AI-mediated tools in primary English language education: a scoping review of pedagogical functions and teacher orchestration

Synthesis: Li and Zhou asked what pedagogical work AI-mediated tools do in primary English language education, what the evidence supports about their use and outcomes, and how teachers orchestrate activity before, during and after tool use. A JBI-method scoping review reported to PRISMA-ScR searched Scopus, Web of Science Core Collection and ERIC to 21 September 2026, without year limits. Of 1,278 records, 379 duplicates were removed, 899 were screened, 454 excluded, 445 sought, 437 full texts retrieved and assessed (8 unavailable), 318 excluded at full text, and 119 reports included: 72 core empirical, 31 adjacent empirical and 16 contextual. Five functions organized the corpus, and performance differences varied with task, comparison condition and outcome measure. The strongest design, a cluster randomized trial with 217 second-grade English learners, found an effect size of .38 on taught words but statistically nonsignificant effects on standardized vocabulary and reading comprehension. With no quality appraisal or risk-of-bias assessment, the review maps the direction and conditions of findings but cannot rank their certainty.

Key Findings

1. The search returned 1,278 records (644 Scopus, 507 Web of Science Core Collection, 127 ERIC); 379 duplicates were removed, 899 were screened, 454 excluded, 445 sought, 437 full texts assessed, 318 excluded at full text, and 119 reports included.
2. The 119 comprised 72 core empirical, 31 adjacent empirical and 16 contextual reports, so 103 were empirical; 89 appeared from 2023 onwards, spanning 2004 to 2026.
3. Five functions organized the corpus: oral interaction and speaking rehearsal was the largest core group at 19 reports, and teacher-facing assessment and materials design drew on 17 core and 16 adjacent reports.
4. In a classroom-cluster randomized trial with 217 second-grade English learners, an intelligent tutoring system improved taught-word knowledge over usual instruction (effect size = .38), while effects on standardized expressive vocabulary, receptive vocabulary and reading comprehension were statistically nonsignificant.
5. In a two-week single-group study of 18 fifth graders, automated pronunciation score changes were statistically nonsignificant for trained and novel sentences, while human ratings showed improved nativeness and intelligibility in both.
6. AI-generated vocabulary cards produced a stronger immediate result, but the comparison group scored higher at the two-week delayed test.

What the included studies measured

The 103 empirical reports did not measure one thing: some reported learner performance, some teacher practice, some system output, so a count of oral turns and a speaking-test score answer different questions. Where performance was measured, the measure carried the result: Baker's trial puts the advantage on the words taught and loses it on standardized comprehension, and Yoo and Ahn show the same sensitivity inside one task. Delayed findings were equally local, surviving one month in two studies and reversing by two weeks in a third.

Feedback uptake, the step from an AI comment to a change in learner behavior, appeared in two studies: writing records of feedback access and revision, and dynamic assessment recording responses to graduated prompts. Wilson and colleagues tracked 3,459 pupils in Grades 3 to 5, including 820 English learners, and found associations between language status and feedback access or revision nonsignificant after adjustment.

A map of directions, not a ranking of effects

The review's claim about what it can establish is narrower than its recommendations: without formal appraisal of methodological quality or risk of bias, the synthesis maps the direction and conditions of reported findings but cannot determine the certainty or comparative strength of evidence for particular forms of instruction. No protocol was prepared before the search, and one reviewer screened everything while a second checked roughly 25 percent: 320 of 1,278 search records, 225 of 899 deduplicated records and 110 of 437 retrieved reports. That was not fully independent duplicate screening.

The corpus further softens comparison: the 31 adjacent reports come from related settings, including mixed-level and teacher-education samples, and the 16 contextual sources are reviews kept for interpretation. Since the five functions are interpretive products whose boundaries remain open to refinement, counts such as 19 core reports describe how the literature distributes itself, not how well a function works.

Orchestration described from practice, connected by argument

The second contribution is a four-stage account of teacher work: pre-task design, in-task regulation, post-task interpretation and subsequent instructional adjustment. The stages were derived inductively from teacher actions, and the review separates direct evidence, an implemented action with an activity or time reference, from partial evidence, a retrospective account covering only part of the link.

The cases are small: a teacher selected books before 11 fifth graders used Reading Companion; teachers interpreted robot recognition failures; 12 primary EFL teachers checked generated language and cultural fit; an eight-month program with 50 teachers was followed by stronger reported self-efficacy. The authors state that individual reports support specific actions while links across stages are conceptual, and that measuring learner responses after an instructional adjustment remains undone. The framework locates where teacher judgment enters AI-mediated activity; it does not show how much that judgment changes outcomes.

What this means for practice

For instructors, the usable claim is that AI support should enter a language task, not replace one. Define the language goal and expected learner production first, then choose the function, participation arrangement and support level, and plan what pupils do when recognition fails: retry, ask a peer, or request adult help. Teacher educators can use the four stages to join preparation, implementation and review of pupils' work, without presenting the framework as evidence that any stage improves learning.

Limitations

• Three English-language databases were searched; the authors note that education, linguistics and regional databases could broaden coverage, especially for non-English research.
• The evidence is concentrated in recent Asian EFL settings with varied tasks and outcome measures, limiting comparison and application to ESL, EAL and differently resourced contexts.
• With no quality appraisal, no protocol and no fully independent duplicate screening, the synthesis cannot establish the certainty of evidence, and age and grade were recorded but not analyzed, so differences across primary grades remain untested.

Citation

Li, M., and Zhou, Y. (2026). AI-mediated tools in primary English language education: a scoping review of pedagogical functions and teacher orchestration. Frontiers in Education, 11, 1968516.