1. Introduction and problem

In the 3-to-6 age span—kindergarten, preschool, CENDI, infant school—a package of three artefacts has taken hold that claim to stand for pedagogical–relational interaction quality. The first is a system that scores CLASS videos: a multimodal or language pipeline that estimates Positive Climate, Negative Climate or Instructional Support and declares that the classroom “already has AI-measured quality.” The second is a wearable that prompts “respond now”: a bug-in-ear device or wristband that increases teachers’ opportunities to respond and presents the alert as if it were relational sensitivity. The third is a model that classifies conversational turns: an algorithm that counts or labels adult–child turns and delivers the count as if it were pedagogical reciprocity. All three are visible and cheap in coordination time. They allow centres to exhibit that they “already do interaction quality with artificial intelligence.” The leap—from obtaining a score, receiving an alert or accumulating turns to asserting that process quality exists—is authorized neither by the evidence on automated observation nor by what Starting Strong and CLASS measure when Emotional Support, Classroom Organization and Instructional Support are present in the daily life of the classroom.

The thesis of this article is restrictive. A system that scores CLASS videos, a wearable that prompts “respond now,” or a model that classifies conversational turns do not constitute pedagogical–relational interaction quality in early childhood education. At this stage, process quality is sensitivity, reciprocity and scaffolding—the serve-and-return of the Center on the Developing Child; the CLASS dimensions Emotional Support (ES), Classroom Organization (CO) and Instructional Support (IS); the OECD Starting Strong anchor of meaningful daily interactions—not a scoring dashboard, an opportunity-to-respond alert, or an automatic turn count. OECD (2021), in Starting Strong VI, analyses curricula from twenty-six countries—fifty-six curricula—and anchors ECEC quality in everyday adult–infant interactions, not in measurement artefacts. That does not authorize translating “there is an automated CLASS score” as “there is process quality.” It authorizes asking what was measured: often, correlation with human judgments on fifteen-minute segments, not the sustained sensitivity when a four-year-old offers a “serve” and the adult responds, expands and scaffolds.

The problem is aggravated by three category confusions that product sheets do not mention and that this article separates rigorously. First: process quality is not pedagogical documentation—portfolios, panels or digital logs that record what happened. Second: it is not social–emotional learning (SEL) as a programme of global emotional competencies. Third: it is not executive functions such as inhibitory control, working memory and cognitive flexibility. Gómez and Strasser (2021) and Gómez-Muzzio and Strasser (2025) are cited here only to distinguish conversational turns from pedagogical interaction quality: turns predict socioemotional development, but a turn classifier is not, by existing, the craft of sensitivity and scaffolding in the room. This work does not recycle articles on documentation, SEL or executive functions. The question is one of process quality: what counts as pedagogical–relational interaction when an early childhood centre “does AI and interaction quality.” Confusing the product—score, wearable, classifier—with the process is the error this work names. Ramakrishnan, Zylich, Ottmar, LoCasale-Crouch and Whitehill (2023) demonstrate the feasibility of multimodal estimation of Positive Climate and Negative Climate; Whitehill and LoCasale-Crouch (2024) connect statement-level judgments with CLASS Instructional Support scores. Pedagogical inference, marked as such: the market for “AI for interaction quality in early childhood” inherits that evaluative distribution and converts it into a promise of quality. It turns automated measurement and the alert into the whole of the relational craft.

The contributions are three: to reconstruct the state of the art that separates process quality (serve-and-return; CLASS ES/CO/IS; OECD Starting Strong) from scoring, alerting and turn counting; to examine three families of empirical cases; and to offer four tests for deciding when a kindergarten may claim that pedagogical–relational interaction quality exists, and not only a video-scoring system, a wearable or a turn classifier.

