1. Introduction and problem

In the 3-to-6-year span —kindergarten, preschool, CENDI, infant school— a package of artifacts has been installed that claim to count as support for sustained shared thinking (SST): a detector/ASR/NLP/ML or multimodal system that scores SST, SSTEW auto-score, dialogic quality index, SST episodes per hour or interaction quality percentile; a GenAI of SST scripts, lesson packs, open-question banks or shared-thinking cards by prompt; a dashboard of SST minutes/hour; a chatbot/robot presented as a substitute that «does SST» without adult mediation; and automatic co-coding of dialogue converted into a ranking of teacher dialogic quality without situated judgment. The artifacts make it possible to display that the center «already does SST with AI». The leap —from score, script, count, substitute-robot or ranking to claiming support— is not authorized by the mappings of AI in ECE nor by pedagogy when there are prolonged dialogues in which adult and child co-construct meaning, with co-presence and pedagogical interpretation, not an «SST-score».

The thesis is restrictive: those artifacts do not constitute support for sustained shared thinking in early childhood. In early childhood education that practice develops with episodes in which two or more people work together in an intellectual way to solve a problem, clarify a concept, evaluate an activity or extend a thought; both parties contribute and the thinking develops (Macha et al., 2024; McLaughlin et al., 2021; Siraj et al., 2023); not an SST detector, a GenAI of scripts, a dashboard of minutes, a robot that «does SST» or a ranking of dialogic quality. AI may support the adult's preparation, documentation and reflection in a subordinate way (Jambunathan, 2025; Heinmäe et al., 2025; Meade y Kwan, 2022); it must not substitute the dialogic relationship nor convert SST into a surveillance metric. Macha et al. (2024) show SST dialogues with children aged 4–6 years that make explicit perspectives on the quality of the center: a practice of co-thinking, not an auto-score. McLaughlin et al. (2021) explore SST with data used by teaching teams: data-informed teaching, not an administrative dashboard. Szymanski et al. (2024) isolate SST speech acts in an experimental task with 4–6-year-olds: a laboratory finding, not a classroom detector. Siraj et al. (2023) publish the SSTEW Scale as a human observation instrument: measure ≠ automatic pedagogy. That authorizes asking what was measured: an SSTEW auto-score or a teacher ranking —not the practice when an adult and a child co-construct meaning and a professional interprets the episode.

The problem is aggravated by ten category confusions. First: it is not dialogic reading as the axis —Jones y Siraj (2023) situate SST in literacy transitions; there may be SST IN an episode with a book; the axis is prolonged co-construction, not the dialogic reading already published—. Second: it is not orality/narration. Third: it is not CLASS nor generic interactions —Howard et al. (2024) find little evidence that global indices predict outcomes; they do not authorize reducing SST to an interaction quality percentile—. Fourth: it is not joint attention. Fifth: it is not ToM / perspective-taking (axis of 9 September). Sixth: it is not prosocial helping/sharing (axis of 8 September). Seventh: it is not conflict negotiation / turn-taking. Eighth: it is not SEL —SSTEW includes emotional well-being as a human subscale (Siraj et al., 2023); the axis is not a SEC auto-score nor an article on well-being—. Ninth: it is not participation/voice —Hildebrandt et al. (2026) anchor agency in high-quality interactions that include SST; this article is not rewritten as voice—. Tenth: it is not executive functions. Mark-making, literacy as axis, sand/water/playdough/loose parts/outdoor, inclusion, well-being, formative assessment, documentation, STEM, motor skills, creativity, free play and symbolic play are not recycled: Cheung et al. (2025) are read for parental SST in joint mathematics, not as STEM; Szymanski et al. (2024) are read for SST speech acts, not as creativity. There is an economy of coordination: the artifacts fit on a slide; the episode in which an adult asks «what would happen if…?», the child hypothesizes and both extend the concept, does not. Contributions: separating craft and artifacts; examining marked cases; and offering four tests of support.

