1. Introduction and problem

In the classroom of three- to six-year-olds a package of three artefacts has been installed that claims to count as social and emotional learning. The first is conversational: a chatbot that “talks about emotions” with the child, or with the educator on the child’s behalf. The second is perceptual: an affect detector that, by camera or microphone, classifies faces, voices or postures into a menu — happiness, sadness, anger, fear, surprise, disgust, neutral — and delivers a dashboard. The third is lexical: an application that names feelings or translates a conflict into a label. All three are visible, auditable and cheap in planning time. They allow the kindergarten to exhibit that it “already does SEL with artificial intelligence.” The enunciative leap is enormous: one moves from classifying, naming or talking to a model to asserting that SEL is in place. That leap is not authorised by the evidence on early-childhood SEL or by the science of affect.

The thesis of this article is restrictive. An emotions chatbot, an affect detector or an application that names feelings does not constitute social and emotional learning. Early-childhood SEL is situated interaction, co-regulation and play; not the classification of a face or dialogue with a model. CASEL defines it as the process through which children and adults acquire and apply knowledge, skills and attitudes to develop healthy identities, manage emotions, feel and show empathy, maintain supportive relationships and make responsible and caring decisions (Chen, Liang & Lin, 2025, citing CASEL, 2020). That process does not occur in front of a camera or inside a chat. It occurs when an educator mirrors negative affect, names what is happening, validates its situational appropriateness and, with a metacognitive prompt, invites the child to imagine how to regulate (Silkenbeumer et al., 2024). It occurs in joint play, a weekly practice of early-childhood centres according to TALIS Starting Strong 2024 (OECD, 2025). It occurs among people who share a conflict, a turn, a material or a wait.

The problem is sharpened by a professional reason that product sheets omit. Starting Strong VI anchors ECEC quality in process quality: everyday interactions with adults, peers, materials and space are the most proximal driver of development, learning and well-being (OECD, 2021). Detector, chatbot and app extract a signal, assign a category and return a report. They may produce a datum. They do not produce co-regulation. There is, in addition, an age threshold. UNESCO’s Guidance for generative AI in education and research proposes a limit — thirteen years, in the 2023 institutional communication — for independent conversations with generative platforms, and requires pedagogical validation (Miao & Holmes, 2023). A four-year-old is not an autonomous user of a model. Treating her as the interlocutor of an emotions chatbot or as the object of an affect detector is not “doing SEL”: it is displacing the educational act toward classification.

This article does not recycle axes already treated in this series. The question is one of pedagogical category: what counts as social and emotional learning when an early-childhood centre “does AI.” The contributions are three: reconstructing the state of the art that separates situated SEL from affective computing; examining three families of cases — a school deployment of emotion recognition, 2025 literature that promises to identify and “intervene” on preschoolers’ emotional states, and what kindergarten classrooms actually do when SEL is present; and offering four tests for deciding when a kindergarten may claim that it educates socioemotionally, and not merely that it classifies, names or chats.

2. State of the art: from situated SEL to the facial label

Four strata that the “socioemotional AI” market tends to mix should be kept apart. The first is early-childhood SEL as social, embodied and situated learning. The second is co-regulation: the adult work of modulating, with body, voice and turn-taking, the state of a child who cannot yet self-regulate. The third is affective computing: systems that infer emotions from biometric or textual data. The fourth is the rights and prohibition framework, which in Europe already treats that inference in educational institutions as an unacceptable practice, save for medical or safety exceptions.

In the SEL stratum, Cipriano and colleagues (2023) offer the widest contemporary synthesis of universal school-based interventions: 424 studies from 53 countries, 252 discrete programmes and 575,361 kindergarten-to-grade-12 students, published between 2008 and 2020. Participants improve, relative to controls, in skills, attitudes, behaviours, school climate and safety, peer relationships, school functioning and academic achievement. Heterogeneity of content, implementation and context moderates effects (Cipriano et al., 2023). That evidence does not authorise treating a detector or a chatbot as “an SEL programme.” The programmes in the meta-analysis are universal school interventions, not classifiers. Hayashi, Liew, Aguilar, Nyanamba and Zhao (2022) further recall that school-based SEL emerged in North America and has not traditionally focused on embodied processes situated in the learner’s context and lived experience. Transferable social-emotional competencies are, they argue, inherently situated and embodied. A menu of six emotions extracted from a frame is not that process.