2. State of the art: from process quality to the artefact on display

It is useful to separate four strata that the market for “AI for interaction quality in early childhood” usually mixes. The first is the construct of process quality for ages 3 to 6 as sensitivity, reciprocity and scaffolding in daily interactions—serve-and-return; Emotional Support, Classroom Organization and Instructional Support; OECD Starting Strong—(Center on the Developing Child at Harvard University, n.d.; OECD, 2021, 2023; Finders et al., 2021; Howard et al., 2024; von Suchodoletz et al., 2023; Paschall et al., 2023). The second is the relational craft that cultivates it—teacher noticing, contingent response, classroom organization and mediated instructional support—(LoCasale-Crouch et al., 2023; NAEYC, 2022). The third is the evidence on automated CLASS observation, bug-in-ear coaching, turn classification and AI mappings in ECE (Ramakrishnan et al., 2023; Whitehill and LoCasale-Crouch, 2024; Green et al., 2023; Ferjan Ramírez et al., 2021; Chen, 2024; Su and Yang, 2022; Ljungcrantz, 2026). The fourth is the rights, systems and developmentally appropriate practice framework that treats the 3–6-year-old as a subject of process interactions, not as a feature vector for a scoring pipeline (UNESCO, 2021; Miao and Holmes, 2023; European Commission, 2022; U.S. Department of Education, 2023).

In the construct stratum, OECD (2021) establishes that ECEC quality is played out in meaningful everyday interactions and analyses how twenty-six countries articulate that mandate across fifty-six curricula. OECD (2023) situates digitalization in uses with pedagogical sense, not in an alert feed that replaces reciprocity. The Center on the Developing Child (n.d.) defines serve-and-return as the sequence in which the infant initiates—serve—and the adult responds contingently, expands and sustains the exchange. Finders, Budrevich, Duncan, Purpura, Elicker and Schmitt (2021), with 684 preschool children, show that CLASS dimensions are not interchangeable: mean Classroom Organization associates positively with mathematics; within-day CO variability associates negatively with literacy; mean Instructional Support associates negatively with language in their models. Status: a finding that process quality is multidimensional and sensitive to variability, not to a single score. Howard, Lewis, Walter, Verenikina and Kervin (2024), in a systematic review of ninety studies and eight hundred seventy associations with children aged 3 to 5, find little support for global indices of interaction quality and better consistency on some dimensions—for example, supporting play—with low support for instructional support even when aligned with cognitive outcomes. Inference: a dashboard that summarizes “CLASS quality” in a number does not, by existing, cover sensitivity, reciprocity and scaffolding.

In the craft stratum, LoCasale-Crouch, Romo-Escudero, Clayback, Whittaker, Hamre and Melo (2023) randomly assign sixty-one toddler educators to control or to the Effective Classroom Interactions for Toddler Educators (ECI-TE) course: those who complete the course show greater noticing skills and higher educator–infant interaction quality, with a medium effect size. Status: a finding of professional development oriented to noticing and enacting effective interactions. It is not a finding that a wearable that pushes “respond now” reproduces that craft. NAEYC (2022) requires developmentally appropriate practice: quality is verified in relationships, not in measurement artefacts. Inference: an automated score does not walk with the infant. An adult who notices the serve and responds does.

In the artefact stratum, Ramakrishnan et al. (2023) present ACORN: multimodal machine learning to estimate CLASS Positive Climate and Negative Climate; they report r = 0.55 for PC and r = 0.63 for NC on fifteen-minute segments from the UVA Toddler corpus (n = 300 segments), with AUC = 0.70 for Positive Climate moments. Whitehill and LoCasale-Crouch (2024) combine Whisper, Llama 2 and bag-of-words for Instructional Support: Pearson correlation up to 0.48 versus human IRR up to 0.55, aggregating statement-level judgments into a fifteen-minute CLASS IS score in toddler and pre-K. Green, Olsen and Nandakumar (2023) test bug-in-ear coaching in Head Start to increase opportunities to respond (OTR) in mathematics with four teachers: a functional relation in two of four; low maintenance. Ferjan Ramírez, Hippe and Kuhl (2021) warn that automatic measures of parent–infant conversational turns are not equivalent to manual judgments. Status: a finding of scoring feasibility, OTR coaching and metrological caution about turns; not a finding that score, alert or classifier constitute process quality. Chen (2024), Su and Yang (2022) and Ljungcrantz (2026) map AI in ECE without equating it to pedagogical–relational interaction quality in kindergarten.