2. State of the art: from the situated practice of co-constructing meaning to the artifact that is displayed

It is useful to separate four strata that the market of «AI for SST in early childhood education» usually mixes. The first is the construct from 3 to 6 years as situated relational practice: prolonged dialogues in which adult and child work together in an intellectual way; open questions, reformulation, conceptual extension, shared hypotheses, emergent metacognition; both parties contribute and the thinking develops (Macha et al., 2024; McLaughlin et al., 2021; Szymanski et al., 2024; Siraj et al., 2023). The second is the craft that cultivates it —an adult who does not present as omniscient; marks uncertainty; invites hypothesizing; co-presence that interprets the episode, not only «there was an SST»— (Meade y Kwan, 2022; McLaughlin et al., 2021; Hildebrandt et al., 2026). The third is the evidence of AI affordances in ECE, robots, teachers' metaphors, GenAI, co-coding and robotic dialogue —without equating them to situated SST pedagogy— (Su y Yang, 2022; Su y Zhong, 2022; Lee et al., 2025; Lim, 2024; Giray, 2026; Jambunathan, 2025; Lee, Lee y Yoon, 2025; Heinmäe et al., 2025; Shin, 2025; Zhang, 2023). The fourth is the framework of human measure, which treats the young child as a subject who thinks with another, not as a vector of SST episodes/hour nor as the addressee of a robot that «scaffolds thinking» (Siraj et al., 2023; Howard et al., 2024; Giray, 2026). Inference: a dialogic quality index does not reformulate «you say that…; and if…?»; an adult who sustains the dialogue does. Su y Yang (2022) map AI in ECE without equating it to situated SST. Lee et al. (2025) propose incorporating robots, not substituting the dialogue. Giray (2026) establishes that the authentic relationship is not replicated by a chatbot. Heinmäe et al. (2025) situate AI as pedagogical advice subordinate to the teaching role.

3. Review method

A critical narrative review was conducted, not a primary meta-analysis. The purpose was not to estimate a homogeneous effect size of the artifacts, but to articulate an argument of pedagogical category with verified sources. Inclusion criteria: (a) 2021–2026, with marked transfer when the sample does not equal 3–6 in an SST classroom, is home/parents, museum, secondary, technical speech HRI, AI curriculum or measure without equating to automatic pedagogy; (b) SST, SSTEW as a human instrument, co-construction dialogue, quality of adult–child interactions, or AI in ECE / robots / GenAI / co-coding / robotic dialogue with artifact/craft relevance; (c) kindergarten, preschool, CENDI or 3–6, or explicit transfer; (d) peer-reviewed journal, chapter, monograph or report with DOI; (e) verifiable DOI or publisher page. Excluded as central object were axes already used in this series —ToM, prosocial, peer-conflict/turn-taking, SEL, participation/voice, CLASS, joint attention, dialogic reading, orality, inclusion, well-being, formative assessment, documentation, mark-making, literacy, sand/water/playdough/loose parts/outdoor, free play, symbolic play, STEM, environmental, motor skills, creativity and inquiry, and executive functions as axis—. Jones y Siraj (2023) are read for SST in transitions, not as literacy nor dialogic reading. Hildebrandt et al. (2026) are read for interaction quality, not as voice. Cheung et al. (2025) are read for parental SST, not as STEM. Szymanski et al. (2024) are read for SST speech acts, not as creativity. Siraj et al. (2023) = human measure ≠ automatic pedagogy. Howard et al. (2024) = review of measure, not a CLASS axis. Kervin et al. (2026) = digital interactions, transfer (museum; 2,5–5 years). Shin (2025) = co-coding in secondary: transfer. Zhang (2023) = ASR of robotic dialogue: artifact, not SST.

The search was executed on 10 September 2026 (slot 09:02 America/Mexico_City, weekday routine) on DOI pages, Crossref, Springer, Elsevier, Wiley, Taylor & Francis, MDPI, PLOS, Hogrefe, NZCER, TLRI, IEEE Xplore, SAGE and publisher sites. Each source was verified against fuentes.md and crossref_check.json. Empirical finding, framework and pedagogical inference were distinguished. No N, d, r, AUC or DOI was invented: when an artifact lacks a verified trial as a display package in CENDI 3–6, it is discussed as a category ceiling. Twenty-one sources from the verified corpus were used.

4. Case 1. An ASR/NLP/ML detector of SST / SSTEW auto-score / dialogic quality index / SST episodes per hour, a GenAI of SST scripts/open-question banks/shared-thinking cards, a dashboard of SST minutes/hour, a chatbot/robot that «does SST», or ChatGPT co-coding of dialogue→teacher ranking do not constitute support for sustained shared thinking

Su y Yang (2022) and Su y Zhong (2022) saturate the ceiling of artifacts when AI in ECE is presented as if it were support for SST. Su y Yang (2022) —a scoping of seventeen studies, 1995–2021— map tools and impacts of AI for children (AI concepts, robotics, creativity, literacy, computational thinking), not that SST episodes/hour cultivate situated co-construction. Su y Zhong (2022) propose an AI curriculum for kindergarten and a social robot as a learning companion of AI principles: a curricular framework, not SST. Inference: «high dialogic quality index = there was support» is a ceiling of analytics. An SSTEW auto-score can coexist with an absence of co-construction. The grammar of display inverts the sequence: first the detector is activated; then it is declared that «there is already SST with AI».

