The adoption of generative AI coding tools has been one of the fastest transitions in the history of software engineering. GitHub Copilot, launched in general availability in June 2022, reached 1.3 million paying subscribers in 2023 and 55,000 enterprise customers. The Stack Overflow Developer Survey 2024 reveals that 76% of developers use or are considering using AI tools in their daily workflow. The category went from zero to ubiquitous in fewer than 36 months.

This mass adoption occurred without direct historical precedent. Previous transitions to new development paradigms — the shift to high-level languages in the 1960s, IDE adoption in the 1990s, framework proliferation in the 2000s — unfolded over decades, allowing the academic community time to document their effects on skills. Generative AI coding tools have not benefited from that delay.

Published studies converge on immediate productivity gains. Peng et al. (2023)¹ document a 55.8% increase in completion speed on the programming tasks defined in their experimental protocol. GitHub Research (2023)² measures a 56% increase in tasks completed per hour on standardised development tasks. These measurements are methodologically sound within their scope — they capture a real and significant short-term effect.

What these studies do not capture is equally significant: none measures the effect on autonomous reasoning capacity at 12, 18 or 24 months of intensive use. The temporal horizon of available studies generally stops at 4–8 weeks. That is sufficient to measure immediate productivity. It is insufficient to detect a progressive atrophy of reasoning capacity.

Three questions remain without formalised answers in existing literature: (a) How can one empirically distinguish amplifying use from substitutive use? (b) Is there an identifiable tipping point — a moment when the relationship to the tool changes qualitatively? (c) Is autonomous code reasoning capacity recoverable after a period of structural dependency, and at what cost?

Section 2 establishes the theoretical framework and distinguishes individual cognitive lock-in from legitimate specialisation. Section 3 presents the four-phase model. Section 4 examines the available evidence base and its limitations. Section 5 sets out implications for organisations, educators and tool publishers. Section 6 proposes a research agenda.

A frontend developer who does not master the Linux kernel is specialised. This specialisation is legitimate: they have made a rational choice to concentrate their attention on a domain, and the reasoning capacity for lower layers is latent — available if the situation requires it, not exercised by choice. The operational criterion of specialisation is reversibility: if the developer decides to learn the Linux kernel, they can do so through standard learning effort. The learning capacity itself is not atrophied.

The theory of distributed cognition (Hutchins, 1995³) and the extended mind concept (Clark & Chalmers, 1998⁴) offer a framework for thinking about the integration of external tools into an individual's cognitive system. Clark & Chalmers argue that a notebook can be part of a person's cognitive system in the same way as their internal memory. This extension is cognitive but non-atrophying — removing the notebook reduces performance without degrading internal reasoning capacity.

The theory of cognitive atrophy through disuse (Salthouse, 1991⁵) documents the degradation of unexercised cognitive capacities. Applied to code reasoning, it predicts that algorithmic reasoning capacity declines if it is not regularly solicited independently of the tool. The concept of cognitive offloading (Risko & Gilbert, 2016⁶) — the delegation of cognitive tasks to external artefacts — is relevant but incomplete: it does not model the irreversible degradation of the delegated capacity, only its temporary deactivation.

Individual cognitive lock-in is the case where cognitive offloading becomes irreversible — where the capacity delegated to the tool can no longer be recovered by simply deactivating the tool, but requires a relearning effort whose cost exceeds available motivation under continuous production pressure.

The structure of ICL is analogous to that of cloud vendor lock-in (Opara-Martins et al., 2016⁷): an excellent tool in nominal cases · a dependency that builds progressively without a visible warning signal · a point of no return that makes the exit cost exceed available motivation. The analogy stops there: responsibility in ICL is tripartite (publisher · organisation · individual) whereas cloud vendor lock-in places responsibility primarily on the supplier side.

Several sources converge on signals consistent with the proposed model, without constituting direct proof of it.

This working paper proposes a theoretical framework ahead of the available empirical base. This position is academically assumed and not concealed. It is consistent with the practice of working papers in cognitive science and innovation economics, where theorisation often precedes longitudinal data by necessity: phenomena with deferred effects cannot wait 24 months of data collection before being theorised.

The value of the proposed model is its falsifiability: the four phases, the associated indicators and PNR-i constitute testable hypotheses through a longitudinal protocol. If future data contradicts the model, the model must be revised. This is the condition of an honest working paper.

Algorithmic fundamentals must be maintained as a non-negotiable baseline, independent of available tools. Generative AI should be introduced as a second-level tool — after mastery of fundamentals — not as primary access to code. The pedagogical analogy is mathematics: one does not learn arithmetic with a calculator before understanding operations. The calculator amplifies an already present skill. It cannot create an absent one.

A pedagogical research direction: design evaluations that deliberately test reasoning capacity without AI tools, regularly and over the duration of training programmes. Not as a sanction of the tool, but as a measure of latent capacity.

Design mechanisms that favour understanding rather than simple acceptance of generated code are both ethically justified and commercially rational in the long term: a developer who understands the code they generate with the tool is a more durable and loyal user than a developer who depends on it structurally without understanding.

Structural cognitive dependency on a proprietary tool raises an unresolved legal question: does it fall under employment law (skills required for the position) · competition law (entry barriers created by tool dependency) · or professional training law? No existing framework explicitly covers this case. It is an open legal research territory that goes beyond the scope of this working paper.

This working paper proposes the first formalised theoretical framework for individual cognitive lock-in through generative AI coding tools. It distinguishes this phenomenon from legitimate specialisation through the criterion of reasoning capacity reversibility. It models four phases with operational indicators for each. It formalises an individual point of no return (PNR-i) analogous to the economic point of no return of cloud vendor lock-in.

The principal contribution is not empirical — the evidence base is explicitly limited — but theoretical: providing a framework that makes the phenomenon observable, measurable and therefore falsifiable. Without a framework, the early signals of drift remain invisible. With a framework, organisations and researchers can design indicators, measurement protocols and preventive interventions.