In the systems stratum, UNESCO (2021) requires human oversight. Miao and Holmes (2023) set pedagogical validation and age thresholds for generative AI. The European Commission (2022) and the U.S. Department of Education (2023) agree on not replacing professional judgment. von Suchodoletz et al. (2023) meta-analyse associations between ECEC quality and child outcomes, reinforcing that quality matters and is not reducible to an artefact. Paschall et al. (2023), with 2 114 Head Start students, identify profiles in which the largest groups positive emotional climate with low instructional support: a finding that dimensions do not travel together. Inference: a four-year-old is neither the input to a scoring pipeline nor the passive recipient of a wearable alert. The child is the protagonist of a serve-and-return that a guaranteeing adult exercises with sensitivity, reciprocity and scaffolding.

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 automated observation, but to articulate an argument of pedagogical category with verified sources. Inclusion criteria: (a) 2021–2026, with transfer explicitly marked when the sample is not ages 3–6; (b) process quality, adult–child interactions, serve-and-return, CLASS (ES/CO/IS), automated observation, bug-in-ear coaching, conversational turns or AI in early childhood education; (c) relevance to kindergarten, preschool, CENDI or ages 3–6; (d) peer-reviewed journal, DOI or report from NAEYC, UNESCO, OECD, the European Commission, the Center on the Developing Child or a department of education; (e) verifiable DOI or editorial page. Axes already used in this series were excluded as central objects—pedagogical documentation, SEL as a substitute programme, executive functions, generative tutoring, gaps, privacy, UDL, STEM, numeracy, formative assessment, family–school, continuing professional development—although some appear as category boundaries.

The search was executed on 28 August 2026 (slot 01:02 America/Mexico_City) on DOI pages, Springer, Elsevier, Wiley, SAGE, IEEE Xplore, OECD iLibrary, UNESDOC, JAIR, NAEYC, Zenodo, the Harvard Center on the Developing Child and editorial sites. Each source was verified against at least one of those pages. Empirical findings, conceptual or normative frameworks, and pedagogical inference marked as such were distinguished. Priority was given to the process quality / documentation / SEL / executive functions distinction, and to Ferjan Ramírez et al.’s (2021) caution about automatic turns versus manual judgments.

4. Case 1. Scoring CLASS videos, alerting “respond now,” or classifying turns does not constitute interaction quality

Ramakrishnan, Zylich, Ottmar, LoCasale-Crouch and Whitehill (2023) publish in IEEE Transactions on Affective Computing the ACORN system that best names the first artefact when it is presented as interaction quality. They train multimodal machine learning to estimate CLASS Positive Climate and Negative Climate. On the UVA Toddler corpus, with three hundred fifteen-minute segments, they report correlation r = 0.55 with Positive Climate and r = 0.63 with Negative Climate, and AUC = 0.70 for detecting Positive Climate moments. Status of the evidence. Empirical finding of automated estimation: the system approximates human judgments on two emotional-climate dimensions. It is not a finding that installing a CLASS scorer develops or guarantees Emotional Support, Classroom Organization and Instructional Support in classroom life, nor that sensitivity, reciprocity and scaffolding are covered. The datum the market does not cite is the ceiling: moderate correlations and an AUC for moment detection, not equivalence with lived process quality. Pedagogical inference, marked as such: this is the gesture a kindergarten copies when it “does quality with AI that scores videos.” It records, it scores, it exhibits the dashboard. What exists is an observation artefact. Process quality, in OECD (2021) and the Center on the Developing Child (n.d.), asks for contingent serve-and-return, not an r of 0.55 with Positive Climate.

