The review followed JBI scoping-review methodology and reporting guidance for PRISMA-ScR, with searches of Scopus, the Web of Science Core Collection and ERIC completed through 21 September 2026. The initial search produced 1,278 records. After removal of 379 duplicates, 899 unique records were screened, 437 full texts were assessed, and the final synthesis retained 119 reports published from 2004 to 2026. Seventy-two were classified as core empirical reports directly involving primary English learners, their teachers or purpose-built classroom systems; 31 were adjacent empirical reports from related educational settings; and 16 were reviews or conceptual sources used for interpretation. Eighty-nine of the included reports appeared from 2023 onward, while the empirical evidence was heavily concentrated in Asian English-as-a-foreign-language environments.

Rather than group tools simply by their technical labels, the researchers classified studies by the pedagogical work performed. Five recurring functions emerged: oral interaction and speaking rehearsal; pronunciation and oral-reading feedback; writing, vocabulary and feedback scaffolding; reading and multimodal meaning-making; and teacher-facing assessment and materials design. Oral interaction was the largest core category, represented by 19 reports, while teacher-facing work included 17 core and 16 adjacent empirical reports. This functional approach shows that very different systems, including speech-recognition applications, educational robots, adaptive tutors and generative models, can occupy the same instructional role. Technology features therefore do not, by themselves, determine the educational function a tool serves.

Speaking applications provided repeated opportunities for oral production, interactional repair and rehearsal. In one primary classroom study, pupils using a chatbot obtained higher speaking-test scores and reported greater willingness to communicate than those receiving conventional instruction. Individual and paired chatbot arrangements both produced favorable speaking results relative to conventional teaching, although the difference between the two AI arrangements was not statistically significant. Their participation patterns differed: individual use supported self-paced rehearsal, whereas paired use generated peer negotiation and shared problem-solving. Speech-recognition failures repeatedly interrupted conversations, but those breakdowns did not simply end the learning activity. Teachers and peers modeled pronunciation, reformulated requests, interpreted unclear responses and supported another attempt, showing how a technical limitation became part of locally focused language instruction.

Pronunciation and oral-reading systems commonly structured learning through cycles of production, system response and another attempt. Reading Progress and Alexa interventions reported higher oral-reading post-test scores than comparison instruction, while a controlled study found that both computer-assisted pronunciation training and teacher-led practice produced improvement, with no statistically significant difference between the groups. One pronunciation difference remained detectable after one month, but the scope of that result was tied to its particular intervention and follow-up interval. Measurement method also mattered. In a study of 18 fifth-grade pupils, automated scores did not change significantly for trained or novel sentences, whereas human raters judged improvements in nativeness and intelligibility for both. Performance on one oral task also did not always change in parallel with performance on another, showing why automated scores cannot be treated as interchangeable with human judgments of spoken English.

Writing and vocabulary studies placed AI within sequences of prompting, production, feedback and revision. A five-week intervention combining large language model support with explicit writing strategies produced higher writing performance than its comparison group, and the authors reported that most advantages remained at a one-month follow-up. Chatbot-assisted dynamic assessment yielded higher immediate and delayed vocabulary scores than non-dynamic assessment and control conditions, with graduated prompts prompting repeated responses that revealed how much mediation each learner required. Other results were more conditional. AI-generated vocabulary cards produced a stronger immediate result, but the comparison group scored higher two weeks later. In a classroom-cluster randomised trial involving 217 second-grade English learners, an intelligent tutoring system improved knowledge of taught words relative to usual instruction, with an effect size of 0.38, yet effects on standardised expressive vocabulary, receptive vocabulary and reading comprehension were not statistically significant.

AI-supported reading combined adaptive applications, conversational agents and multimodal environments, but the measured benefits again depended on what was tested. An eight-week metaverse study with fifth-grade pupils produced higher reported reading interest and flow than silent extensive reading, while a four-week mobile-AI intervention produced higher reading post-test performance than conventional instruction. In adjacent evidence from 67 children aged five to eight, an AI-guided e-book group scored higher on immediate and delayed comprehension, but vocabulary and retelling scores were similar across conditions. Shared-reading studies also showed that adult intervention mattered: parents stepped in during recognition problems and pauses, and children’s proficiency, response timing and conversational-agent prompts shaped how dialogue proceeded. These findings position AI reading tools as elements of a social activity rather than isolated delivery systems.

Teacher-facing applications concentrated on generating and revising instructional materials, preparing assessments and interpreting automated or AI-generated feedback. Primary English teachers reported using generative AI to develop lesson ideas, textual and visual resources and assessment items, but they also checked language accuracy, age suitability, curricular relevance and cultural fit before use. In a design task involving 12 teachers, participants corrected generated language and cultural content and revised instructions for their pupils, treating AI output as draft material requiring professional review. Automated and generative feedback was more mixed: one study found that ChatGPT identified more errors and provided more detailed comments than educators on English essays, while feedback on Greek essays included false error flags. Automated and human writing scores had similar predictive validity for English learners in another study, although both were less closely related to external writing-test performance for this group than for other pupils.

Across studies, teacher orchestration emerged as a four-stage process. During pre-task design, teachers selected and adapted materials, set language demands, prepared prompts and arranged individual or collaborative participation. During in-task regulation, they responded to recognition failures, learner difficulty, unsuitable output and unequal participation, often by modeling language or guiding another attempt. Post-task interpretation required them to decide what information from learner production and system output actually meant. A recognition failure might reflect pronunciation, acoustics, task design or system behaviour, while generated feedback might combine useful guidance with errors. Subsequent instructional adjustment then translated that interpretation into changed tasks, feedback or support in later teaching. The framework is conceptual rather than a tested causal model: individual reports document specific teacher actions, but the links between stages are derived from cross-study analysis.

For primary classrooms, the evidence supports choosing the tool after defining the language goal, expected learner production, participation pattern and necessary support. Tasks should include a response route for system failure, allowing pupils to retry, seek peer assistance or request adult help without losing sight of the language purpose. AI literacy can be embedded in ordinary English lessons by asking pupils to judge whether outputs fit the task, protect personal information and seek human assistance when responses are uncertain. Teacher preparation should likewise include output review, classroom troubleshooting and reflection after implementation. These implications remain bounded by the evidence: the review searched only three databases, concentrated on Asian EFL settings, did not formally assess methodological quality or risk of bias, and included substantial contextual evidence. It mapped reported outcomes without determining the certainty or comparative strength of evidence for any particular AI intervention. The central result is therefore not that AI reliably improves primary English learning, but that carefully designed language tasks, supported use and professional judgement are conditions under which its contribution may vary.

Subject of Research: Scoping review of AI-mediated pedagogical functions and teacher orchestration in primary English language education

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

Article References: Li, M., & Zhou, Y. (2026). AI-mediated tools in primary English language education: a scoping review of pedagogical functions and teacher orchestration. Frontiers in Education, 11, Article 1968516. https://doi.org/10.3389/feduc.2026.1968516

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1968516

Keywords: primary English education, artificial intelligence, scoping review, pedagogical functions, teacher orchestration, AI literacy, speech recognition, writing feedback, adaptive tutoring, generative AI, AI-mediated, tools