1. Introduction and problem

In the 3-to-6 years band — kindergarten, preschool, CENDI, infant school — a package of three artefacts now claims to count as mediation of oral language and narration. The first is a generated story: a text with characters, conflict and a “closing” that a model produces from a prompt. The second is a chatbot that “converses” with the child: an agent that reads, asks questions and gives feedback in a machine turn. The third is an app that corrects speech: a recogniser that transcribes, scores pronunciation and declares “progress.” All three are visible and cheap. They allow a centre to exhibit that it “already works on orality with artificial intelligence.” The enunciative leap is enormous: one moves from generating an output to affirming that there is oral-language mediation. That leap is authorized neither by how a 3- to 6-year-old talks with an agent nor by what early childhood frameworks call oral language.

The thesis of this article is restrictive. A model-generated story, a chatbot that “converses” with the child or an app that corrects speech does not constitute oral-language mediation. Oral language in early childhood education is situated dialogue, listening and expansion by the adult; not a model output. Aistear, updated in 2024, locates the Communicating theme in sharing experiences, thoughts and interests with growing competence, and states that emergent literacies rest on a strong oral-language foundation (Government of Ireland, 2024). NAEYC asks for frequent, extended conversations and uses observation both for planning and for moment-to-moment interactions (NAEYC, 2020). The input that best predicts language learning is interactionally supportive, linguistically adapted and conceptually challenging for the child’s level (Rowe & Snow, 2020). None of those frameworks authorizes treating a generated text, a chatbot turn or a scored transcript as the act of mediating speech.

The problem is sharpened by a professional reason that product sheets do not mention. In early childhood education, the object is not an exhibit-ready story or a “fluency” metric. It is process quality: everyday interactions with adults, peers, materials and space (OECD, 2021). A model does not listen to this child, in this room, with this half-formed utterance. It produces a plausible output for a generic speaker. Inference, marked as such: the market for “AI for orality” inherits a category error — confusing producing a text or a correction with mediating speech — and automates it where oral language is, by definition, situated dialogue.

This paper does not recycle axes already treated in this series. The question is one of pedagogical category: what counts as mediation of oral language and narration when an early childhood centre “does AI.” The contributions are three: separate model output from dialogue, listening and expansion; examine three families of cases; and offer four tests for deciding when a kindergarten mediates speech, and not only produces a story, a chat or a correction.

2. State of the art: from model output to orality as situated craft

Four strata that the “AI for orality” market usually mixes should be kept apart. The first is what early childhood education calls oral language: adult language practices, dialogic reading with expansion, and process quality as interaction. The second is evidence on agents, chatbots and AI narrative support: almost all of it measured as comprehension, vocabulary or engagement during the session, not as sustained mediation of this group. The third is automatic recognition of child speech: the first act of any app that “corrects.” The fourth is the rights and process-quality framework, which treats the 3–6-year-old as a subject of human oversight and the educator as the one who listens and expands.

In the early childhood orality stratum, Jiang, Kaplan and Ko-Wong (2025) meta-analyse 118 articles from 104 studies, with 411 effect sizes, 13,572 teachers and 112,533 children from preschool to third grade: the quality of teachers’ language practices is positively associated with children’s language development. Hadley, Barnes, Wiernik and Raghavan (2022) group, across 30 studies and 539 correlations, eleven types of teacher talk into two registers: emergent academic and bridge. Hadley, Barnes and Hwang (2023) synthesise 54 studies into four strands (conceptual, interactive, linguistic, management or literal); conceptual talk in shared reading and elicitations in small groups are associated with oral growth. Houen, Thorpe, van Os, Westwood, Toon and Staton (2022), in 33 studies with 2- to 5-year-olds, find that questions do not guarantee back-and-forth conversation: rich conversations depend on ongoing responsiveness. Dong, Chow, Mo, Miao and Zheng (2024), with 364 kindergarteners over twelve weeks, show that PEER (prompt, evaluate, expand, repeat) outperforms prompts with minimal feedback also on expressive vocabulary and interest. Weadman, Serry and Snow (2023) observe, among Australian early childhood teachers, wide variability and a predominance of closed prompts. Inference: mediating oral language is not asking questions or producing a text. It is hearing this utterance and expanding it.

