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-2026-2803
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T17:25:58.084124Z
status: published
lens: g_space
sector: education
headline: Artificial intelligence-assisted vocabulary teaching: a meta-analysis study
dek: Introduction: The literature reports promising findings on the use of artificial intelligence (AI) to enhance vocabulary learning performance. However, this body of research is still relatively recent, important gaps remain, and findings vary across studies. Therefore, synthesizing the available evidence is essential to reach more generalizable conclusions and to gain deeper insight into AI-supported vocabulary instruction. Methods: This study employed a meta-analytic design to examine the effects of AI-supporte…
gain_title: A meta-analysis of 18 controlled studies with 1,823 learners found AI-supported vocabulary teaching beat control instruction on vocabulary tests, with a pooled effect of g = 1.309 (95% CI 0.876 to 1.742).
problem_title: (none)
trace_subject: (none)
gain_reading: A meta-analysis of 18 controlled studies with 1,823 learners found AI-supported vocabulary teaching beat control instruction on vocabulary tests, with a pooled effect of g = 1.309 (95% CI 0.876 to 1.742).
gain_evidence: The pooled effect size estimate for the use of AI-supported interventions in vocabulary learning success is 1.309 [95% CI (0.876, 1.742), p  < 0.05] | 18 studies yielding 23 effect sizes and involving a total of 1,823 participants were included in the analysis
problem_reading: (none)
problem_evidence: (none)
quick_read: A meta-analysis in Frontiers in Artificial Intelligence pooled 18 experimental or quasi-experimental studies, published 2015 to 2025, with 1,823 learners from preschool to university. Learners taught with AI-supported vocabulary tools scored higher than control groups, with a pooled effect of g = 1.309, which the authors call very large.

The size of the benefit depended on context. Effects were larger for second-language learning (g = 1.447) than first-language learning (g = -0.321, not significant), and in schools (g = 1.606) than outside them (g = 0.758). Against traditional teaching the effect was g = 1.633; against multimedia-supported teaching it fell to g = 0.802. Preschoolers showed no clear benefit (g = 0.029).

The studies differed greatly from one another (I² = 93.868%), so the pooled figure is an average rather than a precise estimate. Small samples, short interventions and limited reporting of randomisation were the most common weaknesses in the included studies, and nearly all second-language studies covered English.
limitation: The analysis rests on only 18 studies, shows very high heterogeneity (I² = 93.868%) not fully explained by the moderators, and drew mainly on English second-language learners and university students.
tag: Evidence-backed gain
key_points: The analysis pooled 23 effect sizes from 18 studies with 1,823 participants, drawn from Web of Science, Scopus and ERIC for 2015 to 2025. | Overall effect: g = 1.309 (95% CI 0.876 to 1.742), with very high heterogeneity (I² = 93.868%). | Second-language studies (k = 2 in the reported subgroup counts) showed g = 1.447, while first-language studies were not significant (g = -0.321); the authors say the small number of first-language studies weakens that comparison. | By education level, primary school showed g = 1.927, university 1.548, secondary school 1.358 and preschool 0.029 (not significant). | AI beat traditional instruction by g = 1.633 but beat multimedia-supported instruction by a smaller g = 0.802. | In-school use (g = 1.606) outperformed out-of-school use (g = 0.758); teacher involvement and intervention length did not differ significantly. | Leave-one-out checks moved the pooled effect between 1.116 and 1.378; Egger's test was p = 0.070 and the fail-safe N was 2,488.
rundown: The authors searched Web of Science, Scopus and ERIC, screening 3,435 records down to 18 studies. Only experimental or quasi-experimental designs with an AI group and a non-AI control group qualified. Two researchers screened and coded independently (Cohen's kappa 0.92), and effect sizes were calculated as Hedges' g under a random-effects model.

Moderator analyses found significant differences by language type, education level, control treatment and setting. Teacher role (teacher-led g = 1.449, assistant 1.446, no teacher 0.855) and duration (long term 1.623, medium 1.159, short 1.036) did not differ significantly. The authors say the teacher result should not be read as AI replacing teachers.

The authors rate the included studies as low to moderate risk of bias using an adapted ROBINS-I approach, and note that several studies contributed multiple effect sizes, which may create dependency. Funnel plots looked broadly symmetrical, but a mild small-study effect could not be ruled out.
sources:
- peer_reviewed | Frontiers in Artificial Intelligence | https://doi.org/10.3389/frai.2026.1806700 | 2026-04-23
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

eb6c7a202bd294e08773124f38d8b9e8d949f9e46b1980bb3396de48736097d0

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000