Systematic Literature review Sunde
Description
This dataset is the reproducibility package for a systematic review of generative AI and durable learning in higher education. It documents the evidence base, screening process, classification framework and summaries used in the review. The review covers higher-education studies published or available online between 1 January 2020 and 15 September 2026. The package contains 21 primary studies reporting 4,549 participants or enrolments. Six studies were published in 2024, six in 2025 and nine in 2026. The PRISMA data show that 6,178 records were identified: 3,412 from Scopus and 2,766 from Web of Science Core Collection. After 2,105 duplicates were removed, 4,073 unique records were screened. Of these, 3,832 were excluded at title and abstract stage, leaving 241 articles. A further 220 were excluded, leaving 21 studies in the synthesis. The main exclusion reason was the absence of an eligible student learning or performance outcome, accounting for 78 cases. Other reasons included non-higher-education samples, insufficient GenAI exposure, ineligible publication types, journal-indexing failures, and unavailable full text or outcome data. The dataset applies a five-category evidence taxonomy separating AI-assisted performance from independent and durable learning. Five studies (23.81%) were classified as AI-assisted performance, six (28.57%) as immediate independent learning, five (23.81%) as delayed retention, four (19.05%) as transfer or generalisation, and one (4.76%) as metacognitive or self-regulatory evidence. Assisted performance measures outcomes while GenAI remains available; immediate learning measures unaided performance soon after use; delayed retention examines persistence over time; transfer assesses application to a novel task or context; and metacognitive evidence captures monitoring and self-regulation. Study-level records contain citation, year, sample size, research design, task or discipline, GenAI condition, evidence category and principal result. Designs include randomised trials, quasi-experiments, controlled pre/post studies, crossover trials and field experiments. Findings are heterogeneous: some studies report gains in task quality, learning, retention or transfer, while others show null, mixed or adverse effects. The package also provides 2025 Journal Impact Factor and SCImago Journal Rank data, eligibility criteria, an evidence hierarchy, reference index, source notes and a risk-of-bias framework. The framework covers methodological and GenAI-specific concerns, including AI access during assessment, assisted versus learner-generated performance, assessment timing, transfer distance, model and prompt reporting, and outcome independence. It does not provide study-level risk-of-bias judgements. Raw database exports and record-level screening files are absent. The dataset reproduces the article’s reported synthesis, tables and figures, but cannot fully reconstruct the original search and screening process from raw records.
Files
Steps to reproduce
This dataset is the reproducibility package for a systematic review of generative AI and durable learning in higher education. It documents the evidence base, screening process, classification framework and summaries used in the review. The review covers higher-education studies published or available online between 1 January 2020 and 15 September 2026. The package contains 21 primary studies reporting 4,549 participants or enrolments. Six studies were published in 2024, six in 2025 and nine in 2026. The PRISMA data show that 6,178 records were identified: 3,412 from Scopus and 2,766 from Web of Science Core Collection. After 2,105 duplicates were removed, 4,073 unique records were screened. Of these, 3,832 were excluded at title and abstract stage, leaving 241 articles. A further 220 were excluded, leaving 21 studies in the synthesis. The main exclusion reason was the absence of an eligible student learning or performance outcome, accounting for 78 cases. Other reasons included non-higher-education samples, insufficient GenAI exposure, ineligible publication types, journal-indexing failures, and unavailable full text or outcome data. The dataset applies a five-category evidence taxonomy separating AI-assisted performance from independent and durable learning. Five studies (23.81%) were classified as AI-assisted performance, six (28.57%) as immediate independent learning, five (23.81%) as delayed retention, four (19.05%) as transfer or generalisation, and one (4.76%) as metacognitive or self-regulatory evidence. Assisted performance measures outcomes while GenAI remains available; immediate learning measures unaided performance soon after use; delayed retention examines persistence over time; transfer assesses application to a novel task or context; and metacognitive evidence captures monitoring and self-regulation. Study-level records contain citation, year, sample size, research design, task or discipline, GenAI condition, evidence category and principal result. Designs include randomised trials, quasi-experiments, controlled pre/post studies, crossover trials and field experiments. Findings are heterogeneous: some studies report gains in task quality, learning, retention or transfer, while others show null, mixed or adverse effects. The package also provides 2025 Journal Impact Factor and SCImago Journal Rank data, eligibility criteria, an evidence hierarchy, reference index, source notes and a risk-of-bias framework. The framework covers methodological and GenAI-specific concerns, including AI access during assessment, assisted versus learner-generated performance, assessment timing, transfer distance, model and prompt reporting, and outcome independence. It does not provide study-level risk-of-bias judgements. Raw database exports and record-level screening files are absent. The dataset reproduces the article’s reported synthesis, tables and figures, but cannot fully reconstruct the original search and screening process from raw records.
Institutions
- Namibia University of Science and TechnologyKhomas Region, Windhoek