In the conversational and narrative AI stratum, Xu, Wang, Collins, Lee and Warschauer (2021) compare, among 90 three- to six-year-olds, reading with an agent and with an adult: the agent matched comprehension, but the adult elicited more production, more lexical diversity and more topical relevance; the agent favoured intelligibility. Xu, Aubele, Vigil, Bustamante, Kim and Warschauer (2022), with 117 three- to six-year-olds, find that the agent replicates the dialogic benefit for comprehension by increasing narrative-relevant vocalisations. Cheng, Yin, Lin, Shi, Zheng, Zhu, Liu, Chen and Dong (2024), with 148 Beijing kindergarteners (M = 70.07 months), find comparable support for comprehension and word learning, moderated by prior language proficiency. Xiao, Zou, Lin, Li and Yang (2025), with 67 five- to eight-year-olds, find better comprehension with the agent, but more narrative-relevant vocalisations with the adult. Zhang, Xu, Wei and Wang (2026) assign 180 children (M = 78.91 months) to three narrative-support conditions: the iterative condition wins during the phase with the system and differences disappear at transfer. He, Gastón-Panthaki, Zhuo, Munsey, Zhang and Warschauer (2025) evaluate StoryPal with 23 four- to seven-year-olds in one session: high engagement; adults asked for a complement, not a replacement. Xu, Branham, Deng, Collins and Warschauer (2021) find that most voice apps do not sustain open-ended dialogue. Xu (2023) concludes that there is comprehension, not equivalence of communicative pattern. Sun, Tan and Lim (2025) and Zhang, Husnin and Kamsin (2026) map the field: benefit concentrates on vocabulary, comprehension and participation, and AI works best under educator direction, not as a substitute. Chen (2024) covers 18 articles and 15,081 two- to eight-year-olds; Su and Yang (2022), 17 studies; Ljungcrantz (2026), 39 studies from 2020–2024. Status: field finding, not everyday mediation. Inference: “more chatbot” is not “more oral language.”

In the recogniser stratum, Kim, Druga, Esmaeili, Woodward, Shaw, Jain, Langham, Hollingshead, Lovato, Beneteau, Ruiz, Anthony and Hiniker (2022) observe 28 five- to ten-year-olds with a commercial assistant: transcription exceeded 84% in unscripted conversation, but the device responded relevantly to content only half the time. Pelfrey et al. (2024) find, in preschool classrooms, higher error rates for child speech than for adult speech. Quan et al. (2026) deploy Vovo, an LLM conversational system, for six weeks with 10 families and 3- to 7-year-olds (M = 5.4): 150 sessions; adult-led sessions produced significantly higher learning outcomes than AI-led sessions, t(9) = 11.83, p < .001, d = 3.74; speech recognition was one of the system’s limits. Inference: an app that “corrects speech” stumbles at the first act. It does not reliably hear this utterance. It cannot expand it.

In the rights and process-quality stratum, the Recommendation on the Ethics of AI requires human oversight and particular attention when children are involved (UNESCO, 2021). Miao and Holmes (2023) set a 13-year threshold for independent conversations with generative platforms. UNICEF (2021) prioritises the best interests of the child. Regulation (EU) 2024/1689 treats as high-risk, in Annex III, systems intended to determine access or admission, to assign, to assess or to determine educational level (Unión Europea, 2024). The European Commission (2022) asks not to replace professional judgement. OECD (2021) anchors quality in everyday interactions; OECD (2023, 2025) locates digitalisation and staff practices as a system lever. Inference: a five-year-old is not the user of an orality chatbot. He or she is the subject of listening that an adult sustains in the act.

3. Review method

A critical narrative review was conducted, not a meta-analysis. The purpose was not to estimate a homogeneous effect size across an agent trial, a narrative-training study and a voice assistant, but to articulate an argument of pedagogical category with verified sources. Inclusion criteria: (a) 2021–2026, with Rowe and Snow (2020) and NAEYC (2020) as the current editions of the input and developmentally appropriate practice frameworks; (b) oral language, narration, dialogic reading, conversational agents, generated stories or automatic speech recognition, or AI in early childhood education ages 3–6; (c) relevance for kindergarten, preschool, CENDI or ages 3–6; (d) peer-reviewed journal, DOI or report from NAEYC, NCCA, UNESCO, OECD, UNICEF, the European Union or the European Commission; (e) verifiable DOI or publisher page. Axes already used in this series were excluded as central objects. The recogniser’s linguistic bias is cited only as a technical limit, not as a multilingualism axis.

