The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2704
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T17:11:58.096335Z
status: published
lens: g_space
sector: education
headline: A meta-analysis of the effects of generative artificial intelligence on learning, cognitive, and linguistic outcomes
dek: This meta-analysis examined the effect of generative artificial intelligence (GenAI) on students' cognitive, learning, and linguistic outcomes. A total of 68 effect sizes extracted from 58 independent primary studies were analyzed using a three-level random-effects model with cluster-robust variance estimation applied at the study level. Seven moderator variables were evaluated: outcome domain, educational level, GenAI tool type, experimental design, cultural orientation, course/subject domain, and comparison-co…
gain_title: A meta-analysis of 58 studies covering 6,941 learners found generative AI use was associated with a large average improvement in learning, cognitive and linguistic outcomes (pooled g = 0.864).
problem_title: (none)
trace_subject: (none)
gain_reading: A meta-analysis of 58 studies covering 6,941 learners found generative AI use was associated with a large average improvement in learning, cognitive and linguistic outcomes (pooled g = 0.864).
gain_evidence: The pooled effect was large, g = 0.864, 95% CI [0.720, 1.009]
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers pooled 68 effect sizes from 58 experimental or quasi-experimental studies, covering 6,941 learners, and found that generative AI support was linked to a large average gain (Hedges' g = 0.864, 95% CI 0.720 to 1.009) across learning, cognitive and linguistic outcomes. Most of the evidence came from university students.

The average hides large differences between studies. Heterogeneity was about 88%, and the 95% prediction interval ran from −0.213 to 1.942, so a new classroom could see no benefit or a negative effect. Tests for small-study bias suggested the true effect could be as low as g of roughly 0.26 to 0.45.

The pattern differed by outcome. Language outcomes showed the biggest gains (g = 1.102) and cognitive outcomes the smallest (g = 0.535). Most of the other differences between subgroups, including educational level and culture, did not hold up after statistical correction.
limitation: The prediction interval of −0.213 to 1.942 means a new setting could show no or negative effect, and tests for small-study bias point to an inflated average; most subgroup differences did not hold after correction.
tag: Evidence-backed gain
key_points: The analysis pooled 68 effect sizes from 58 studies with 6,941 participants; 73.5% of effect sizes came from higher education or adult learners and 10.3% from primary-age children. | Effect sizes by outcome: linguistic g = 1.102 (23 effect sizes), learning g = 0.861 (29) and cognitive g = 0.535 (16); the difference was significant before but not after Holm correction (adjusted p = .062). | Primary-age learners showed g = 1.311 against 0.855 for higher education and 0.612 for secondary students, but only seven effect sizes covered primary learners and the difference lost significance after correction (adjusted p = .233). | Studies from vertical-collectivist settings (g = 0.926, 61 effect sizes) showed larger effects than horizontal-individualist ones (g = 0.454, 6 effect sizes), a difference that did not survive Holm correction (adjusted p = .233). | Gains were larger against passive baselines such as business-as-usual teaching (g = 1.014) than against active alternatives (g = 0.765), though this difference was not statistically significant (p = .161). | ChatGPT-based tools accounted for 40 of 68 effect sizes (g = 0.959) against 0.739 for other or multi-model tools, a difference that was not significant. | Egger's test indicated funnel-plot asymmetry (p = .002); PET and PEESE corrections gave g = 0.258 and 0.450, and a selection model gave 0.361 with a confidence interval that included zero. | Removing studies one at a time moved the pooled estimate only between g = 0.834 and g = 0.888.
rundown: The authors searched Web of Science and Scopus in May 2026, starting from 3,254 records and 2,925 after removing duplicates. They kept peer-reviewed English-language experimental or quasi-experimental studies that reported enough statistics to calculate an effect size, leaving 58 studies.

They modelled effect sizes nested within studies using a three-level random-effects model. Positive values meant outcomes favoring the AI condition, and reduced cognitive load counted as positive. Eleven effect sizes came from one-group pre-post designs and 57 from independent-group comparisons; restricting the analysis to the group comparisons gave g = 0.882.

Outcome domain and educational level reached conventional significance before correction for multiple tests, but neither survived it. Cultural orientation remained significant in two of three adjusted models (p = .032 and .041), though the third model, adding experimental design and control type, did not (p = .115) and had few residual degrees of freedom.

The countries in the dataset skewed heavily: China supplied 22 effect sizes and Taiwan 9, while only 6 effect sizes came from horizontal-individualist settings such as Germany, the USA and Norway. The authors could not fit a model combining culture, educational level and subject because those studies had no primary-level or STEM samples.
sources:
- peer_reviewed | Frontiers in Education | https://doi.org/10.3389/feduc.2026.1971476 | 2026-10-21
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

016f9583b88f882cdf380d72bdf43c23e5a4b37a6323849ced29461399e898c3

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000