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-2026-2719
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T17:14:08.390896Z
status: published
lens: g_space
sector: education
headline: The Impact of AI on Learners’ Self-Efficacy: A Meta-Analysis
dek: With the rise of generative artificial intelligence, the application of AI in learning environments has received widespread attention. Although empirical studies have explored the effect of AI on self-efficacy, the results have not been consistent. This study conducted a meta-analysis on the results from 23 empirical studies on the impact of AI use on self-efficacy. These studies were published between January 2005 and February 2025 and indexed in one or more of the three major educational research databases: We…
gain_title: A meta-analysis of 23 studies found AI use in learning raised students' self-efficacy by a medium effect size of 0.758 (95% CI 0.470 to 1.045), strongest in natural science and medical subjects.
problem_title: (none)
trace_subject: (none)
gain_reading: A meta-analysis of 23 studies found AI use in learning raised students' self-efficacy by a medium effect size of 0.758 (95% CI 0.470 to 1.045), strongest in natural science and medical subjects.
gain_evidence: AI had a significant positive impact on self-efficacy in learning contexts (effect size of 0.758) | the total effect size is 0.758, and the 95% confidence intervals range from 0.470 to 1.045
problem_reading: (none)
problem_evidence: (none)
quick_read: A meta-analysis pooling 23 empirical studies published between January 2005 and February 2025 found that AI use in learning settings had a medium, significant positive effect on learners' self-efficacy, with a pooled effect size of 0.758 (95% CI 0.470 to 1.045). Self-efficacy means a learner's belief in their ability to do a specific task.

The size of the effect depended on the subject and on the role AI played. Natural science (1.310) and medicine (1.013) showed the largest effects, while engineering showed no significant effect. AI as an intelligent learning tool had the highest effect (0.883); AI used in a mixed tutor-and-tool role had a lower one (0.450).

Results varied widely between studies: 16 effect values were positive and significant, 2 were negative and significant, and 5 were not significant. Only university students showed a significant effect, with K-12 and other learners not reaching significance, and only one included study was set online.
limitation: Only 23 studies were included, with relatively small samples, some subgroups had fewer than five observations, and heterogeneity was very high, so unmeasured factors may explain much of the variation.
tag: Evidence-backed gain
key_points: The analysis screened 1,433 articles from Web of Science, Scopus and ERIC, and 23 studies met the inclusion criteria. | Pooled effect size was 0.758 (95% CI 0.470 to 1.045); across the 23 effect values, 16 were positive and significant, 2 negative and significant, and 5 not significant. | Discipline significantly moderated the effect (Q = 10.348, p = 0.035): natural sciences 1.310, medicine 1.013, engineering not significant (p = 0.894). | AI's role also mattered (Q = 3.991, p < 0.05): intelligent learning tools scored 0.883, mixed tool-and-tutor roles 0.450. | University students showed a significant effect of 0.813, while K-12 and other learners did not reach significance, though the difference between learner levels was not statistically significant. | Research setting, type of AI and study duration did not significantly moderate the effect, and all durations showed positive effects. | Leave-one-out checks kept the pooled effect between 0.598 and 0.815, and fail-safe N and Egger's test (p = 0.073) did not point to substantial publication bias.
rundown: The authors searched three education databases using terms such as AI, chatbot, intelligent tutoring system and machine learning combined with self-efficacy. After removing duplicates and screening titles, abstracts and full texts, 23 studies remained. Two coders independently coded the studies with 86.96% agreement and rated quality, with every study scoring 11 to 15 points.

They used Hedges' g with a random-effects model because heterogeneity was very high (Q = 213.115, I2 = 89.677). Moderator analyses covered learner level, research setting, discipline, type of AI, AI role and study duration. Only discipline and AI role produced significant differences between subgroups.

The authors suggest engineering results may differ because AI gives direct answers, which could hinder the internalisation that builds self-efficacy. They recommend educators design activities that promote student-AI interaction and account for differences between disciplines, and they call for more experimental research, noting few experimental studies exist.
sources:
- peer_reviewed | Behavioral Sciences | https://doi.org/10.3390/bs16010158 | 2026-01-22
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

3a1932fced985fc95e59f84d2db6284261a97888ce801554a3412b322ae82cde

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000