The search was executed on 26 August 2026 on DOI pages, Springer, Elsevier, Wiley, SAGE, Taylor & Francis, ACM, OECD iLibrary, UNESDOC, UNICEF, EUR-Lex, NAEYC and NCCA. Each source was checked against at least one of those pages. The analysis distinguished empirical finding, conceptual or normative framework, and pedagogical inference, marked as such.

4. Case 1. A chatbot that “converses” is not the adult who listens and expands

Xu et al. (2021) publish the trial that best illustrates the leap the thesis rejects. Ninety three- to six-year-olds read with a conversational agent or with an adult. The agent’s guided conversation was as supportive of comprehension as the adult’s. The communicative pattern was not the same: the adult elicited longer, more lexically diverse and more topically relevant responses; the agent favoured intelligibility. Status. Finding of comprehension and verbal pattern in one session, not of everyday mediation over a year. It is not a finding that the agent “converses” in the sense of situated dialogue. Pedagogical inference, marked as such: this is the gesture a CENDI copies when it “does AI for orality.” Story comprehension is taken as oral language, a machine turn is produced, and mediation is declared. What the trial authorizes is narrower: an agent can sustain dialogic reading well enough for comprehension not to drop in that session. It does not authorize saying that the kindergarten mediated speech. Intelligibility is not expansion. A child who clearly articulates a short answer has not been heard in lexicon or topic.

Xu et al. (2022) saturate the picture with 117 three- to six-year-olds (50% girls; 37% White, 31% Asian, 21% multi-ethnic). The agent replicated the dialogic benefit for comprehension by increasing narrative-relevant vocalisations. Status: engagement and comprehension in one session, not longitudinal oral development. Cheng et al. (2024) move the design to 148 Beijing kindergarteners (83 girls; M = 70.07 months): the chatbot was comparable on comprehension and words; the effect was moderated by prior language, not by age or reading. Status: N = 148, not inflated to every CENDI. Inference: a chatbot “works” better with those who already have more language. That is not mediation of this group. It is selection of those who can already answer the prompt.

Xiao et al. (2025) bring the contrast to a language model: 67 five- to eight-year-olds. The agent group outperformed the adult on comprehension, with comparable benefits for vocabulary and retelling. The agent produced more behavioural engagement and visual attention; the adult, more affective engagement and more narrative-relevant vocalisations. Status: ages 5–8, English as an additional language; transfer to 3–6 is inference. This is not turned into a study of parental mediation: the object is the child’s speech pattern with an adult versus an agent. Inference: if “orality” is looking at the screen, the agent wins. If “orality” is producing relevant narrative language, the adult wins. The thesis is played on the second criterion.

Dong et al. (2024) name the mechanism the chatbot does not replace. In 364 kindergarteners over twelve weeks, PEER — prompt, evaluate, expand, repeat — outperformed prompts with minimal feedback on expressive vocabulary and interest. Houen et al. (2022) recall that asking is not enough: rich conversation depends on ongoing responsiveness. Hadley et al. (2023) and Jiang et al. (2025) locate that responsiveness in teacher talk, not in a machine turn. Xu, Branham et al. (2021) show that current voice interfaces do not sustain open-ended dialogue. Inference: a chatbot that asks questions and gives pre-assembled feedback has executed the prompt. It has not evaluated this utterance or expanded it. PEER without evaluate and without expand is not dialogic reading. It is a spoken questionnaire.

5. Case 2. A generated story, or narrative support that does not transfer, is not mediated narration

Zhang et al. (2026) publish in Learning and Instruction the study that best names, in this corpus, the difference between performance with the system and narration when the system is switched off. One hundred and eighty children in North China (M = 78.91 months, SD = 7.05; 97 boys) were assigned to passive-active retelling, autonomous storytelling with AI feedback or iterative storytelling with AI feedback. Story structure, complexity and internal-state terms were measured at baseline, with support and at transfer. After controlling for baseline performance, age and duration, the iterative condition outperformed the others during the phase with the system and reported more enjoyment. At transfer, without the support, there were no significant differences among conditions. Status. Finding from narrative training; mean age slightly exceeds 6 years, so transfer to 3–6 is marked as inference. It is not a finding that “AI teaches narration.” Pedagogical inference, marked as such: a centre cannot properly call this narrative development. A child who narrates better while the model returns versions has produced a joint output. Mediation is verified when the child narrates without the system, before an adult who listens.

