Source-linked AI summary
When AI performs, what remains? An integrative review of generative AI and cognitive capital in individuals and organizations
Giuseppe Riva, Stefania La Rocca, Brenda K. Wiederhold, Andrea Gaggioli
TL;DR
Generative AI’s productivity benefits are increasingly documented, but less is known about what knowledge, skills and judgment remain in people and organizations over time. This integrative narrative review develops a multilevel cognitive-capital framework from conceptual, experimental and field work. It concludes that AI can amplify or erode cognitive capital, and that short-term output alone does not reveal which is happening.
Problem
Productivity measures do not show what knowledge, skills and judgment people and organizations retain over time after adopting generative AI.
Source: Introduction · p. 2Method
The integrative narrative review combines conceptual frameworks with experimental, field and observational evidence to develop four mechanisms and a multilevel model.
Source: Introduction · p. 2, Four mechanisms linking AI use to… · p. 5Results
Generative AI can amplify or erode cognitive capital, and short-term outputs do not reveal which; assisted performance is not retained competence, and effects depend on technology, design and use.
Source: Conclusions · p. 14Takeaways & Limitations
Cognitive capital offers a way to distinguish system performance from human capability and frame AI adoption around measuring, renewing and protecting organizational capacity.
Source: Conclusions · p. 14The review is integrative rather than systematic, so selection bias cannot be excluded; much experimental evidence is short-term or setting-specific, and organizational effects are largely inferred from individual-level studies.
Source: Limitations and research agenda · p. 13
Abstract
from arXiv · showhide
Generative artificial intelligence is rapidly entering knowledge work, and its short-term productivity benefits are increasingly documented. Less is known about what happens, over time, to the knowledge, skills and judgment that remain in people and organizations. This integrative narrative review addresses that question through the construct of cognitive capital: the accumulated, renewable and depletable stock of knowledge, cognitive and metacognitive capabilities, and practices that enables an individual or an organization to learn, interpret information, exercise judgment, solve problems and adapt. We distinguish individual and organizational cognitive capital from the capability of the human-AI system, so that gains in joint performance and losses in human capability can be observed at the same time. Drawing on experimental, field and conceptual work, we propose four mechanisms through which interaction with AI affects cognitive capital: substitution of practice, scaffolding and feedback, metacognitive calibration, and organizational aggregation. Seven principles summarize the current evidence, including that assisted performance differs from retained competence, that the same technology can support or displace learning depending on design and use, and that individual gains can coexist with reduced collective diversity. For each mechanism we derive testable propositions, and we discuss five design levers for organizations as proposals to be validated rather than established practice. We conclude that cognitive capital offers a multilevel frame for governing AI adoption and outline a research agenda centered on longitudinal, multilevel measurement.
Introduction
AI can improve immediate productivity, but those gains do not establish what knowledge or capability people retain. This review examines when delegation supports human development versus replacing practice over time.
Introduction
- AI’s productivity gains measure assisted performance, not what remains when the system is withdrawn.A rollout to 5,172 customer-support agents raised issues resolved per hour by 15% on average.Source: Introduction · p. 2
- A proposed System 0 describes algorithmic systems pre-organizing information before deliberation, while AI outputs become knowledge through human interpretation.The authors note that their argument does not depend on this framework.Source: Introduction · p. 2
- Delegation can extend capability or replace its exercise, with the two effects difficult to distinguish in the short term but potentially diverging over time.Whether delegation sustains development or preserves strategic capabilities is an empirical question.Source: Introduction · p. 2
- The paper is an integrative narrative review combining conceptual frameworks with experimental, field and observational evidence to propose a multilevel model and research agenda, not a validated theory.Source: Introduction · p. 2
Review approach
The review uses iterative searches and thematic synthesis to integrate empirical and conceptual work, aiming for conceptual integration rather than exhaustive retrieval or quantitative pooling.