Lee et al. (2025) explore a practical framework for incorporating AI humanoid robots in early childhood education —appropriate concepts, child–AI relationship, activities and challenges—: a framework, not a trial that the robot «does SST». Lee, Lee y Yoon (2025) —six early childhood teachers; 20–30-minute interviews— find that all six considered AI useful for learning and classroom management, but all emphasized human interactions and expressed concern about training and about AI making the teacher obsolete. Status: qualitative finding of perceptions. Inference: that voice prevents selling a chatbot that «scaffolds thinking» without mediation. Lim (2024) —137 undergraduate students in ECE; metaphor analysis— identifies positive conceptions (play and experience, essence of the future, innovation, convenience, assistant teacher) and negative ones (two-facedness, complexity). Inference: «assistant teacher» does not authorize an SST substitute-robot; «two-facedness» names the risk that the dashboard erases. Giray (2026) argues that the role of AI in ECE for 0–7 years remains highly contested and that high-quality development requires authentic relationships that AI does not replicate: a critical framework. Heinmäe et al. (2025) situate AI as a digital resource of pedagogical advice, with the teaching role of sustaining active engagement: a framework of subordination, not of substitution.

Jambunathan (2025) summarizes the use of GenAI (ChatGPT, Claude, Dall-E, among others) in early childhood teacher education for lesson plans and teaching tasks: adult preparation. Marked inference: producing materials by prompt may be subordinate preparation; SST support begins when adult and child co-construct in the episode, not when an open-question bank closes the dialogue as a recipe. Shin (2025) saturates the co-coding anchor and must be read with marked transfer. It analyzes 1 287 secondary science utterances with ChatGPT-4o; Cohen's kappa of 0,56; the greatest disagreement on metacognition. Status: feasibility finding in a single-researcher case. Transfer: secondary, not CENDI 3–6; science dialogue ≠ early childhood SST; maximum discordance precisely on the trait that SST claims as emergent. Inference: converting automatic co-coding into a ranking of «teacher dialogic quality» of 3–6 without situated judgment is illegitimate even before transferring the stage. Zhang (2023) saturates the dialogue-robot anchor: enhancement of children's speech with two-stage attention; an improvement of 3,2 dB in SSNR. Status: technical ASR finding. Transfer: intelligibility for the machine, not SST pedagogy. Restrictive inference: ASR/NLP detector ≠ adult who questions and reformulates; GenAI of shared-thinking cards ≠ open episode; dashboard of SST minutes/hour ≠ observation; chatbot/robot that «does SST» ≠ co-presence; co-coding→ranking ≠ pedagogical judgment. The artifacts lack, in the verified corpus, pedagogical-equivalence trials with the craft in CENDI 3–6; no d, r or AUC of «support» is invented. They share a grammar of substitution: the score speaks for the dialogue; the GenAI closes; the dashboard substitutes observation; the robot erases mediation; the ranking converts process into a percentile without judgment.

5. Case 2. What the kindergarten does do when there is support for SST: situated relational craft and what Macha, Meade, McLaughlin, Szymanski, Techapoonpon, Cheung, Jones/Siraj, human SSTEW, Hildebrandt/Macha Beyond Voice and Howard measure —with Kervin in digital transfer

The floor of the craft is not a score: it is an episode in which adult and child work together in an intellectual way; open questions, reformulation, conceptual extension, shared hypotheses; both parties contribute and the thinking develops; co-presence that interprets the dialogue —not only «there was an SST»—; and, when there is uncertainty, the adult marks it instead of presenting as omniscient (Macha et al., 2024; McLaughlin et al., 2021; Szymanski et al., 2024; Siraj et al., 2023; Hildebrandt et al., 2026). Inference: there is support when that practice is protected; not when an SSTEW auto-score, a GenAI of scripts or a ranking is displayed. The craft is recognized in the episode: the child offers a hypothesis; the adult reformulates («you think that…; what would happen if…?») and sustains that the thinking be extended; the robot or the GenAI, if they appear, are subordinate means to the adult, not substitutes.