He et al. (2025) saturate the picture with StoryPal: an LLM agent that generates questions and feedback. Twenty-three four- to seven-year-olds, in one session, showed high verbal engagement, with distinct patterns between English-dominant and bilingual speakers. Adults asked that it not replace reading with an adult. Status: usability, N = 23, one session; not narrative development. The bilingual pattern is cited only as a technical limit, not as an axis. Inference: a kindergarten that declares “we already narrate with AI” because the agent asked questions has measured session engagement. It has not measured narration.

Xu (2023) and Xu et al. (2022) recall that agents’ documented benefit is mainly comprehension in the session, not a narration of one’s own. Rowe and Snow (2020) require three dimensions at once: interactive, linguistic and conceptual. A generated story can simulate the linguistic dimension and fail the other two: there is no contingency with this child and no challenge anchored in what was just said. Aistear (Government of Ireland, 2024) asks that emergent literacies rest on oral language. NAEYC (2020) asks for extended conversations. OECD (2021) measures process quality in interaction, not in the story file. Inference: generating the story the night before and reading it to the group is not mediating narration. It is substituting the child’s voice with the model’s fluency. Early childhood narration is verified when this child tells what was seen or invented, and an adult listens and expands. A PDF with a moral “closing” is not that act.

Sun et al. (2025), Chen (2024), Su and Yang (2022) and Ljungcrantz (2026) saturate the map: the field measures vocabulary, alphabet and comprehension, not whether this child narrates tomorrow without the system. Zhang et al. (2026, IOJET) conclude that AI works better under educator direction, not as a substitute. Inference: the gap is not of products. It is of category. Output is measured because output is what a model knows how to produce.

6. Case 3. An app that “corrects speech” stumbles at hearing; hearing is not expanding

Kim et al. (2022) publish the case that best illustrates the third artefact. Twenty-eight five- to ten-year-olds interacted with a commercial voice assistant: requests, recitation and unstructured conversation. The device transcribed correctly more than 84% of the time in unscripted conversation. It responded relevantly to content only half the time. Appropriate responses rose with age and when discourse was standardised. In free conversation, the most common goal was to build a relationship with the device. Status: laboratory, ages 5–10, N = 28; transfer to 3–6 is inference, and the lower band is the most unfavourable. It is not a finding that the recogniser “corrects” speech. Pedagogical inference, marked as such: a correction app inherits this ceiling. If it hears form and does not answer sense, there is no dialogue. There is transcription. A four-year-old rehearsing a sentence does not need a score. The child needs someone to take up the sense and return it slightly expanded.

Pelfrey et al. (2024) saturate the picture in the preschool classroom: automatic processing shows higher error rates for child speech than for adult speech. Quan et al. (2026) deploy an LLM conversational system: across 150 sessions over six weeks, adult-led sessions outperformed AI-led sessions (d = 3.74); speech recognition, nonverbal cues and phoneme instruction were limits. Status: N = 10 families, ages 3–7; not generalised to CENDI and not turned into a home article. The object is that the adult hears and the system stumbles. Inference: declaring “speech correction” when the recogniser fails with 3- to 6-year-old speech is not mediation. It is a metric the child cannot contest.

Houen et al. (2022) and Hadley et al. (2022, 2023) recall what an adult would do in the same instant: respond, elicit, conceptualise. Jiang et al. (2025) associate that quality of teacher talk with language development. Miao and Holmes (2023) exclude children under 13 as independent interlocutors. A four-year-old is not the user of the correction app. Inference: if AI enters a kindergarten, the only place it does not contradict early childhood orality is the adult’s workshop — possible words, a topic — subject to pedagogical validation and discardable on hearing this child. It does not enter as listener or corrector.