Whitehill and LoCasale-Crouch (2024) saturate the portrait of the first artefact from Instructional Support. They combine Whisper transcription, Llama 2 and bag-of-words; they aggregate statement-level judgments into a fifteen-minute CLASS IS score in toddler and pre-K. Pearson correlation reaches up to 0.48 versus human interrater reliability up to 0.55. Status: a finding of connection between global prediction and statement-specific feedback. It is not a finding that a language model constitutes instructional scaffolding. Inference, marked as such: approaching human IRR on a segment does not authorize declaring “there is already quality Instructional Support.” It authorizes saying there is estimation. Finders et al. (2021) further show that IS and CO do not behave the same with respect to outcomes: a single score flattens that heterogeneity.

Green, Olsen and Nandakumar (2023) name the second artefact. In a Head Start programme, they use bug-in-ear coaching to increase opportunities to respond in mathematics with four teachers. They document a functional relation in two of four and low maintenance. Status: a finding that an alert or in-ear coaching can increase OTR in some cases, with fragile maintenance. It is not a finding that a wearable saying “respond now” is relational sensitivity or serve-and-return. Inference: OTR is an academic response opportunity; sensitivity is noticing the child’s serve—verbal, gestural, playful—and returning it with affective and cognitive contingency. Ferjan Ramírez, Hippe and Kuhl (2021) close the third artefact with a metrological warning: automatic measures of conversational turns are not equivalent to manual measures. Status: a finding of disagreement between automatic and manual. Inference: a turn classifier can count acoustic events; it does not demonstrate pedagogical reciprocity or scaffolding. Chen (2024), Su and Yang (2022) and Ljungcrantz (2026) confirm growth of AI in ECE without evidence that score, alert or classifier replace the craft. Inference: a CENDI that delivers “AI-generated CLASS quality” or “enough turns according to the model” has done measurement or counting. Process quality is verified in whether the adult responded to the serve, organized the room with warmth and scaffolded thinking. It is not verified in the scorer’s AUC.

5. Case 2. What the kindergarten does do when process quality is present: sensitivity, reciprocity and scaffolding

The Center on the Developing Child at Harvard University (n.d.) defines serve-and-return as a current interaction framework: the infant initiates; the adult notices, responds, names, waits and sustains the exchange. Status: an institutional conceptual framework verifiable on the centre’s editorial page. It is not an AI finding. Pedagogical inference, marked as such: this is the object an early childhood centre may call pedagogical–relational interaction quality. Reciprocity is not a turn count: it is affective and cognitive contingency. A wearable that pushes “respond now” without reading the serve may increase OTR and, at the same time, interrupt the child’s rhythm. A CLASS scorer may rate positive climate on video and still have failed to capture whether the adult expanded the child’s thinking in the play corner.

OECD (2021) saturates the framework from comparative policy: across twenty-six countries and fifty-six curricula, Starting Strong VI holds that quality is played out in meaningful daily interactions. OECD (2023) adds that digitalization should empower without replacing those interactions. Status: a systems framework. Inference: a kindergarten cannot treat “we already measure CLASS with AI” as equivalent to having transformed everyday interactions. LoCasale-Crouch et al. (2023) supply the empirical bridge of the craft: the ECI-TE course improves noticing and interaction quality among toddler educators (N = 61; medium effect), precisely because it forms the professional gaze, not because it automates the score. Status: an RCT finding of professional development. Inference: noticing is a condition of sensitivity; a dashboard does not notice for the teacher.