Macha et al. (2024) saturate the dialogic-practice anchor: a childcare center in Berlin; 33 children aged 4–6 years; four group discussions of three to four children; two in-depth transcripts; dialogue according to SST rules combined with QuaSi. Empirical finding: the children expressed in an explicit way ideas, wishes, likes and dislikes about the quality of the center; the dialogue developed with equal contributions; the children enacted their agency and formulated reasoned opinions. Marked inference: making perspectives explicit through SST ≠ an SST detector nor «voice» as an already published axis; the study underlines that the research situation is already pedagogical. Meade y Kwan (2022) —Daisies, Aotearoa New Zealand; SSTEW as a professional-development resource (TLIF)— describe the weaving of SST into Te Whāriki: data knowledge for practice, not an auto-score. Inference: using SSTEW with kaiako judgment ≠ SSTEW auto-score as pedagogical assessment.

McLaughlin et al. (2021) saturate the data-informed-teaching anchor: two kindergartens (Linton y Makino); CEOS, video, LENA, SSTEW and ECERS-E by trained human observers. Findings: SST and positive learning interactions occurred less frequently than other types; time with a teacher correlated positively with SST episodes; the teams used the data to affirm and challenge their practice and to plan SST with more intentionality. Status: mixed finding from two exploratory cases. Restrictive inference: LENA quantifies turns and words, not the quality of thinking —the report warns of this—; SSTEW and CEOS here serve the team's reflection, not a dashboard of SST minutes/hour. Converting those tools into «shared-thinking compliance» inverts the meaning of the project.

Szymanski et al. (2024) —N = 51 children aged 4–6 years (n = 17 per condition); between-subjects design (instructive, SST, neutral)— find more innovative behavior (more new cards; fewer exact copies) in the SST condition than in the instructive one. Status: experimental finding. Marked inference and of the authors' limits: a subset of SST speech acts was isolated (uncertainty markers: «I think», «could», «perhaps»), not a complete sustained dialogue; it does not equal a classroom detector nor a GenAI of scripts; it is not converted into a creativity axis. Techapoonpon et al. (2025) —one-arm quasi-experiment; n = 55 parents of children aged 4–6 years; six weeks of SST activity books— report an increase in parental empowerment at three and six weeks, with persistence at three months. Transfer: home, not a CENDI classroom; the outcome is perceived empowerment, not SST observed in the kindergarten; the authors warn that they did not measure SST skill in a direct way. Inference: self-directed activity books ≠ GenAI open-question banks as a classroom recipe.

Cheung et al. (2025) —466 parents; kindergarten children; a scale of parental SST in joint mathematics; three factors (exchange of ideas about solution processes, child-centered atmosphere, mathematical thinking engagement)— find that constructivist conceptions predicted the use of those strategies, associated differentially with numeracy and with approach/avoidance motivation. Marked distinction: parental SST, not a STEM axis nor a classroom SSTEW auto-score. Jones y Siraj (2023) argue the contribution of SST to literacy transitions in the English curriculum. Inference: SST may occur around a text; that does not recycle dialogic reading nor authorize sustained dialogue lesson packs by prompt.

Siraj, Kingston y Melhuish (2023) publish the SSTEW Scale for 2–6 years as a human-observation Quality Rating Scale: confidence and self-regulation, socioemotional well-being, language and communication, learning and critical thinking —including SST items in narration/books and in investigation—. Status: instrument framework. Contrast inference: SSTEW is human observation for quality improvement; not an auto-score, not an automatic dialogic quality index, not an article on well-being nor on dialogic reading. Howard et al. (2024) —90 studies; 870 associations; children aged 3–5 years— find little evidence that global ECEC quality indices relate to child outcomes, and only partial consistency even when the dimension aligns with the outcome. Inference: a dashboard interaction quality percentile does not sign situated SST. Hildebrandt et al. (2026) show that agency is enacted above all in high-quality participatory interactions —reciprocity, co-construction of meaning—, not in mere formal structures. Marked distinction: SST as interaction quality, not a participation/voice axis. Kervin et al. (2026) —36 families; a child aged 2,5–5 years; digital play in a museum; themes of expertise, control and pace— find that high adult control can facilitate an apparently passive engagement. Transfer: museum, not CENDI; it includes 2,5 years. Inference: adult mediation of control and pace is the craft; a robot that «does SST» inverts it.