The rights framework closes the case. UNESCO (2021) requires human oversight. UNICEF (2021) requires development. The European Commission (2022) requires not replacing professional judgement. Annex III of Regulation (EU) 2024/1689 includes automated assessment and determination of level (Unión Europea, 2024). Inference: transcribing, scoring and treating speech as level in a CENDI is not an orality innovation. In that territory it is a high-risk practice. Miao and Holmes’s (2023) threshold is not met by lowering the user’s age. It is respected by not turning the early childhood child into the interlocutor of a system that does not hear what matters.

7. Inferential framework: four tests for claiming oral-language mediation, not an output

The framework that follows is pedagogical inference of this article, anchored in the cases and in the verified instruments. It is not a new international standard. It distinguishes four tests. If a kindergarten, preschool, CENDI or infant school does not pass them, it cannot declare that a generated story, a chatbot that “converses” or a speech-correction app constitutes oral-language mediation.

7.1. Test of this utterance, not of the generic speaker. Xu et al. (2021) show that the adult, not the agent, elicited topical relevance and lexical diversity. Cheng et al. (2024) show that the chatbot favours those who already have more language. Rowe and Snow (2020) require linguistic adaptation and conceptual challenge for this level. NAEYC (2020) asks for conversations with these children. Inference: evidence of mediation is verified in features of this utterance — what was said, what was not reached, what was returned slightly expanded — present in the next turn. If the “evidence” that the centre mediates speech is a prompt with age and story theme, the centre has generated an output for a generic speaker. It has not mediated.

7.2. Test of listening and expansion, not of the prompt. Dong et al. (2024) show that evaluating and expanding, not only asking, mark the expressive difference. Houen et al. (2022) show that the question does not guarantee conversation. Hadley et al. (2023) and Jiang et al. (2025) locate quality in teacher talk that responds. Kim et al. (2022) show a device that transcribes and does not answer sense. Inference: a chatbot that walks a question tree does not pass the test. It may be a material. Mediation begins when someone hears and expands.

7.3. Test of transfer without the system, not of performance with the system. Zhang et al. (2026) show gains during the AI-support phase and no condition differences when support is withdrawn. He et al. (2025) measure one session. Xu et al. (2022) measure comprehension in the session. Inference: if “orality” vanishes when the agent is switched off, there was no oral-language mediation. There was co-production of an output. Early childhood narration includes this child telling tomorrow, without the model, to an adult who is still there.

7.4. Test of the adult who hears, not of the model’s fluency. Miao and Holmes (2023) exclude children under 13 as independent interlocutors. Pelfrey et al. (2024) and Quan et al. (2026) document the recogniser’s ceiling. The European Commission (2022) and UNESCO (2021) require not replacing professional judgement. OECD (2021, 2025) locates quality in staff interactions. Government of Ireland (2024) reserves Communicating for a slow relational environment. Inference: the 3–6-year-old is not the user of a generated story, a chatbot or a correction app. The child is the speaker of a morning that an adult sustains. If AI enters, it enters as the adult’s materials workshop, subject to pedagogical validation. It does not enter as interlocutor, narrator or corrector.

The framework admits generative AI as the adult’s workshop (He et al., 2025; Zhang et al., 2026, read against the thesis: a discardable input, not an interlocutor); cycles of reading, asking, evaluating and expanding (Dong et al., 2024; Houen et al., 2022); and conceptual and interactive teacher talk as a form of mediation (Hadley et al., 2023; Jiang et al., 2025). It rejects declaring oral-language mediation by a generated story, a chatbot that “converses” or an app that corrects speech (Xu et al., 2021, 2022; Cheng et al., 2024; Xiao et al., 2025; Zhang et al., 2026; Kim et al., 2022).

8. Discussion

Three tensions organise the discussion. The first is between producing an output and mediating speech. It is a finding that an agent can match or exceed an adult’s comprehension in a session (Xu et al., 2021, 2022; Cheng et al., 2024; Xiao et al., 2025). It is a framework that early childhood oral language is situated dialogue, listening and expansion (Rowe & Snow, 2020; Dong et al., 2024; Houen et al., 2022; Government of Ireland, 2024; NAEYC, 2020). It is not a finding that matching comprehension produces the expressive, topical speech the adult elicited (Xu et al., 2021; Xiao et al., 2025) or the language practices Jiang et al. (2025) associate with development. The three artefacts measure what engineering knows how to measure and declare what only situated craft would authorize.