Finders et al. (2021), with 684 children, show that Emotional Support, Classroom Organization and Instructional Support have differentiated relations with readiness: mean CO with mathematics; CO variability with literacy (negative); mean IS with language (negative in their specifications). Howard et al. (2024) reinforce, across ninety studies, that global indices of interaction quality have little support and that specific dimensions—such as supporting play—show greater consistency, while instructional support has low support even when aligned with cognitive outcomes. Status: empirical and synthesis findings on multidimensionality and limits of IS as a predictor. Inference: when process quality is present in early childhood, there are ES (emotional climate, sensitivity, respect), CO (organization, productivity, behaviour management) and IS (concept development, quality feedback, language modelling) exercised in real time by adults; not a flattened average. Paschall et al. (2023), with 2 114 Head Start children, find a majority profile of positive emotional climate with low instructional support: the dimensions are not bought as a package. von Suchodoletz et al. (2023) meta-analyse that ECEC quality associates with child outcomes, without authorizing reducing that quality to a scoring artefact. NAEYC (2022) requires developmentally appropriate practice: the 3–6-year-old is a subject of relationship, not of a pipeline. Inference: sensitivity, reciprocity and scaffolding are verified in the playground, at the table and in transitions—when the adult sustains serve-and-return—not in the wearable log or the turn-classifier JSON.

6. Case 3. Process interactions are not documentation, nor SEL, nor executive functions

The first category boundary is pedagogical documentation. A digital panel, a portfolio or an anecdotal record may preserve traces of what happened; they do not, by themselves, constitute the quality of process interactions that OECD (2021) anchors in the daily encounter. Status: pedagogical inference anchored in the Starting Strong framework, marked as such. This article does not make documentation a central object: it names it to forbid the equivalence. A system that scores CLASS videos is not meaning-making documentation either: it is automated psychometric estimation. Confusing “we have evidence on screen” with “there is reciprocity in the room” is the category error the market exploits.

The second boundary is SEL. Gómez and Strasser (2021), with 43 dyads, show that conversational turns at 18 months predict socioemotional competencies at 30 months. Gómez-Muzzio and Strasser (2025), with 33 cases, show that turns at 30 months explain between 14.4 % and 20.3 % of socioemotional variance at 77 months. Status: empirical finding on turns and socioemotional development. Pedagogical inference, marked as such and with restrictive use: these studies are retained only to distinguish. A turn classifier may inherit the predictive fame of turns for SEL and sell itself as “interaction quality.” It is not. Process quality in this article is sensitivity, reciprocity and scaffolding in CLASS ES/CO/IS and serve-and-return, not a socioemotional competencies programme nor a count-based SEL proxy. This work does not recycle an SEL article: it uses turn evidence to show the illegitimate leap from automatic counting to a declaration of pedagogical–relational quality. Ferjan Ramírez et al. (2021) reinforce the cut: automatic ≠ manual.

The third boundary is executive functions. This article does not convert interaction quality into inhibitory control, working memory or cognitive flexibility. Paschall et al. (2023) and Finders et al. (2021) are retained for CLASS profiles and dimensions, not for EF. von Suchodoletz et al. (2023) are retained for quality–outcome associations, not for executive training. Inference: when a product promises “improves executive attention because it measures turns” or “develops self-regulation because it scores CLASS,” another construct has been crossed. Process quality may associate empirically with many outcomes; it is not identical with them. Chen (2024), Su and Yang (2022) and Ljungcrantz (2026) map AI in ECE; Miao and Holmes (2023) set age limits and validation. Inference: the only AI use coherent with ages 3–6 remains on the side of the adult who notices, responds and scaffolds—as support for noticing training or deferred professional feedback—under pedagogical validation. It does not enter as an autonomous scorer, a wearable that replaces judgment, or a turn classifier that declares quality fulfilled.

7. Inferential framework: four tests for affirming that process quality exists, not an artefact

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 system that scores CLASS videos, a wearable that prompts “respond now,” or a model that classifies conversational turns constitute pedagogical–relational interaction quality.

7.1. Test of contingent serve-and-return, not of the climate score. The Center on the Developing Child (n.d.) defines the serve–response–expansion sequence. Ramakrishnan et al. (2023) estimate Positive/Negative Climate with moderate r and AUC = 0.70. Inference: evidence of quality is verified in whether the adult noticed the child’s serve and responded with contingency. If the centre’s “evidence” is an ACORN dashboard or a Positive Climate score, the centre has done estimation, not reciprocity.

