When AI performs, what remains? An integrative review of generative AI and cognitive capital in individuals and organizations

Synthesis: Riva, La Rocca, Wiederhold and Gaggioli (2026) ask what happens over time to the knowledge and judgment that stay with people once generative AI takes over the effortful steps. Their organizing construct is cognitive capital: the accumulated, renewable and depletable stock of knowledge, capability and practice that lets a person or an organization learn, interpret, judge and adapt. The review's analytic move is to separate individual capital, organizational capital and the capability of the human-AI system, so that a gain in joint performance can be observed alongside a loss in human capability rather than hidden by it. Four mechanisms — substitution of practice, Scaffolding and feedback, metacognitive calibration, and organizational aggregation — carry seven principles and a set of falsifiable propositions. Nothing here is measured by the authors; the contribution is a frame and a measurement agenda built from studies that measured other things.

Key Findings

1. Assisted performance is not retained competence. The review's first principle separates what a person can do with the tool from what remains without it, and treats the difference as the outcome of interest rather than a caveat.
2. Productivity evidence measures the assisted condition only. The motivating example is a staggered rollout to 5,172 customer-support agents where an AI assistant raised issues resolved per hour by 15% on average — a joint-system gain that says nothing about the agents' unaided capacity.
3. Offloading a task differs from offloading the practice that builds skill. The distinction is the review's load-bearing one: using a resource to reduce demand is often adaptive, while removing the exercise that produces the skill is not.
4. The same models can support or displace learning depending on design. In a randomized crossover trial with 194 undergraduate physics students, a tutor designed on pedagogical principles produced different outcomes from a solution-providing assistant built on the same underlying capability.
5. Calibration depends on expertise. Among 319 knowledge workers, greater confidence in generative AI was associated with less self-reported critical thinking, while greater self-reported critical thinking predicted more critical evaluation of AI output.
6. Agreeableness cuts both ways. An assistant that always confirms the user builds trust while weakening belief revision; one that always challenges erodes trust and invites disuse, so sycophancy is a design parameter rather than a defect to be eliminated.
7. Individual gains can coexist with reduced collective diversity. A meta-analysis of 106 experiments found human-AI combination does not reliably beat the better of the two alone, and the review adds that output quality never reveals the process that produced it.

Why a new construct was needed

The existing vocabulary splits the phenomenon in ways that lose it. Cognitive Offloading describes the individual act of moving work to an external resource; skill formation describes the slow accumulation of capability; organizational learning describes aggregation across people. Each explains part of what happens when an assistant starts drafting the memos, and none can state the trade-off the review wants to name, because the trade-off is precisely between a joint-system gain and an individual stock.

Cognitive capital is offered as the common denominator: a stock that can grow through deliberate practice, be drawn down through disuse, and be renewed, with different dynamics at the individual and organizational levels. Treating it as a stock rather than a trait is what makes depletion a measurable hypothesis instead of a metaphor.

The four mechanisms

Substitution of practice is the mechanism that removes the exercise. Scaffolding and feedback is the mechanism that can add capability, and the review is careful to note it only does so under design conditions the literature has begun to specify. Metacognitive calibration governs whether a person notices the difference between performing with help and knowing how — the point at which confidence and competence come apart.

Organizational aggregation is the newest and thinnest. A firm can accumulate capital in shared documents, routines and review practices that outlive any individual, and it can also let a uniform assistant homogenize the drafts its staff produce. The review's claim that individual gains can coexist with reduced collective diversity rests on this mechanism, and it is the one with the least direct evidence.

From frame to falsifiable claims

The propositions are the part a research program can use. Each mechanism yields predictions designed to be tested rather than asserted, and the review labels them separately (P1 to P4c) to keep them from being read as findings.

Five design levers follow for organizations, presented explicitly as proposals to be validated under stated conditions: choosing the collaboration configuration task by task, and weighing expected benefit against risks and costs, is the first. The review is honest that its levers are reasoned from mechanism rather than demonstrated by trial, and that no organization has yet run the longitudinal, multilevel measurement its own frame requires.

What this means for practice

• Instructors. Design the unaided task as part of the course rather than as an audit at the end. If the skill is built by the exercise the assistant performs, the assessment that omits the assistant is the only one that can detect the substitution.
• Instructional designers. Choose the configuration per task. The review's first lever is not a general preference for friction but a task-by-task judgment about whether the exercise is the point or the obstacle.
• Administrators. Measure the withdrawn condition before claiming a productivity gain. The 15% figure is real and describes a joint system; whether the workforce can still do the work unaided is a separate measurement nobody in the reviewed literature ran.
• Researchers. Treat cognitive capital as a stock to be measured longitudinally. The frame's own conclusion is that single-timepoint performance studies cannot test it, because depletion and renewal are time-dependent.

Limitations

• The paper is an integrative narrative review with no data, sample or measured outcome of its own; its four mechanisms are interpretive syntheses of studies that measured other constructs.
• The organizational mechanism rests on the thinnest evidence in the review, and the claim that individual gains can reduce collective diversity is supported by argument and analogy rather than by a direct organizational study.
• The five design levers are proposals the authors explicitly ask to have validated, not established practice, and none has been tested in a controlled organizational setting.
• The evidence base is drawn largely from knowledge work and higher education; no claim is made about schooling, and the construct has no validated instrument.
• The review's selection of literature is not systematic: there is no registered protocol, no search-date bound and no quality appraisal, so the seven principles are an expert synthesis rather than an estimate.

Citation

Riva, G., La Rocca, S., Wiederhold, B. K., & Gaggioli, A. (2026). When AI performs, what remains? An integrative review of generative AI and cognitive capital in individuals and organizations. arXiv:2610.38647.