Taken together: there are systems that make children's thinking explicit with SST, weave SSTEW into professional development, use data to plan SST, isolate SST speech acts, stimulate parental SST, argue SST in transitions, publish a human scale, review the measure of interactions or describe digital play with transfer. None —by itself— signs an ASR detector, a GenAI of scripts, a dashboard of minutes, a substitute-robot or a teacher ranking. The evidence is real in its domain; «score / script / robot / co-coding = support for SST» is an illegitimate category inference.

6. Case 3. SST ≠ dialogic reading as axis, orality-narration, generic CLASS, joint attention, ToM, prosocial, peer-conflict, SEL, participation-voice or executive functions

Ten boundaries protect the axis. First: dialogic reading —Jones y Siraj (2023) situate SST in literacy transitions; Siraj et al. (2023) include an SST item through books; there may be co-construction IN an episode with a book; the axis is prolonged dialogue, not the dialogic reading already published—. Inference: «we work dialogic reading = there is SST» does not sign; a GenAI open-question bank is not SST either. Second: orality/narration —there is speech that extends concepts; this article is not converted into orality—. Third: CLASS / generic interactions —Howard et al. (2024) show the fragility of global indices; McLaughlin et al. (2021) distinguish SST from general or informational interactions; SST is not reduced to an interaction quality percentile—. Fourth: joint attention —sustaining an intellectual dialogue is not joint attention as axis—. Fifth: ToM / perspective-taking —there may be attribution of a point of view IN an SST episode; the axis is not that of 9 September—. Sixth: prosocial helping/sharing —there may be cooperation IN a dialogue; the axis is not that of 8 September—. Seventh: conflict negotiation / turn-taking —SST requires turns, but it is not a pedagogy of the turn—. Eighth: SEL —Siraj et al. (2023) measure emotional well-being as a human subscale; Hildebrandt et al. (2026) anchor attunement in interaction quality; the axis is not a SEC auto-score nor an article on well-being—. Ninth: participation/voice —Macha et al. (2024) and Hildebrandt et al. (2026) show agency in SST dialogue; voice is not recycled as axis—. Tenth: executive functions —only a peripheral contrast—. Giray (2026), Lee, Lee y Yoon (2025) and Heinmäe et al. (2025) require human mediation. Inference: coherent AI in 3–6 remains on the side of the adult (preparation, documentation, reflection; Meade y Kwan, 2022; Jambunathan, 2025; McLaughlin et al., 2021), not as a detector, a GenAI of lesson packs, a dashboard, a robot that «does SST» or a ranking without judgment. SST may touch, laterally, a book, speech, emotion or agency; it is not allowed to be reduced to those axes. A center that «works literacy», «scores CLASS» or «gives voice» has not demonstrated by that label alone the craft of co-constructing meaning. Cheung et al. (2025) are read for parental SST, not as STEM; Szymanski et al. (2024) are read for SST speech acts, not as creativity; Kervin et al. (2026) are read for digital mediation, not as free play.

7. Inferential framework: four tests to claim that there is support for sustained shared thinking, not an artifact

The framework that follows is a 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 the artifacts constitute support for sustained shared thinking in early childhood.

7.1. Test of situated practice (prolonged dialogue of co-construction; open questions, reformulation, conceptual extension, shared hypotheses; both parties contribute), not of the ASR/NLP/ML detector nor of the chatbot/robot that «does SST». Macha et al. (2024), McLaughlin et al. (2021), Szymanski et al. (2024) and Siraj et al. (2023) define the craft/artifact contrast. Su y Yang (2022), Lee et al. (2025), Zhang (2023, transfer) and Giray (2026) map AI, robots and ASR as affordance or speech engineering, not as situated SST. Inference: if the «evidence» is SST-score, SSTEW auto-score, dialogic quality index, SST episodes per hour or a robot that «scaffolds thinking» without an adult, the center did analytics or HRI, not support.

7.2. Test of adult co-presence that questions, reformulates and extends —not only «there was an SST»—, not of the GenAI of SST scripts / lesson packs / open-question banks / shared-thinking cards. Meade y Kwan (2022), McLaughlin et al. (2021), Macha et al. (2024) and Hildebrandt et al. (2026) situate practice and mediation. Jambunathan (2025) and Heinmäe et al. (2025) admit GenAI as the adult's preparation or advice. Techapoonpon et al. (2025) and Cheung et al. (2025) situate SST with adults in the home, not classroom recipes. Inference: a script by prompt does not sign the episode in which the adult marks uncertainty and the child hypothesizes.