Review approach
- The review asks when generative AI interaction develops, maintains or depletes cognitive capital, distinct from changes in human–AI system performance.Searches were iterative and targeted, with final search and bibliographic verification on 19 October 2026.Source: Review approach · p. 2
- Empirical studies were selected for methodological detail on AI-assisted performance, learning, retention, transfer, judgment, metacognition or collective diversity.Conceptual contributions and formal models were included when they clarified the construct or proposed mechanisms.Source: Review approach · p. 2
- The review included conference papers, selected preprints and working papers, and adjacent non-generative AI studies, whose transfer to generative systems remains unverified.Source: Review approach · p. 3
- Iterative thematic synthesis yielded four mechanisms and seven principles; propositions and design options are interpretive extensions requiring validation, and no quantitative pooling was undertaken.The search also sought evidence challenging inevitable depletion, including learning gains, retained productivity during outages, stronger subsequent learning and improved decision accuracy.Source: Review approach · p. 3
Defining cognitive capital
The review defines cognitive capital as a renewable and depletable stock of human knowledge, capabilities and practices, considered separately from human–AI system capability. Its multilevel framing distinguishes individual and organizational stocks and proposes criteria for assessing whether AI-mediated work renews or depletes them.
Defining cognitive capital
- Cognitive capital is the accumulated, renewable and depletable stock of knowledge, cognitive and metacognitive capabilities, and practices enabling learning, interpretation, judgment, problem solving and adaptation.Source: Defining cognitive capital · p. 3
- Individual cognitive capital includes acquired knowledge and capacities such as attention, memory, reasoning and self-monitoring, while organizational capital integrates expertise with shared knowledge, routines and knowledge-use mechanisms.Organizational cognitive capital exceeds the sum of employees’ capabilities.Source: Defining cognitive capital · p. 3
- AI-assisted performance alone does not establish growth in cognitive capital; growth requires retained knowledge, transferable competence or improved collective capacity to learn and act.Declines in those capacities count as depletion even when productivity metrics do not show them.Source: Defining cognitive capital · p. 3
- The framework distinguishes individual and organizational cognitive capital from human–AI system capability; Table 1 presents three objects of assessment in AI-mediated work.The distinction makes visible cases where system performance improves while human capacities needed to supervise or replace it weaken.Source: Defining cognitive capital · p. 4, Table 1 · p. 4
- Cognitive capital connects human capital, absorptive capacity and distributed cognition around capabilities available for subsequent learning, judgment and adaptation; whether it adds discriminant validity remains an empirical question.Its proposed added value would be supported if its measures explained subsequent learning, transfer or resilience beyond existing constructs and current performance.Source: Defining cognitive capital · p. 4
- The review links human, machine and hybrid cognition to evidence that practice performance can diverge from learning, while a formal model suggests accurate recommendations may reduce shared general knowledge under some conditions.The model describes a theoretical possibility, not an observed organizational trajectory.Source: Defining cognitive capital · pp. 3–4
Four mechanisms linking AI use to cognitive capital
The review organizes evidence into four proposed mechanisms—substitution of practice, scaffolding and feedback, metacognitive calibration, and organizational aggregation—while treating joint human–AI performance as a cross-cutting assessment. Its seven principles summarize evidence that mostly measures individual performance over minutes to weeks, with organizational effects largely inferred.
Mechanism 1: Substitution of practice
- Removing AI after practice can expose weaker retained competence, especially after answer-seeking [22]; offloading is risky when it replaces the practice that builds skill.Field evidence is mixed: customer-support agents retained part of their productivity gains during outages. Survey findings are associative, and essay support improved scores without significantly different knowledge gain or transfer.Source: Mechanism 1: Substitution of practice · pp. 5–6
Mechanism 2: Scaffolding and feedback
- Active-learning tutoring and hint-based support improved learning outcomes, while concept-explanation use accompanied larger delayed gains than having AI write [28].A framework proposes reciprocal, reliable, transparent, personalized support grounded in calibrated trust; scaffolding should fade as competence grows.Source: Mechanism 2: Scaffolding and feedback · p. 6
Mechanism 3: Metacognitive calibration
- Expertise shaped assisted and retained gains, while agreeable models could earn trust yet effects on judgment varied by context; these studies do not establish long-term cognitive growth [25].The colonoscopy finding involved non-generative AI and was observational, so it does not establish causal deskilling.Source: Mechanism 3: Metacognitive calibration · pp. 6–7
Joint performance across the three objects
- Human–AI combinations outperformed humans alone on average but underperformed the better of human or AI alone, especially in decision tasks.Combination tended to help when humans outperformed AI alone.Source: Joint performance across the three… · p. 7
Mechanism 4: Organizational aggregation
- Individual gains can coincide with reduced collective diversity: AI-generated ideas made stories more similar, while rewarding originality reduced verbatim adoption without restoring unassisted-writing diversity [43].Across 41.3 million papers, AI adopters gained publications and citations as AI-augmented science contracted toward data-rich topics; this evidence spans earlier AI generations, not just current chatbots. Organizations may gain efficiency while losing variety.Source: Mechanism 4: Organizational aggregation · pp. 7–8
From mechanisms to testable propositions and design levers
The review turns four proposed mechanisms into falsifiable propositions and presents five organizational design levers as proposals for validation, not established practice. Collaboration configurations also differ in their implications for cognitive capital.