7.2. Test of Emotional Support, Classroom Organization and Instructional Support in real situations, not of the flattened average. Finders et al. (2021) document differentiated ES/CO/IS relations and within-day variability. Howard et al. (2024) document little support for global indices and low support for IS. Paschall et al. (2023) document profiles with positive climate and low IS. Whitehill and LoCasale-Crouch (2024) reach Pearson up to 0.48 on automated IS. Inference: process quality requires the three dimensions exercised, not a number that collapses them. A model that approximates IS does not demonstrate concept scaffolding at the work table.

7.3. Test of sensitivity and teacher noticing, not of the OTR alert. LoCasale-Crouch et al. (2023) improve noticing and interaction quality through training. Green et al. (2023) increase OTR with bug-in-ear in 2/4 teachers, with low maintenance. Inference: “respond now” may raise academic response opportunities without raising sensitivity to the child’s serve. Sensitivity is reading the other; OTR is prompt frequency.

7.4. Test of category distinction and professional judgment, not of the product catalogue. Ferjan Ramírez et al. (2021) prevent equating automatic turns with manual judgments. Gómez and Strasser (2021) and Gómez-Muzzio and Strasser (2025) are used only to show that turns predict SEL, not that a classifier is process quality. UNESCO (2021), Miao and Holmes (2023), the European Commission (2022), the U.S. Department of Education (2023), NAEYC (2022) and OECD (2021, 2023) require human oversight, pedagogical validation, developmentally appropriate practice and process interactions. Inference: a centre cannot treat the infant as a feature vector for scoring, alerting or counting. Process quality is not fulfilled by scoring the video better. It is fulfilled by responding to the serve, organizing the room with warmth and scaffolding thinking with adults who sustain the craft.

The framework admits the digital when it is subordinated to teacher education and validated professional observation (LoCasale-Crouch et al., 2023; Whitehill and LoCasale-Crouch, 2024, as support for feedback, not as declared quality). It rejects declaring interaction quality by CLASS scorer, alert wearable or turn classifier (Ramakrishnan et al., 2023; Green et al., 2023; Ferjan Ramírez et al., 2021).

8. Discussion

Three tensions organize the discussion. The first is between exhibiting a measurement artefact and exercising process quality. It is a finding that ACORN estimates emotional climate with moderate correlations (Ramakrishnan et al., 2023); that LLMs and BoW approximate Instructional Support with Pearson up to 0.48 (Whitehill and LoCasale-Crouch, 2024); and that bug-in-ear can move OTR for some teachers (Green et al., 2023). It is a framework that quality is played out in daily interactions (OECD, 2021; Center on the Developing Child, n.d.; NAEYC, 2022). It is not a finding that score, alert or classifier produce the craft that LoCasale-Crouch et al. (2023) form as noticing. The three artefacts measure or push what engineering knows how to deliver and declare what only reciprocity would authorize.

The second is between turn counting and pedagogical reciprocity. Ferjan Ramírez et al. (2021) show automatic–manual disagreement. Gómez and Strasser (2021) and Gómez-Muzzio and Strasser (2025) show the predictive value of turns for SEL—restrictive use here. Inference: insisting that the kindergarten “already has interaction quality” because the model counts enough turns is inverted pedagogy. A correlational proxy with socioemotional development is made to stand for Emotional Support, Classroom Organization and Instructional Support. The infant becomes a turn emitter; the adult, a supervisor of the classifier.

The third is between global index and multidimensionality. Howard et al. (2024) and Finders et al. (2021) advise against treating quality as a single number. Paschall et al. (2023) show unbalanced profiles. Inference: the only AI use coherent with ages 3–6 remains on the side of the adult who notices and responds, under pedagogical validation and the guidelines of UNESCO, the European Commission and the U.S. Department of Education. An autonomous scorer because the system needs a product is not that use.

