Preregistration: Purposes of Generative AI Use and Computational Problem-Solving Process — A 62-Country Analysis of PISA 2025 Learning-in-a-Digital-World Log Data

Abstract

Preregistration for a confirmatory analysis of PISA 2025 Learning-in-a-Digital-World (LDW) log data, deposited before the researcher had access to any unmasked estimate for the confirmatory sample. RESEARCH QUESTION. Are the purposes for which 15-year-olds use generative AI for schoolwork (summarising, preliminary research, drafting text, helping them learn) associated with process engagement, transfer, and disengagement in an interactive digital learning task, once prior ability is controlled? The focal hypothesis is a null-after-adjustment hypothesis, tested by equivalence testing rather than by null-hypothesis significance testing. CONFIRMATORY SAMPLE (fixed at deposit). 62 countries and economies, 181,378 LDW respondents, 155,454 analysable students, 362,678 student-by-unit observations, covering 78.8% of the weighted population of LDW-administering countries. The country list is deposited and will not be revised. WHY THIS DEPOSIT EXISTS. The study uses openly available secondary data, so anyone could inspect the results before formulating hypotheses. The only credible evidence that the analysis plan predates the results is a public, timestamped deposit. At the time of this deposit the six focal predictors (ST438Q01-Q04, AIUSESCH, ST436Q16DA) had been randomly permuted within country in every analysis run, so no true predictor-outcome association had been seen for any of the 62 confirmatory countries. An exploratory pilot on three countries (Albania, Azerbaijan, Argentina) was conducted and seen beforehand; those three countries are excluded from the confirmatory sample. This is disclosed in full in section 8 of the deposited document and in field 18 of the deposited form answers. CONTENTS. (1) Registration document and form answers, in PDF and Markdown. (2) Research plan and three versions of the model specification, with the revision history and the point at which each revision was made. (3) Four data audit reports. (4) Masked dry-run outputs: 294 model estimates, random-effects meta-analysis across 62 countries, a 252-specification validation, a 12-seed permutation diagnostic, the country readiness table, and the verification of which core domain each booklet administered. (5) Complete analysis code, implemented in numpy, pandas and scipy, with self-tests. (6) SHA-256 manifest of every file. The analysis code will not be modified after this deposit; any post-deposit change will be reported with old and new hashes.