The second is between the generic speaker and this utterance. Cheng et al. (2024) show moderation by prior language. Kim et al. (2022) show that the device responds better when the child standardises discourse and is older. Zhang et al. (2026) show support that does not transfer. Inference: insisting that “the story is already there” or “the chatbot already conversed” while no one has heard this utterance is inverted pedagogy. Rowe and Snow (2020) place contingency at the start. The model writes first and, if at all, hears later.

The third is between the adult’s workshop and the child’s interlocutor. He et al. (2025) show that a complement, not a replacement, is requested. Miao and Holmes (2023) exclude children under 13. Quan et al. (2026) show a large effect in favour of the adult. Inference: the only AI use that does not contradict early childhood orality remains on the adult’s side, with pedagogical validation and a right to silence the system. A word list can be material. A model that “converses” with a four-year-old is not mediation.

9. Limits

This review is narrative. It does not apply PRISMA or estimate combined effects. Xu et al. (2021, 2022) are reading sessions, not a kindergarten year. Cheng et al. (2024) are Chinese kindergarten, N = 148. Xiao et al. (2025) are ages 5–8 in English as an additional language; they are not used as a family or multilingualism article. Zhang et al. (2026) have a mean age of 78.91 months; transfer to 3–6 is inference; the object is not generic scaffolding, but the non-transfer of narration. He et al. (2025) are N = 23 and one session. Kim et al. (2022) are ages 5–10, N = 28. Pelfrey et al. (2024) measure transcription error. Quan et al. (2026) are 10 families; they are cited for the adult/AI contrast and the recogniser, not as a home object. Dong et al. (2024) are dialogic reading with adults, not AI. Houen et al. (2022), Hadley et al. (2022, 2023), Jiang et al. (2025), Weadman et al. (2023) and Rowe and Snow (2020) are of teacher talk and input. Chen (2024), Su and Yang (2022), Ljungcrantz (2026), Sun et al. (2025) and Zhang et al. (2026, IOJET) map AI and early language. NAEYC, Aistear, UNESCO, UNICEF, the OECD and the European Union are framework. Miao and Holmes (2023) are used for the age threshold, not for teacher competence. The recogniser’s bias toward adult speech and standardised discourse is cited as a technical limit. No Latin American trials of AI and oral-language mediation in kindergarten that measure situated expansion and transfer without the system were located. Section 7 inferences are category hypotheses, not implementation evidence.

10. Conclusions

A model-generated story, a chatbot that “converses” with the child or an app that corrects speech does not constitute oral-language mediation in an early childhood centre. The verified evidence does not authorize that declaration. Three- to six-year-olds comprehend as well or better with an agent, but the adult elicited more production, lexical diversity, topical relevance and narrative vocalisations (Xu et al., 2021, 2022; Xiao et al., 2025); a comparable chatbot is moderated by those who already have more language (Cheng et al., 2024). One hundred and eighty children narrate better with iterative AI support and cease to differ when support is withdrawn (Zhang et al., 2026); in one session, adults ask that it not replace the adult (He et al., 2025). An assistant answers sense half the time (Kim et al., 2022); the recogniser fails more with child speech (Pelfrey et al., 2024); adult-led sessions outperform AI-led sessions (Quan et al., 2026). PEER wins when the adult evaluates and expands, not when the adult only asks (Dong et al., 2024). The quality of teacher talk is associated with the language development of more than one hundred thousand children (Jiang et al., 2025). The field of AI and early language does not equate that with everyday mediation (Chen, 2024; Su & Yang, 2022; Ljungcrantz, 2026; Zhang et al., 2026). When there is mediation in early childhood education, there is an adult who hears and expands: interactive, linguistic and conceptual input (Rowe & Snow, 2020), conceptual and interactive talk (Hadley et al., 2023), ongoing responsiveness (Houen et al., 2022), oral language as a foundation (Government of Ireland, 2024) and moment-to-moment interactions (NAEYC, 2020; OECD, 2021). Current law requires human oversight, the child’s best interests, a 13-year threshold for independent conversations with generative platforms and, in the European Union, treating as high-risk AI that determines access, admission, assignment, assessment or level (UNESCO, 2021; UNICEF, 2021; Miao & Holmes, 2023; Unión Europea, 2024; European Commission, 2022).