9. Limits

This review is narrative. It does not apply PRISMA or estimate primary combined effects. Ramakrishnan et al. (2023) estimate PC/NC, not the three full CLASS dimensions as lived quality. Whitehill and LoCasale-Crouch (2024) reach correlations below or near human IRR, not pedagogical equivalence. Green et al. (2023) are four Head Start teachers with low maintenance. LoCasale-Crouch et al. (2023) are sixty-one toddler educators in an online-course RCT, not AI scoring. Finders et al. (2021) are 684 in a U.S. context. Howard et al. (2024) synthesize ninety heterogeneous studies. Paschall et al. (2023) are Head Start profiles. Gómez and Strasser (2021) and Gómez-Muzzio and Strasser (2025) are small samples (N = 43 and N = 33) and are used only to distinguish turns/SEL from process quality. Ferjan Ramírez et al. (2021) are parent–infant turns, with marked transfer to the classroom. Chen (2024), Su and Yang (2022) and Ljungcrantz (2026) map AI in ECE, not kindergarten CLASS quality. NAEYC, UNESCO, OECD and the Center on the Developing Child are framework sources. No Latin American AI trials comparing adult-mediated serve-and-return with an autonomous scorer or alert wearable in CENDI settings were located. The inferences in section 7 are pedagogical category hypotheses, not implementation evidence.

10. Conclusions

A system that scores CLASS videos, a wearable that prompts “respond now,” or a model that classifies conversational turns do not constitute pedagogical–relational interaction quality in an early childhood education centre. Verified evidence does not authorize that declaration. ACORN estimates Positive Climate and Negative Climate with r = 0.55 and r = 0.63 on three hundred fifteen-minute segments, with AUC = 0.70 (Ramakrishnan et al., 2023). Whisper, Llama 2 and BoW reach Pearson up to 0.48 on Instructional Support versus human IRR up to 0.55 (Whitehill and LoCasale-Crouch, 2024). Bug-in-ear moves OTR in two of four teachers, with low maintenance (Green et al., 2023). Automatic turn measures are not equivalent to manual ones (Ferjan Ramírez et al., 2021). By contrast, when process quality is present in early childhood, there are sensitivity, reciprocity and scaffolding: serve-and-return (Center on the Developing Child, n.d.); daily interactions across twenty-six countries and fifty-six curricula (OECD, 2021); differentiated and variable CLASS dimensions (N = 684; Finders et al., 2021); noticing formed in toddler educators (N = 61; LoCasale-Crouch et al., 2023); little justification for global indices and low support for IS in a synthesis of ninety studies (Howard et al., 2024); profiles that separate emotional climate from instructional support (N = 2 114; Paschall et al., 2023). Process interactions are distinguished from documentation, from SEL—although turns predict SEL (Gómez and Strasser, 2021; Gómez-Muzzio and Strasser, 2025)—and from executive functions. Current guidelines require process interactions, human oversight, pedagogical validation and not replacing professional judgment (OECD, 2021, 2023; NAEYC, 2022; UNESCO, 2021; Miao and Holmes, 2023; European Commission, 2022; U.S. Department of Education, 2023).

Where the sources do not measure a kindergarten, this article does not assert it. Where they measure scoring, OTR alerts or turn classification, it does not translate them into pedagogical–relational interaction quality. Accompanying girls and boys aged three to six in process quality is exercising sensitivity to the serve, reciprocity in the return and scaffolding in Emotional Support, Classroom Organization and Instructional Support. Everything else is an observation dashboard, a wearable alert and algorithmic turn counting. It is not pedagogical–relational interaction quality, and it must not be presented as what it is not.

NEXTECH.IA Editorial Laboratory / Engineer Mitre.