7.3. Test of human mediation and of the interpretation of SST as relational practice, not of the dashboard of SST minutes/hour nor of ChatGPT co-coding converted into a ranking of teacher dialogic quality. Siraj et al. (2023) anchor human measure. Howard et al. (2024) anchor the limits of global indices. McLaughlin et al. (2021) use LENA and SSTEW for the team's reflection, not for surveillance. Shin (2025, transfer) shows kappa 0,56 and maximum disagreement on metacognition. Inference: high SST minutes/hour do not by themselves raise scaffolding. Subordinate preparation of the adult is legitimate; the scorer that converts the child or the teacher into a compliance vector is not.

7.4. Test of category distinction and of professional judgment, not of the product catalogue. SST ≠ dialogic reading, orality-narration, CLASS, joint attention, ToM, prosocial, peer-conflict, SEL, participation-voice or FE. Jones y Siraj (2023), Hildebrandt et al. (2026), Howard et al. (2024) and Siraj et al. (2023) prevent those confusions. Giray (2026) and Lee, Lee y Yoon (2025) require human mediation. Inference: support is fulfilled with prolonged dialogue of co-construction, co-presence that questions/reformulates/extends and agency of both parties —not by scoring SST nor by substituting the adult with a conversational agent. The framework admits digital subordinate to the adult's preparation, documentation and reflection (Meade y Kwan, 2022; McLaughlin et al., 2021; Jambunathan, 2025; Heinmäe et al., 2025). It rejects declaring support by the artifacts (Giray, 2026; Su y Yang, 2022; Shin, 2025; Zhang, 2023). The four tests are read together.

8. Discussion

Three tensions. First: displaying the artifacts versus exercising support. The cases of sections 5–6 —with marked transfers and distinctions— sustain the contrast; Su y Yang (2022) and Su y Zhong (2022) map AI without equivalence to situated SST. Second: automated assessment or substitute-robot versus relational pedagogy —declaring support by SSTEW auto-score, dialogic quality index or by a system that enhances speech for the machine is inverted pedagogy; Siraj et al. (2023) and Howard et al. (2024) show that even a human process measure is not the craft, and that global indices predict little—. Third: GenAI of shared-thinking cards / a chatbot that «does SST» / co-coding→ranking versus craft (Giray, 2026; Lee, Lee y Yoon, 2025; Jambunathan, 2025; Shin, 2025; Zhang, 2023): selling a conversational agent as scaffolding of thinking in 3–6 confuses product with practice. McLaughlin et al. (2021) confirm that data serve when the team interprets them; Meade y Kwan (2022) confirm SSTEW as PLD, not as an auto-score; Shin (2025) confirms that human–model disagreement concentrates on metacognition; Giray (2026) subordinates any use to the human relationship. The four tests of section 7 read these tensions. Subordinate digital is legitimate toward the adult (preparation of open questions; SSTEW for the human observer; Jambunathan, 2025, as teacher education; Heinmäe et al., 2025, as pedagogical advice); inverting the sequence is not. If the center shows prolonged dialogues of co-construction, an adult who questioned, reformulated and extended, and children who hypothesized with agency, it may speak of support; if it only shows detectors, GenAI scripts, SST minutes/hour, robots that «do SST» or rankings without judgment, it speaks of artifacts.

9. Limits

This review is narrative. It does not apply its own PRISMA nor estimate combined effects. No d, r or AUC is invented. Macha et al. (2024) are a small study (33 children; four discussions; two transcripts). Meade y Kwan (2022) report one center and SSTEW as PLD. McLaughlin et al. (2021) are two kindergartens in Aotearoa New Zealand; CEOS and LENA do not by themselves measure the quality of thinking. Szymanski et al. (2024) isolate speech acts in N = 51: the authors warn that they do not portray a complete dialogical SST. Techapoonpon et al. (2025) are a one-arm quasi-experiment in the home, without a direct measure of SST skill. Cheung et al. (2025) measure parental SST in mathematics: home, not classroom. Jones y Siraj (2023) are a transitions chapter: they are not converted into a dialogic-reading axis. Siraj et al. (2023) are an instrument, not a Latin American classroom trial. Hildebrandt et al. (2026) are read for interaction quality, not as voice. Howard et al. (2024) review 2000–2022: they do not evaluate SST detectors. Kervin et al. (2026) are a museum and 2,5–5 years: transfer. Lim (2024) measure 137 future teachers. Lee, Lee y Yoon (2025) measure six teachers. Jambunathan (2025) is a trainer's experience. Heinmäe et al. (2025) are a pedagogical-advice chapter. Lee et al. (2025), Su y Yang (2022) and Su y Zhong (2022) map AI/robots/curriculum, not the display package in CENDI. Shin (2025) is secondary. Zhang (2023) is conference ASR. Giray (2026) is a critical perspective. No verified trials of the artifacts as a single package of «SST with AI» in a 3–6 classroom were located; they are discussed as a category ceiling. The inferences of section 7 are category hypotheses, not implementation evidence.

