capability may be calibrated review and justified refusal. AI-first design is therefore not thin by definition; passive acceptance is.

Transfer, degradation, and diagnostic assignments

A transfer problem tests whether reasoning survives changes in surface features. A degraded-AI condition tests recovery when support becomes incomplete or wrong. An error-injection exercise tests diagnosis. A simulation can expose rare exception-handling capability without waiting for an expensive real failure. Williams’s RungBuilder architecture develops a related applied logic around transfer testing and bounded entrustment, while its effectiveness remains an empirical question (Williams, 2026).

The most organizationally distinctive intervention is a capability-revealing assignment. Managers can allocate real work partly for its information value. Lateral moves, supervised stretch projects, rotations, bounded exception cases, temporary responsibility, and complex reviews can generate production while allowing an under-observed worker to demonstrate capability. Bidwell and Keller’s account of lateral mobility and Minni’s evidence on manager-driven worker-job reallocation show why assignment can be a discovery mechanism rather than merely a destination for already-known talent (Bidwell & Keller, 2024; Minni, 2026).

This produces a different response to uncertainty than simply withholding responsibility. A manager uncertain about an employee’s exception judgment can allocate a bounded case with supervision and explicit escalation rights. The organization learns by assigning rather than only assigning after it has learned.

No single evidence architecture should dominate all work. Routine AI-first production best measures the combined system’s ordinary output. Human-first sequence preserves one kind of provenance. Active critique can reveal supervisory capability. Transfer and degraded-system tasks can expose portability or recovery. Diagnostic assignments can combine production, formation, and revelation. Organizations therefore face a portfolio problem rather than a universal rule.

The governing principle is: when routine production stops revealing a consequential capability, create another valid occasion on which that capability can be exercised and observed.

8. Distribution and Organizational Consequences

The formation–revelation feedback has distributive consequences, but the theory does not establish demographic effects. It predicts a structural mechanism: when current capability evidence becomes weaker, prior signals become more influential in allocation.

Any pre-existing inequality embedded in prior evaluations, credentials, status, social visibility, or access to diagnostic assignments can therefore acquire greater allocative importance when current work becomes less informative. Call, Jiang, and Idso’s AI-specific Matthew Effect supplies a complementary mechanism through which status can shape value capture under AI augmentation; the present theory specifies how weaker capability evidence can increase the relative weight of such priors in developmental allocation (Call et al., 2026).

The important scarce resource is therefore not only developmental work but revealing work. Employees differ in access to assignments through which higher-order capability becomes visible. A worker can repeatedly perform valuable AI-mediated work while receiving few opportunities to produce evidence about independent diagnosis, judgment under ambiguity, customer responsibility, or leadership. The absence of such evidence can subsequently rationalize continued exclusion from those assignments.

The justice pressure is specific: who receives the opportunities through which capability can become both stronger and more visible; what it costs to obtain alternative evidence; and whether workers have a credible route to challenge an inference produced from low-provenance performance. These questions should not be inflated into an unsupported demographic claim. They concern differential access to formation, revelation, and appeal.

Allocation efficiency

Worker-job matching depends on information about workers. Minni demonstrates that managers create persistent productivity gains by reallocating workers toward jobs that better fit their specific skills. If AI changes the evidentiary value of performance without corresponding improvement in assessment, current task productivity can rise while future talent-allocation efficiency falls (Minni, 2026).

Succession robustness

Organizations can also become uncertain about who is prepared for roles requiring judgment beyond routine AI-mediated production. The uncertainty can arise because capability was not developed, because it was developed but poorly observed, or because both occurred simultaneously. Restoring developmental work will not solve an observability problem if the organization still cannot distinguish who acquired the capability. Better assessment will not solve a formation problem if workers never receive experiences through which capability can grow.

This distinction matters most when succession decisions are made from cumulative records of prior assignments. A sparse record of difficult work can mean that a worker lacks capability, that managers previously lacked confidence, or that the worker simply lacked access to diagnostic opportunities. Without alternative evidence, those explanations are observationally similar. The theory therefore predicts that organizations can face succession uncertainty even when sufficient capability exists somewhere in the workforce but has not been made legible through prior allocation.

Concentration of judgment

A locally rational response to uncertainty is to send difficult exceptions to already trusted workers. Doing so minimizes immediate production risk. Repeated routing, however, can concentrate the experiences that develop and reveal high-level judgment among workers whose capability was already easiest to observe. No task must disappear. Development can become selectively concentrated through allocation itself.