From mechanisms to testable propositions and design levers
- P1–P4c specify testable predictions for substitution, scaffolding, calibrated friction, and organizational aggregation.They address delayed learning and transfer, error detection and overreliance, output quality and solution diversity, and resilience during outages.Source: From mechanisms to testable… · pp. 11–12
- Support, synergy, and symbiosis differ in who controls or contributes to work, with symbiosis complicating accountability but not necessarily reducing autonomous competence.Synergy here means an interactive configuration, not proof that the combination outperforms its components.Source: From mechanisms to testable… · p. 12, Table 4 · p. 12
- Five proposed design levers cover task-by-task configuration, sequencing human attempts before AI, setting friction, reviewing how AI proposals are transformed, and monitoring unassisted competence and organizational indicators.The authors frame these as proposals to validate; friction may increase load or abandonment, and sequencing may slow work where competence is not being built.Source: From mechanisms to testable… · pp. 12–13
Implications for organizations
Organizations are urged to monitor what people and the organization can still do, alongside human–AI system performance. Talent development warrants particular attention because immediate benefits and risks to skill formation can coexist.
Implications for organizations
- Organizations rarely track whether people’s competence and shared knowledge strengthen or weaken with AI, so monitoring should include capability when systems change or fail.The proposed distinction separates individual and organizational capability from the performance of the human–AI system.Source: Implications for organizations · p. 13
- Talent development requires explicit attention because less experienced workers receive the largest immediate benefits while skill formation may be vulnerable during acquisition.The review also notes evidence that exposure to capable generative AI may weaken motivation to create independently, and calls for shared organizational responsibility to assess whether adoption builds or erodes cognitive capital.Source: Implications for organizations · p. 13
Limitations and research agenda
The review’s evidence base has methodological and scope limits, and its research agenda calls for validated measures and longitudinal, multilevel workplace studies. The proposed mechanisms and design options remain to be tested.
Limitations and research agenda
- The integrative rather than systematic review may have selection bias, and much experimental evidence comes from short tasks, education, or single occupations.Immediate post-withdrawal effects may not indicate durable change; some evidence is non-generative, observational studies cannot establish causality, and organizational effects are inferred from individual-level studies.Source: Limitations and research agenda · p. 13
- Four priorities are validated measures of the three capability objects, longitudinal field studies, workplace tests of the propositions, and multilevel studies of aggregation and resilience.The agenda includes examining solution diversity and organizational resilience when systems fail.Source: Limitations and research agenda · p. 13
Conclusions
The review reframes AI adoption around what human and organizational capabilities remain, rather than productivity alone. It offers cognitive capital as a multilevel lens and turns the concern into testable hypotheses, not a validated theory.
Conclusions
- Cognitive capital distinguishes retained human and organizational capability from human–AI system performance, which can improve even as human capacities weaken.Assisted output alone does not establish retained competence; AI effects depend on design and use, and individual gains do not automatically become organizational assets.Source: Conclusions · p. 14
- Cognitive capital connects learning science, organizational research, and extended-mind theory around what remains available when systems change, fail, or are replaced.The review argues that fluent outputs can invite acceptance without the effort through which competence develops and judgment is calibrated.Source: Conclusions · p. 14
- As AI agents take on more multistep work, the gap between human–AI system capability and unaided human capability may widen.The capacities needed to supervise, correct, or replace systems may be among the first to erode.Source: Conclusions · p. 14
- The review organizes its account into four mechanisms, seven principles, and testable propositions that remain hypotheses rather than a validated theory.These proposals aim to make concerns about AI and human capability measurable for organizations and researchers.Source: Conclusions · p. 14