References

  1. Center on the Developing Child at Harvard University. (s. f.). Serve and return. https://developingchild.harvard.edu/key-concept/serve-and-return/
  2. 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
  3. 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
  4. Ferjan Ramírez, N., Hippe, D. S., y Kuhl, P. K. (2021). Comparing automatic and manual measures of parent–infant conversational turns: A word of caution. Child Development, 92(2), 672–681. https://doi.org/10.1111/cdev.13495
  5. Finders, J. K., Budrevich, A., Duncan, R. J., Purpura, D. J., Elicker, J., y Schmitt, S. A. (2021). Variability in preschool CLASS scores and children’s school readiness. AERA Open, 7. https://doi.org/10.1177/23328584211038938
  6. Gómez, E., y Strasser, K. (2021). Language and socioemotional development in early childhood: The role of conversational turns. Developmental Science, 24(5), e13109. https://doi.org/10.1111/desc.13109
  7. Gómez-Muzzio, E., y Strasser, K. (2025). Conversational turns at early childhood predicts socioemotional development at school age. Social Development, 34(4), e70019. https://doi.org/10.1111/sode.70019
  8. Green, A. L., Olsen, A. A., y Nandakumar, V. (2023). Increasing teacher opportunities to respond in a Head Start program using a bug-in-ear coaching model. Early Childhood Education Journal, 52(6), 1107–1119. https://doi.org/10.1007/s10643-023-01498-4
  9. 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, 6. https://doi.org/10.1007/s10648-023-09832-3
  10. 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
  11. LoCasale-Crouch, J., Romo-Escudero, F., Clayback, K., Whittaker, J., Hamre, B., y Melo, C. (2023). Results from a randomized trial of the Effective Classroom Interactions for toddler educators professional development intervention. Early Childhood Research Quarterly, 65, 217–226. https://doi.org/10.1016/j.ecresq.2023.06.005
  12. Miao, F., y Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535
  13. NAEYC. (2022). Developmentally appropriate practice in early childhood programs serving children from birth through age 8 (4.ª ed.). NAEYC. https://www.naeyc.org/resources/pubs/books/dap-fourth-edition
  14. OECD. (2021). Starting Strong VI: Supporting meaningful interactions in early childhood education and care. OECD Publishing. https://doi.org/10.1787/f47a06ae-en
  15. OECD. (2023). Empowering young children in the digital age (Starting Strong). OECD Publishing. https://doi.org/10.1787/50967622-en
  16. Paschall, K. W., Barnett, M. A., Mastergeorge, A. M., Li, X., y Vasquez, M. B. (2023). A new look at teacher interactional quality: Profiles of individual teacher–child relationship and classroom teaching quality among Head Start students. Early Education and Development, 34(5), 1172–1190. https://doi.org/10.1080/10409289.2022.2094159
  17. Ramakrishnan, A., Zylich, B., Ottmar, E., LoCasale-Crouch, J., y Whitehill, J. (2023). Toward automated classroom observation: Multimodal machine learning to estimate CLASS Positive Climate and Negative Climate. IEEE Transactions on Affective Computing, 14(1), 664–679. https://doi.org/10.1109/TAFFC.2021.3059209
  18. 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
  19. UNESCO. (2021). Recommendation on the ethics of artificial intelligence. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000381137
  20. U.S. Department of Education, Office of Educational Technology. (2023). Artificial intelligence and the future of teaching and learning: Insights and recommendations. U.S. Department of Education. https://www.ed.gov/sites/ed/files/documents/ai-report/ai-report.pdf
  21. von Suchodoletz, A., Lee, D. S., Henry, J., Tamang, S., Premachandra, B., y Yoshikawa, H. (2023). Early childhood education and care quality and associations with child outcomes: A meta-analysis. PLOS ONE, 18(5), e0285985. https://doi.org/10.1371/journal.pone.0285985
  22. Whitehill, J., y LoCasale-Crouch, J. (2024). Automated evaluation of classroom Instructional Support with LLMs and BoWs: Connecting global predictions to specific feedback. Journal of Educational Data Mining, 16(1), 34–60. https://doi.org/10.5281/zenodo.10974824