10. Conclusions

The artifacts —an ASR/NLP/ML detector or scores of SST / SSTEW auto-score / dialogic quality index / SST episodes per hour / shared-thinking compliance / interaction quality percentile; a GenAI of SST scripts/lesson packs/open-question banks/shared-thinking cards; a dashboard of SST minutes/hour; a chatbot/robot that «does SST» without adult mediation; ChatGPT co-coding converted into a teacher ranking— do not constitute support for sustained shared thinking in early childhood education. Su y Yang (2022) and Su y Zhong (2022) confirm AI affordances and curriculum without equivalence to situated SST. Lee et al. (2025), Lim (2024), Giray (2026), Jambunathan (2025), Lee, Lee y Yoon (2025) and Heinmäe et al. (2025) set frameworks, metaphors, perceptions, teacher preparation and pedagogical advice, not a substitute-robot nor a GenAI of a closed recipe. Shin (2025) and Zhang (2023) are read as an artifact ceiling with transfer. The empirical cases (sections 5–6) require reading SST, human SSTEW and interaction quality by what they measure and by what they do not equal as situated pedagogy. When there is support there are prolonged dialogues of co-construction, co-presence that questions, reformulates and extends, and pedagogical interpretation (Macha et al., 2024; McLaughlin et al., 2021; Siraj et al., 2023). SST is distinguished from dialogic reading, orality-narration, CLASS, joint attention, ToM, prosocial, peer-conflict, SEL, participation-voice and FE.

Where the sources do not measure a kindergarten, this article does not claim it. Where they measure mappings of AI, robots, GenAI of training, secondary co-coding, robotic ASR, SST in discussions of quality, SSTEW as PLD, data-informed teaching, SST speech acts, parental activity books, mathematical SST in the home, literacy transitions, a human scale or digital play in a museum, it does not translate them into support via SST-score, script by prompt, dashboard of minutes, substitute-robot or teacher ranking. Accompanying girls and boys from three to six years in sustained shared thinking is to exercise prolonged dialogues in which adult and child co-construct meaning, open questions, reformulation, conceptual extension, shared hypotheses, emergent metacognition, scaffolding that does not present as omniscient, and co-presence that interprets SST as situated relational practice, not as an «SST-score». The rest is a detector, a GenAI of scripts, a dashboard of minutes, a robot that «does SST» and a ranking. It is not support for sustained shared thinking in early childhood education, and it must not be presented as what it is not.

Laboratorio Editorial de NEXTECH.IA / Ingeniero Mitre.

