The Rented Self: How AI-Enabled Tools Decouple Performance From Becoming

Synthesis: de Barba (2026) argues that AI-enabled tools can decouple performance from becoming — learners can produce strong work without developing the capability, or the self, that producing it once required. Where cognitive offloading and decremental research describe what learners stop doing, her construct describes what they stop becoming. Learning has two outcomes: a capability, and a domain self — what a learner notices, cares about, judges by, and does when stuck — which only the work of acquisition builds. The rented self is the part of a learner's domain self whose capabilities reside in an AI-enabled tool and are available only on the provider's terms. Crossing where a capability resides with whom the learner credits for it yields four configurations and two costs: calibration costs when learners misjudge where their capability resides, and self costs whenever a capability that builds the domain self is rented, even knowingly. Ten propositions, a measurement program pairing decremental designs with an N=me approach, and implications for design, assessment and equity follow.

Key Findings

1. Two outcomes, not one. A learning episode produces a capability and a domain self; the domain self has four facets — noticing, caring, judging by, and acting when stuck — and capabilities that build it are core capabilities. Renting a capability that is not core is often efficient; renting a core one is the case the construct is about.
2. The construct crosses two dimensions. Locus is where a capability resides — internalized when the learner can perform it unaided, AI-resident when performance depends on the tool — and is indexed by the gap between assisted and independent performance on matched content. Attribution is whom the learner credits for it.
3. Crossing them gives four configurations. Owned competence (internalized, self-credited — the aim of education), unclaimed competence (internalized but credited to the tool), visible rental (AI-resident and known), and hidden rental (AI-resident but self-credited).
4. Each configuration carries a named cost. Unclaimed competence carries a calibration cost — the learner "continues to pay rent on a home they already own"; visible rental carries a self cost, because the capability gap becomes visible while the self gap cannot (a self never built leaves nothing to notice); hidden rental carries both, since an internal attribution for an externally produced success removes the signal that would prompt learning.
5. Hidden rental is the consequential case. de Barba sets it against the flawed apprenticeship, which describes learners who appear competent to others; hidden rental describes learners who appear competent to themselves.
6. Rental is not extension. Capability in a commercial tool is available on terms the learner does not set — access, price, model behavior and continuity belong to a provider — so it fails the reliable-availability condition that extended-mind accounts presuppose. Institutional licenses relocate the landlord rather than remove it, since access typically ends with enrollment.
7. The mechanisms land inside Winne's models of self-regulated learning. Noticing, monitoring and evaluative judgment are exercised by the system, and the standards by which the learner judges are borrowed rather than built; when an impasse is resolved by search, the resolution happens without the learner facing the difficulty that would have built persistence or strategy.
8. Costs vary by stage of domain learning. The paper predicts the largest self and calibration costs in acclimation, the most frequent transitions between configurations in competence, and declines in locus for previously owned capabilities concentrated in proficiency — where learners who arrive with less domain knowledge also spend longer and are least able to identify which capabilities are core.
9. Ten propositions, each with an example prediction and a disconfirmation. P1 Self cost regardless of attribution; P2 Operations, not AI use; P3 Supplied standards and evaluations; P4 Interest without knowledge growth; P5 Asymmetric calibration errors; P6 Self-sustaining hidden rental; P7 Choice versus designation; P8 Costs vary by stage; P9 Stage mismatch; P10 Provider-set standards.
10. Measurement needs two levels of evidence. Decremental research (Panadero and Broadbent, 2025) detects what disappears across groups; Winne's (2026) N=me approach detects what happens for each learner. The construct needs within-learner indices — locus, attribution, calibration and self-building — computed per capability rather than per student.

The two costs, and why accurate attribution is not enough

The account separates two failures that assessment usually blurs. A calibration cost is an error of self-understanding: learners may underclaim a capability they hold or overclaim one they do not. A self cost is not an error at all. It is incurred by renting a core capability whether or not the learner knows what they are doing, because the operations that build the domain self are never performed — standards are not applied, problems are not encountered, difficulty is not met. Accurate attribution can therefore repair calibration and cannot repair the self.

That asymmetry drives the practical split. Unclaimed competence resembles the attributional core of the impostor phenomenon, in which high achievers credit success to external causes, and AI makes the misattribution newly likely by supplying a ready external cause at the moment success occurs — a success that then does less for efficacy. Hidden rental is worse because it is invisible to the learner: there is little reason to build what you believe you already own.