Where the sources do not measure a kindergarten, this article does not claim it. Where they measure session comprehension or a more fluent story, it does not translate them into mediation. Accompanying three- to six-year-olds in oral language is to dialogue in situation, to hear this utterance and to expand it. The rest is a model output. It is not mediation.

Laboratorio Editorial de NEXTECH.IA / Ingeniero Mitre.

References

References

  1. Chen, J. J. (2024). A scoping study on AI affordances in early childhood education: Mapping the global landscape, identifying research gaps, and charting future research directions. Journal of Artificial Intelligence Research, 81, 701–740. https://doi.org/10.1613/jair.1.16882
  2. Cheng, X., Yin, L., Lin, C., Shi, Z., Zheng, H., Zhu, L., Liu, X., Chen, K., y Dong, R. (2024). Chatbot dialogic reading boosts comprehension for Chinese kindergarteners with higher language skills. Journal of Experimental Child Psychology, 240, 105842. https://doi.org/10.1016/j.jecp.2023.105842
  3. Dong, Y., Chow, B. W.-Y., Mo, J., Miao, X., y Zheng, H.-Y. (2024). Effects of dialogic reading elements on children’s language development. Journal of Research in Reading, 47(2), 181–200. https://doi.org/10.1111/1467-9817.12447
  4. European Commission. (2022). Ethical guidelines on the use of artificial intelligence (AI) and data in teaching and learning for educators. Publications Office of the European Union. https://doi.org/10.2766/153756
  5. Government of Ireland. (2024). Aistear: The Early Childhood Curriculum Framework. National Council for Curriculum and Assessment. https://www.curriculumonline.ie/early-childhood/aistear-2024/
  6. Hadley, E. B., Barnes, E. M., y Hwang, H. (2023). Purposes, places, and participants: A systematic review of teacher language practices and child oral language outcomes in early childhood classrooms. Early Education and Development, 34(4), 862–884. https://doi.org/10.1080/10409289.2022.2074203
  7. Hadley, E. B., Barnes, E. M., Wiernik, B. M., y Raghavan, M. (2022). A meta-analysis of teacher language practices in early childhood classrooms. Early Childhood Research Quarterly, 59, 186–202. https://doi.org/10.1016/j.ecresq.2021.12.002
  8. He, K., Gastón-Panthaki, A., Zhuo, D., Munsey, J., Zhang, M., y Warschauer, M. (2025). StoryPal: Supporting young children’s dialogic reading with large language models. En Proceedings of the 24th Interaction Design and Children (pp. 494–511). ACM. https://doi.org/10.1145/3713043.3728867
  9. Houen, S., Thorpe, K., van Os, D., Westwood, E., Toon, D., y Staton, S. (2022). Eliciting and responding to young children’s talk: A systematic review of educators’ interactional strategies that promote rich conversations with children aged 2–5 years. Educational Research Review, 37, 100473. https://doi.org/10.1016/j.edurev.2022.100473
  10. Jiang, Y., Kaplan, B., y Ko-Wong, L. (2025). Does teacher talk matter too? A meta-analysis of partial correlations between teachers’ language practices and children’s language development from preschool to third grade. Review of Educational Research. https://doi.org/10.3102/00346543251339131
  11. Kim, M. K., Druga, S., Esmaeili, S., Woodward, J., Shaw, A., Jain, A., Langham, J., Hollingshead, K., Lovato, S. B., Beneteau, E., Ruiz, J., Anthony, L., y Hiniker, A. (2022). Examining voice assistants in the context of children’s speech. International Journal of Child-Computer Interaction, 34, 100540. https://doi.org/10.1016/j.ijcci.2022.100540
  12. Ljungcrantz, L. (2026). The interaction of AI and early childhood education. A state-of-the-art review 2020–2024. Early Childhood Education Journal, 54, 3565–3581. https://doi.org/10.1007/s10643-025-02079-3
  13. Miao, F., y Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535
  14. NAEYC. (2020). Developmentally Appropriate Practice (DAP) position statement. National Association for the Education of Young Children. https://www.naeyc.org/resources/position-statements/dap
  15. OECD. (2021). Starting Strong VI: Supporting meaningful interactions in early childhood education and care. OECD Publishing. https://doi.org/10.1787/f47a06ae-en