This distinction matters for intervention. If developmental tasks have disappeared, organizations need substitutes for learning. If judgment-intensive tasks remain but are routed mainly to already trusted workers, organizations need better allocation rules and bounded diagnostic access for under-observed workers.

9. Empirical Test

The theory is useful only if capability observability, developmental allocation, and formation–revelation feedback can be separated empirically from ordinary productivity, attribution, and selection effects.

Construct validation must begin by defining the target capability before outcomes are observed. Researchers should specify the capability, the evidence available to the evaluator, the consequential allocation decision, and an external criterion against which discrimination and calibration can be assessed. Defining “high observability” as whatever later predicts capability well would make the construct tautological.

The criterion must also be independent of the same artifact used to form the original inference. If managers judge capability from an AI-assisted report and researchers validate those judgments with another rating of the same report, the design measures agreement about artifacts rather than inference about capability. Strong tests require transfer, changed task conditions, deliberately misleading AI, degradation, or another outcome whose connection to the focal capability is theoretically warranted.

The strongest initial design is a multistage experiment.

Stage 1 · Generate different evidence architectures

Workers complete domain-valid tasks under randomized conditions such as no AI, AI first, human judgment first followed by AI, and AI first with structured active critique. Separately, workers complete a validated task measuring the focal capability. The benchmark should test transfer, exception handling, diagnosis, justified disagreement, or another capability genuinely relevant to later assignment rather than simply repeat the production task without the tool.

Stage 2 · Measure observability

Experienced managers or domain experts receive randomized evidence portfolios: final AI-assisted output only; output plus disclosure of AI use; capability-relevant provenance; a pre-AI judgment; an active-review trace; or a direct assessment. Evaluators estimate each worker’s capability and confidence.

A distinctive test of Proposition 1 holds final output quality constant while varying the provenance available to evaluators. If managers form materially different and more accurate estimates of the same underlying capability, the evidence architecture has changed observability rather than simply performance.

Stage 3 · Allocate scarce work

Managers allocate a fixed portfolio of consequential opportunities: developmentally rich assignments, mentoring, autonomy, exception-handling responsibility, high-status work, lateral mobility, and promotion recommendations. Prior status or reputation can be manipulated orthogonally. The derived prior-substitution implication predicts that those historical signals should exert greater relative influence when current capability evidence is weak and less influence when valid direct evidence is supplied.

Selection is the main identification challenge. In real organizations, capable workers are precisely the people managers choose for difficult work, so a positive relationship between challenge and later capability can reflect selection rather than development. Random project assignment, manager rotation, staggered policy change, or other plausibly exogenous variation is therefore crucial. Li and colleagues’ random early-career project assignments and Minni’s manager rotations illustrate the kind of identification needed.

Stage 4 · Observe formation and revelation

Participants complete the allocated experiences. Researchers later measure both capability and the accuracy of subsequent capability beliefs. This distinction is essential. If an assignment improves skill without improving later inference, it supports formation but not revelation. If it improves belief accuracy without changing skill, it supports revelation but not formation. If it changes both, the dual mechanism receives support.

False-positive and false-negative trajectories should not be averaged away. An initial false positive can lead to failure, correction, or capability growth that makes the initial overestimate less false. A false negative can persist because the worker never enters an environment in which corrective evidence would be produced. The empirical signature of informational path dependence is therefore not simply later performance; it is whether initial inference error predicts subsequent access to evidence capable of correcting that error.

Identifying Formation–Revelation Feedback

The decisive empirical claim is not that assignment predicts later capability. It is that an initial capability inference changes allocation, allocation changes both later capability and the availability of evidence capable of correcting that inference, and those changes affect the next capability estimate. A credible design therefore requires four elements: baseline capability, managerial belief before allocation, plausibly exogenous variation in developmental or revealing assignment, and later measures of both capability and evidence or belief accuracy.

Random project assignment, manager rotation, or staggered workflow redesign can provide leverage. A false-negative path supports observability foreclosure only if an underestimated worker receives less revealing work and the resulting record becomes less informative than the record of a comparable worker who receives a diagnostic opportunity. A false-positive path is supported when richer assignment either exposes the initial error or develops the worker enough that the original overestimate shrinks.

Repeated manager forecasts can separate formation from revelation. Improved capability without more accurate beliefs indicates formation; improved belief accuracy without capability growth indicates revelation; movement in both identifies the dual mechanism. Field studies should also measure evidence availability itself—complex cases, exception states, client exposures, review episodes—so that persistence in belief can be distinguished from persistence caused by differential access to corrective evidence.

