The AI Writes the Code, the Student Writes the Model: A Theory and Measurement Programme for Learning by Construction with Generative AI

Synthesis: This theoretical paper develops Code-to-Learn with Generative AI (CtL-GenAI), a framework for upper-secondary students (roughly fifteen to eighteen) who build software with generative AI and are meant to learn the subject the software models (physics, chemistry, a local environmental problem) rather than to learn coding itself. Its organizing claim is that the AI may write the code while the student must write the model. The paper defines a model authorship construct (four facets, four ordered levels, observable indicators, a partial-credit measurement model), partitions the outcome space into supported performance, unsupported performance and delayed transfer (defining the assistance-removal gap), recasts verification of AI output as signal detection, audits twenty-four studies, and states six hypotheses with designs, measures and falsification criteria. For instructors it yields six design principles and a rubric for judging what the student, not the AI, contributed; for researchers, a measurement model to test.

Key Findings

1. The performance–learning dissociation. Three randomized studies (Bastani et al., 2025; Fan et al., 2025; Liu et al., 2026) found assistance raised performance while available and lowered it once withdrawn.
2. The unit of analysis is the task, not the tool. The same tool produces opposite outcomes under different task structures, so task design is predicted to moderate the assistance-removal gap.
3. The field has measured the wrong outcome. Of twenty-four audited studies, nearly all measured performance with the AI available; one reported a transfer sub-measure and none measured it after withdrawal.
4. Model authorship is a measurable construct. How far a learner, not the system, originates an artifact's domain-level content is scored across four facets and four levels, with refutable assumptions.
5. Verification is a detection problem. Separating sensitivity to flawed output from the threshold a learner sets for rejecting it splits over-reliance into two failure modes needing different responses.

The performance–learning dissociation

Watch for Generative AI assistance that raises performance now and lowers it later. Nearly a thousand Turkish high-school mathematics students with a standard GPT interface gained sharply in practice (48%) but scored about 17% worse on an exam without it; a guardrailed version removed the deficit (Bastani et al., 2025). A randomized study of 117 university writers found ChatGPT most improved essay scores while adding no knowledge, a bypassed metacognition named metacognitive laziness (Fan et al., 2025). RCTs with 1,222 participants found in-the-moment performance improved yet people became worse and less persistent once withdrawn (Liu et al., 2026). AI Literacy asks what students should understand about these systems, not how to arrange the lesson.

Theoretical foundations

What to protect. Constructionism explains why students should write code rather than converse with a chatbot. Productive failure places specification before prompting: a meta-analysis of 166 comparisons found Problem Solving-before-instruction beat instruction-first designs (mean d = 0.36; Sinha & Kapur, 2021). Cognitive load theory gives the rule delegate the incidental, protect the essential, while self-regulated learning and metacognitive monitoring reframe prompting and revising. The ICAP model ranks engagement from passive to interactive, and scaffolding theory (Vygotsky's zone of proximal development) requires help calibrated and withdrawn.

Model authorship and the outcome space

Ask who wrote the model, not who wrote the code. Model authorship concerns domain-level content (code-complete work with no view of the physics scores low) and is a property of a task, not a trait. Four facets (specification, conceptual model, verification, interpretation) score at four ordered levels: delegated, shepherded, co-constructed, authored. Whether a natural-language specification is modeling is answered by moving precision to a validation phase; novices explaining generated code succeeded on only a third of tasks (Nguyen et al., 2025). Competence is read under three regimes (supported, unsupported immediately after, delayed transfer at least two weeks later); the assistance-removal gap is their difference, predicted to shrink as model authorship rises.

Verification, design principles, and the evidence audit

Design for withdrawal. Six principles operationalize it: problem or concept first, student as specifier, glass-box prompting, verification against the world (only where prior knowledge lets students tell correct from flawed output apart), productive friction, and scaffold fading with reflection. They run through five phases (Frame, Specify, Co-construct, Validate, Reflect and Share), as in a ninth-grade hydrangea-color Simulation the model coded. The audit of twenty-four studies (2021–2026, Scopus) is gap-locating, not a systematic review; its dominant research question has been the tool: the same setup lowered achievement and Self-Efficacy (Yang et al., 2025) while error-reflection modules raised higher-order thinking and a Socratic chatbot beat a no-chatbot control (Zhang et al., 2025; Kholis et al., 2026).

What this means for practice

• Instructors. Move Assessment off the artifact: collect specifications, predict-then-test records, oral defenses, or process portfolios.
• Instructors. Require a verifiable specification in the learner's own words before any code, and treat candidate specifications as a dosage variable.
• Instructors. Judge artifacts against the domain, not whether they run: a Benchmark had generative AI answering secondary physics items at over 80% accuracy while sounding plausible when wrong.
• Designers. Withhold the AI at set moments and shrink its role across the unit; support that never recedes produces dependence, not competence.
• Administrators. Fund the role shift, which makes prompt-engineering skill a precondition, and confirm students can already tell correct from flawed output.

Limitations

• The paper is a theoretical proposal: none of its six hypotheses has been tested as stated, and its partial-credit model is specified but not calibrated (two raters, 20% of cases, kappa ≥ 0.70).
• The audit of 24 studies (2021–2026) is gap-locating, not a systematic review: one database (Scopus), conference proceedings excluded, 20% double-screened, no quality appraisal.
• The dissociation rests on three randomized studies (nearly a thousand Turkish mathematics students, Bastani et al., 2025; 117 writers, Fan et al., 2025; 1,222 participants, Liu et al., 2026), none using construction tasks.
• The author's boundary conditions narrow the scope: domains must supply an external criterion (physics, chemistry, ecology and epidemiology do; history, literature and ethics largely do not).

Citation

Gousopoulos, D. (2026). The AI writes the code, the student writes the model: A theory and measurement programme for learning by construction with generative AI. Preprints.org.