References

  1. Cheung, S. K., Kwan, J. L. Y., Chan, W. W. L., Kum, B. H. C., y Ho, P. L. (2025). Parents' use of sustained shared thinking during joint mathematics activities with young children: An investigation of its measurement, antecedents, and outcomes. British Journal of Educational Psychology, 95(2), 363–383. https://doi.org/10.1111/bjep.12722
  2. Giray, L. (2026). A critical stand against artificial intelligence in early childhood education. Discover Artificial Intelligence, 6, Article 768. https://doi.org/10.1007/s44163-026-01510-x
  3. Heinmäe, E., Puksand, H., y Kollom, K. (2025). AI in action: Pedagogical advice using artificial intelligence as a digital resource in early childhood education. En S. Papadakis (Ed.), AI in early education: Integrating artificial intelligence for inclusive and effective learning (pp. 41–58). Wiley. https://doi.org/10.1002/9781394352821.ch03
  4. Hildebrandt, A., Macha, K., Lonnemann, J., y Hildebrandt, F. (2026). Beyond voice: High-quality interactions and children’s agency in early childhood education. Early Childhood Education Journal. https://doi.org/10.1007/s10643-026-02237-1
  5. Howard, S. J., Lewis, K. L., Walter, E., Verenikina, I., y Kervin, L. K. (2024). Measuring the quality of adult–child interactions in the context of ECEC: A systematic review on the relationship with developmental and educational outcomes. Educational Psychology Review, 36(1), Article 6. https://doi.org/10.1007/s10648-023-09832-3
  6. Jambunathan, S. (2025). Integrating artificial intelligence into early childhood teacher education. Contemporary Issues in Early Childhood. https://doi.org/10.1177/14639491251340141
  7. Jones, P., y Siraj, I. (2023). The contribution of ‘sustained shared thinking’ to successful literacy transitions in English curriculum. En A. Thwaite, A. Simpson y P. Jones (Eds.), Dialogic pedagogy: Discourse in contexts from pre-school to university. Routledge. https://doi.org/10.4324/9781003296744-6
  8. Kervin, L. K., Lewis, K. L., Day, N., y Howard, S. J. (2026). Adult and child interactions when using digital technologies. Education Sciences, 16(8), Article 1260. https://doi.org/10.3390/educsci16081260
  9. Lee, J., Jo, J., Lee, J. O., y Kim, S. H. (2025). Incorporating humanoid artificial intelligence (AI) robots into early childhood education. Early Childhood Education Journal, 53(8), 2849–2857. https://doi.org/10.1007/s10643-024-01690-0
  10. Lee, J., Lee, J. O., y Yoon, J. (2025). Exploring perceptions of early childhood teachers on the use of artificial intelligence in early childhood education. Journal of Early Childhood Teacher Education. https://doi.org/10.1080/10901027.2025.2600033
  11. Lim, E. M. (2024). Metaphor analysis on pre-service early childhood teachers’ conception of AI (Artificial Intelligence) education for young children. Thinking Skills and Creativity, 51, Article 101455. https://doi.org/10.1016/j.tsc.2023.101455
  12. Macha, K., Hildebrandt, F., Wronski, C., Lonnemann, J., y Urban, M. (2024). Making it explicit – Sustained shared thinking dialogue as a way to explore children's perspectives on quality in German early childhood education and care. British Educational Research Journal. https://doi.org/10.1002/berj.4054
  13. McLaughlin, T., Cherrington, S., McLaughlin, C., Aspden, K., Hunt, L., y Gifkins, V. (2021). Data, knowledge, action: Exploring sustained shared thinking to deepen young children’s learning. Teaching and Learning Research Initiative. https://doi.org/10.18296/tlri.0021
  14. Meade, A., y Kwan, M. (2022). Weaving in data knowledge about sustained shared thinking enhances early childhood education practice. Early Childhood Folio, 26(2), 25–31. https://doi.org/10.18296/ecf.1112
  15. Shin, E. (2025). Co-coding classroom dialogue: A single researcher case study of ChatGPT-assisted analysis in science education. Journal of Computer Assisted Learning, 41(4), Article e70089. https://doi.org/10.1111/jcal.70089
  16. Siraj, I., Kingston, D., y Melhuish, E. (2023). The Sustained Shared Thinking and Emotional Well-being (SSTEW) Scale. Routledge. https://doi.org/10.4324/9781003379867
  17. Su, J., y Yang, W. (2022). Artificial intelligence in early childhood education: A scoping review. Computers and Education: Artificial Intelligence, 3, Article 100049. https://doi.org/10.1016/j.caeai.2022.100049
  18. Su, J., y Zhong, Y. (2022). Artificial intelligence (AI) in early childhood education: Curriculum design and future directions. Computers and Education: Artificial Intelligence, 3, Article 100072. https://doi.org/10.1016/j.caeai.2022.100072
  19. Szymanski, L., Hildebrandt, F., y Wronski, C. (2024). Sustained Shared Thinking fördert das innovative Verhalten vier- bis sechsjähriger Kinder. Frühe Bildung, 13(3). https://doi.org/10.1026/2191-9186/a000642
  20. Techapoonpon, K., Pruttithavorn, W., Sripatanaskul, M., Limpiti, N., y Yoong, K. (2025). The efficacy of promoting sustained shared thinking through the use of activity books on parental empowerment; A quasi-experimental study. PLOS One, 20(7), Article e0328537. https://doi.org/10.1371/journal.pone.0328537
  21. Zhang, H. (2023). A robot dialogue system for preschool education based on AI artificial intelligence. En 2023 International Conference on Integrated Intelligence and Communication Systems (ICIICS). IEEE. https://doi.org/10.1109/iciics59993.2023.10421599