The theory is materially weakened if AI-mediated work remains equally or more predictive of the relevant capability under low-provenance conditions; provenance information does not improve inference; capability beliefs do not affect developmental allocation; developmental assignments do not affect later capability; assignments generate no additional information about capability; weaker current evidence does not increase reliance on priors; or direct assessment already recovers the relevant capability so accurately and cheaply that ambiguity in routine work is allocatively irrelevant.

The final falsifier is decisive: if future work rarely requires the independent capability the organization is attempting to observe, the measurement problem is economically unimportant.

10. Discussion

The theory makes two extensions connected by one recursive mechanism.

First, it extends employer learning and internal-labor-market theory by making observation technology endogenous to work technology. Existing theory explains that firms learn about worker ability, workers accumulate human capital, and jobs are allocated from those beliefs. Generative AI changes what a performance observation can mean because the same visible output can arise from different combinations of human judgment, orchestration, and machine contribution. Employer learning therefore depends not only on the number of performance signals available but on how work architecture generates those signals.

This extension matters because allocation generates the next round of information. Internal labor markets are repeated systems of learning through allocation, not systems that learn completely and only then allocate. Managers can create value not only by matching known talent to jobs but by assigning work through which uncertain talent becomes knowable. Parker, Van den Broeck, and Holman’s work on the multilevel antecedents of job design reinforces this point: managers and organizational contexts partly shape the work architectures that subsequently generate the evidence managers use.

Second, the theory adds an informational consequence to work design. Work architecture determines not only what people do, what they learn, and what they produce, but what the organization can infer about them. The production-efficient workflow and the talent-informative workflow need not be identical. A highly reliable AI-first process may maximize routine throughput while revealing little about rare recovery capability. A diagnostic assignment may sacrifice some immediate efficiency but improve subsequent allocation. An active-review workflow may achieve both.

This creates a portfolio problem rather than a universal prescription. Organizations need not make every work episode diagnostic. They need enough valid revealing episodes to support the consequential decisions whose stakes justify the cost of measurement.

Formation–revelation feedback connects these extensions. Developmental allocation is the mechanism through which altered observation becomes altered capability: managers allocate from beliefs, and those assignments subsequently change both the worker and the evidence available for the next belief. This distinguishes the argument from task-removal accounts. A firm can preserve excellent developmental assignments yet reproduce expertise poorly if those assignments are allocated from weak capability evidence; conversely, it can automate traditional junior production without the same problem if alternative learning experiences are paired with valid allocation evidence.

The formation–revelation feedback also explains why opportunity can compound without a deliberately discriminatory rule. A worker denied a rich assignment loses both development and an opportunity to generate future evidence. The next manager observes a record partly produced by the earlier allocation. Opportunity can therefore accumulate epistemically as well as developmentally.

The strongest alternative is optimistic: AI may improve observability. Process traces can reveal error detection, escalation, revision, uncertainty, and transfer more clearly than conventional final artifacts. Models can generate standardized diagnostic scenarios cheaply. If those measurements predict consequential capability and enter allocation appropriately, observability can increase even as AI contributes more to production. The risk identified by the theory is therefore an asymmetry of adaptation: production systems can change faster than the assessment and allocation systems built around them.

That alternative also identifies the theory’s practical upside. Organizations need not choose between AI productivity and human legibility. They can redesign evaluation around the human capabilities that remain consequential in the future role and use the technology itself to create calibrated diagnostic states. AI can become part of the measurement architecture rather than merely a source of measurement ambiguity.

A firm can become better at producing today’s work while becoming worse at learning who should perform tomorrow’s work. It can avoid that tradeoff only if observation evolves with production.

That is why “human in the loop” is too crude a design principle. A human can remain formally present while revealing little independent judgment. Another worker can encounter AI first and reveal exceptional expertise through critique, refusal, or recovery. The relevant questions are which human capabilities the future role still requires, what the workflow makes observable about those capabilities, and how the resulting evidence shapes access to the experiences through which capability develops.

Conclusion

Organizations do not allocate future work from capability itself. They allocate from beliefs about capability. When AI changes the evidence from which those beliefs are formed, it can change who receives the assignments through which capability subsequently develops and becomes visible.

Assignment is therefore both a developmental intervention and a measurement environment. A worker misread today can receive a different opportunity tomorrow, become differently capable because of that opportunity, and enter the next evaluation with a different evidentiary record.

The challenge is not to preserve unaided work indiscriminately, but to preserve sufficient decision-relevant evidence about human capabilities that still matter.

When machines improve the work, organizations need new ways to know the worker.