  16. OECD. (2023). Empowering young children in the digital age (Starting Strong). OECD Publishing. https://doi.org/10.1787/50967622-en
  17. OECD. (2025). Results from TALIS Starting Strong 2024: Strengthening early childhood education and care. OECD Publishing. https://doi.org/10.1787/20af08c0-en
  18. Pelfrey, G. L., Justice, L. M., Gonzalez Villasanti, H., y Foster, T. J. (2024). Evaluating the accuracy of automated processing of child and adult language production in preschool classrooms. Frontiers in Psychology, 15, 1322665. https://doi.org/10.3389/fpsyg.2024.1322665
  19. Quan, S., Tu-Shea, X., Ding, Y., Du, Y., Zheng, Q., y Gerdich, L. E. (2026). Conversational AI in children’s home literacy learning: Effectiveness, advantages, challenges, and family perception. Computers and Education: Artificial Intelligence, 10, 100549. https://doi.org/10.1016/j.caeai.2026.100549
  20. Rowe, M. L., y Snow, C. E. (2020). Analyzing input quality along three dimensions: Interactive, linguistic, and conceptual. Journal of Child Language, 47(1), 5–21. https://doi.org/10.1017/S0305000919000655
  21. Su, J., y Yang, W. (2022). Artificial intelligence in early childhood education: A scoping review. Computers and Education: Artificial Intelligence, 3, 100049. https://doi.org/10.1016/j.caeai.2022.100049
  22. Sun, H., Tan, J., y Lim, M. C. (2025). AI and early language learning: A scoping review. AI, Brain and Child, 1, 4. https://doi.org/10.1007/s44436-025-00005-3
  23. UNESCO. (2021). Recommendation on the ethics of artificial intelligence. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000381137
  24. UNICEF. (2021). Policy guidance on AI for children 2.0. UNICEF Office of Global Insight and Policy. https://www.unicef.org/innocenti/reports/policy-guidance-ai-children
  25. Unión Europea. (2024). Reglamento (UE) 2024/1689 del Parlamento Europeo y del Consejo, de 13 de junio de 2024, por el que se establecen normas armonizadas en materia de inteligencia artificial. Diario Oficial de la Unión Europea, L 2024/1689. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  26. Weadman, T., Serry, T., y Snow, P. (2023). Oral language and emergent literacy strategies used by Australian early childhood teachers during shared book reading. Early Childhood Education Journal, 51, 1335–1348. https://doi.org/10.1007/s10643-022-01381-8
  27. Xiao, F., Zou, E. W., Lin, J., Li, Z., y Yang, D. (2025). Parent-led vs. AI-guided dialogic reading: Evidence from a randomized controlled trial in children’s e-book context. British Journal of Educational Technology, 56(5), 1784–1813. https://doi.org/10.1111/bjet.13615
  28. Xu, Y. (2023). Talking with machines: Can conversational technologies serve as children’s social partners? Child Development Perspectives, 17(1), 53–58. https://doi.org/10.1111/cdep.12475
  29. Xu, Y., Aubele, J., Vigil, V., Bustamante, A. S., Kim, Y.-S., y Warschauer, M. (2022). Dialogue with a conversational agent promotes children’s story comprehension via enhancing engagement. Child Development, 93(2), e149–e167. https://doi.org/10.1111/cdev.13708
  30. Xu, Y., Branham, S., Deng, X., Collins, P., y Warschauer, M. (2021). Are current voice interfaces designed to support children’s language development? En Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (pp. 1–12). ACM. https://doi.org/10.1145/3411764.3445271
  31. Xu, Y., Wang, D., Collins, P., Lee, H., y Warschauer, M. (2021). Same benefits, different communication patterns: Comparing children’s reading with a conversational agent vs. a human partner. Computers & Education, 161, 104059. https://doi.org/10.1016/j.compedu.2020.104059
  32. Zhang, N., Xu, J., Wei, R., y Wang, Y. (2026). From scaffolding to transfer: The impact of AI-powered support on children’s narrative skills. Learning and Instruction, 104, 102345. https://doi.org/10.1016/j.learninstruc.2026.102345
  33. Zhang, Y., Husnin, H. B., y Kamsin, I. F. B. (2026). The application of artificial intelligence technology in early childhood language development. International Online Journal of Education and Teaching, 13(2), 1–16. https://www.iojet.org/index.php/IOJET/article/view/2306