A worked course design

The paper's illustrative case is a first-year history course where both students log their chatbot exchanges. The teacher designates the core in week one — weighing the reliability of a source and building an argument from evidence are to be performed by students, while formatting references and checking grammar are declared peripheral — and explains the distinction by pointing at what historians do. The course then keeps the operations with the learner (core requests return questions and hints; a draft precedes any feedback request), has students compare before receiving by rating their own paragraph against the rubric first, and returns the indices as a private profile of predicted versus actual performance on unassisted checkpoints.

Assessment is a validity problem, not a cheating problem

Neither student cheated, and a grader reading the essays could not tell which of them had done the learning — which is why de Barba follows Dawson et al. (2024) in locating AI's central problem for assessment in validity rather than misconduct, and then specifies why: a product is valid evidence of development only if the learner performed the operations that produced it. Process-based assessment shifts evidence from products to the conditions of their production — unassisted checkpoints matched to the core, drafting histories, explanations of one's own work, and the accuracy of learners' own predictions as evidence of evaluative judgment. The failure modes are named too: an audience changes what a process shows, and the resulting evidence depends on what the author calls a design of trust. Open learner models, with the scrutability principle that learners can inspect and question what a model says about them, are offered as one direction.

Equity: who pays the rent

The ownership condition makes rental an equity question, and rent is paid in three currencies that are not distributed evenly. Money: access is tiered by price, and institutional licenses equalize it only during enrollment, so after graduation rented capability persists only for those who can keep paying while owned competence goes with the graduate — "rental converts part of that promise into a subscription". The self: self costs concentrate in acclimation, where learners with less domain knowledge spend longer and are least able to tell which capabilities are core, which makes designating the core an equity measure as well as a pedagogical one. Standards: rented standards carry the model's reference points, and P10 predicts more displacement for learners whose cultural and linguistic reference points differ from the model's. It refuses a one-sided reading: for learners writing in an additional language, or for whom a disability makes certain operations costly, rental can be the condition of participation rather than a substitute for it.

What this means for practice

• Instructors. Name the core capabilities in the first week, and say why they are core in the terms of the discipline, so that renting the rest becomes a deliberate choice rather than a default.
• Instructors. Ask learners to predict and self-rate before they see AI feedback; the accuracy of those predictions is evidence of evaluative judgment, and the paper treats calibration itself as assessable.
• Assessment designers and tool developers. Put unassisted checkpoints matched to the core into the course and pair products with drafting histories or short explanations of how the work was built — a product alone is valid evidence only if the learner performed the operations behind it — and return locus, attribution and calibration to learners as their own inspectable profile.
• Administrators. Treat the designation of core capability as an equity decision, because self costs concentrate among the learners with the least prior domain knowledge, and licenses that end at graduation relocate the provider rather than removing it.
• Researchers. Test the ten propositions with within-learner designs: alternate assisted and unassisted conditions within the same person and track self-building indices, rather than comparing learners who used AI with learners who did not.

Limitations

• It is a preprint of a conceptual paper, and none of the ten propositions has been tested. Every claim in it is an argument from the literature plus a stated prediction, so the empirical status of the construct is currently unknown.
• The construct's boundary against existing accounts is contested by its own analysis. The paper situates the rented self against cognitive offloading, self-regulatory erosion, metacognitive laziness, algorithmic deskilling and the extended mind, and acknowledges the jingle-jangle risk that a new label can be attached to something already described.
• Three frameworks with incompatible criteria of loss supply its foundations. Developmental internalization treats dependence as incomplete learning, the extended mind treats reliable coupling as a legitimate cognitive system, and the paper does not resolve which criterion governs a judgment about a given learner.
• The measurement demands are heavy and untested. Within-learner indices require repeated assisted and unassisted measurement per capability plus experience sampling of predictions and attributions, and the paper supplies sample predictions rather than validated instruments.
• The stage predictions rest on prior literature rather than new evidence. The claims that self and calibration costs are largest in acclimation, that transitions cluster in competence, and that locus declines concentrate in proficiency follow from the Model of Domain Learning, not from data on AI-era learners.
• The account may not transfer evenly across providers. Its central mechanic — capability available only on the provider's terms — assumes commercial, changeable model behavior whose standards P10 leaves to future cross-cultural measurement.

Citation

de Barba, P. G. (2026). The rented self: How AI-enabled tools decouple performance from becoming [Preprint]. Zenodo.