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-2026-2714
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T17:13:29.582872Z
status: published
lens: g_space
sector: education
headline: Effectiveness of intervention using artificial intelligence for students with disabilities: systematic review and meta-analysis
dek: Background Algorithmic systems do not sufficiently capture the contextual nuances of disability, thereby risking model bias and reducing ecological validity. Hence, this review aims to investigate the various pedagogical and ethical concerns that have resulted from the integration of artificial intelligence (AI) into special education. Methods We conducted a systematic review and meta-analysis to investigate the effects of intervention using AI on students with disabilities. For this purpose, 16 papers were sele…
gain_title: A meta-analysis of 16 studies of students with disabilities found AI-based interventions linked to large improvements (Hedges' g = 1.263, 95% CI 0.950–1.576), with 14 of 16 studies showing significant effects, though only one study was randomized.
problem_title: (none)
trace_subject: (none)
gain_reading: A meta-analysis of 16 studies of students with disabilities found AI-based interventions linked to large improvements (Hedges' g = 1.263, 95% CI 0.950–1.576), with 14 of 16 studies showing significant effects, though only one study was randomized.
gain_evidence: The overall ES for AI-based interventions was large | the ESs of 14 studies (87.5%) were statistically significant
problem_reading: (none)
problem_evidence: (none)
quick_read: A systematic review and meta-analysis by researchers at Jeonju University in South Korea pooled 16 studies of AI-based interventions for students with disabilities. The overall effect was large (Hedges' g = 1.263, 95% CI 0.950–1.576), and 14 of the 16 studies (87.5%) reported statistically significant improvements.

Effects differed by outcome and group. Learning (ES 1.99) and cognitive outcomes (ES 1.92) improved most, and students with intellectual disabilities, learning disabilities and ADHD did better than students with autism (ES 0.59). Social outcomes showed the smallest gain (ES 0.43). Interventions delivered by teachers (ES 1.60) and researchers (ES 1.58) outperformed those run by caregivers (ES 0.41).

The evidence is thin. Only one of the 16 studies was a randomized trial, 15 used pre-post designs, and none met What Works Clearinghouse standards without reservation. The authors say the results do not establish a causal effect, and prediction intervals were wide in most outcome domains.
limitation: Only one of the 16 included studies was a randomized trial and most used non-randomized pre-post designs, so the results do not establish a causal effect; the interventions and populations varied widely, and subgroup findings are exploratory.
tag: Evidence-backed gain
key_points: The review searched eight databases from inception to October 13, 2025, and pooled 16 English-language studies with quantitative outcomes. | The pooled effect was large (Hedges' g = 1.263, 95% CI 0.950–1.576); 14 of 16 studies were significant, 13 of them with large effects and one with a small effect. | Learning (ES 1.99) and cognitive (ES 1.92) outcomes improved most; social outcomes (ES 0.43) and self-efficacy (ES 0.75) improved least. | By disability, students with intellectual disabilities (ES 1.56) and ADHD (ES 1.92) showed larger effects than students with autism (ES 0.59). | Interventions led by teachers (ES 1.60) or researchers (ES 1.58) had larger effects than those led by caregivers (ES 0.41); the authors say this cannot be put down to implementer type alone. | Only one study was a randomized trial; 15 relied on pre-post designs, and 10 of 16 failed the What Works Clearinghouse design standards. | Learning disabilities were the most common target (five studies, 31.3%), followed by autism (four studies) and intellectual disabilities (three studies). | Only one study reported implementer training and one reported a fidelity check, so how well the interventions were delivered is largely undocumented.
rundown: The included studies covered a wide mix of AI tools: adaptive arithmetic practice, a cognitive retraining platform, ChatGPT-based reading and conversation exercises, a text-to-speech reader, wearable glasses giving emotion cues, facial-expression feedback games, a mental health chatbot and an AI-supported swimming simulation. Sample sizes ranged from 6 to 205 participants, aged 2 to 19.3 years. In the 12 studies reporting gender, 706 participants were 55.7% male and 44.3% female.

Settings were mostly schools, classrooms and computer labs (eight studies), with three studies in centers and one at home. Session length ran from 4 to 60 minutes and duration from 1 to 48 weeks, most often 3 to 12 weeks. Researchers delivered the interventions in nine studies, teachers in three and caregivers in two.

The authors judged methodological quality using What Works Clearinghouse standards: six studies met them with reservation, 10 did not, and none met them fully. Publication-bias checks, including Egger's regression (p = 0.346), found no sign of bias, and a random-effects model was used because of significant heterogeneity (Q = 171.212).

The authors recommend randomized trials, larger moderator analyses and better documentation of training and fidelity. They also suggest that social skills, particularly for students with autism, may need human mediators rather than automation alone.
sources:
- peer_reviewed | Frontiers in Public Health | https://doi.org/10.3389/fpubh.2026.1936957 | 2026-11-06
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

dda057ee8aa20f8d06cb3b791ea0c7477578f2fa8a1e41b9c097f78a4bde626f

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000