In the co-regulation stratum, Kostøl and Kovač (2024) define it as warm, receptive and supportive interactions between adult and child that provide guidance and modulation of emotions, behaviours and thoughts. In 24 clips of eight dyads (children aged 2–8), they identify three basic elements: the adult response to an initiative, negotiation, and attunement; the last defines the quality of the other two. Silkenbeumer et al. (2024) take that definition into the ECEC classroom. With video in 19 groups, 48 teachers and 213 children aged 2–6, they find that initial emotion coaching — mirroring the expression of negative affect, labelling the feeling and validating its situational appropriateness — and co-regulation through metacognitive prompts are associated with children’s independent self-regulation. Mänty et al. (2022) design, on the same premise, a 32-week collaborative programme for 450 professionals from 60 centres in seven municipalities in Northern Finland: the object is not an emotion dashboard, but the shared capacity to offer conscious, consistent co-regulation in everyday interactions. Richard, Cavadini, Dalla-Libera, Angonin, Alaria, Lafay, Berger and Gentaz (2025), with 1,285 children aged 3–5, show that emotion comprehension improves with age, and that labelling facial expressions without context remains the most complex task. Inference, marked as such: if even human labelling of decontextualised faces is not the most accessible route at ages 3–5, a system that “names feelings” from a frame cannot be presented as the privileged form of kindergarten SEL.

In the affective-computing stratum, the hypothesis of a reliable, specific mapping between emotional states and facial configurations does not hold with the robustness the product needs. Le Mau, Hoemann, Lyons, Fugate, Brown, Gendron and Barrett (2021) analyse 604 photographs of professional actors posing emotional states in context-rich scenarios. Unsupervised and supervised machine learning find that the actors portray those states with variable facial configurations; only three categories — fear, happiness and surprise — are portrayed with moderate reliability and specificity. Observers’ emotion inferences also vary in a context-sensitive way. Crawford (2021) summarises the underlying scientific disagreement: there is no reliable evidence that AI detects emotions, and yet the emotion-recognition industry was projected toward USD 37 billion by 2026. The pedagogical gesture this article contests — treating classification as SEL — is fed by that industry, not by classroom evidence.

In the rights stratum, paragraph 128 of UNESCO’s Recommendation calls for awareness policies on the anthropomorphisation of AI technologies and on technologies that recognise and mimic human emotions, particularly in robot–human interaction and especially when children are involved (UNESCO, 2021). Paragraph 127 requires that one be able to tell whether one is interacting with a living being or with a system imitating the human, and to refuse that interaction. UNICEF (2021) requires support for development and the best interests of the child. Regulation (EU) 2024/1689, Article 5(1)(f), prohibits placing on the market, putting into service for that specific purpose, or using AI systems to infer emotions in the workplace and in educational institutions, except for medical or safety reasons (European Union, 2024). The February 2025 guidelines specify that educational institutions cover all types and levels, and that a general class “well-being” dashboard does not fall under the medical exception (European Commission, 2025). The 2022 ethical guidelines and the U.S. Department of Education report coincide in not replacing professional judgement or the teacher (European Commission, 2022; U.S. Department of Education, 2023). Inference: the kindergarten cannot treat a four-year-old as a chatbot user or as the object of a detector, nor call SEL what European law already classifies, in its territory, as a prohibited practice.

3. Review method

A critical narrative review was conducted, not a meta-analysis. The purpose was not to estimate a homogeneous effect size across a commercial deployment, computer-vision papers and classroom observations, but to articulate an argument of pedagogical category with verified sources. Inclusion criteria: (a) 2021–2026; (b) SEL, co-regulation or socioemotional teaching in early childhood, or emotion recognition and affective computing in schools or preschools; (c) relevance to kindergarten, preschool, CENDI or ages 3–6; (d) peer-reviewed journal, DOI, or UNESCO, OECD, UNICEF, European Union, European Commission or ministry report; (e) verifiable DOI or publisher page. Axes already used in this series, including leadership and centre direction, were excluded.

The search was executed on 25 August 2026 on DOI pages, Springer, Elsevier, Frontiers, Nature, MDPI, Wiley, IEEE, SPIE, OECD iLibrary, UNESDOC, UNICEF, EUR-Lex, the European Commission and CASEL. Each source was checked against at least one of those pages. The corpus was organised into school deployments of emotion recognition; 2025 systems that promise to identify and intervene; and evidence of SEL and co-regulation in the room.

The analysis distinguished three enunciative statuses. Empirical finding: what was observed or measured in the sample. Conceptual or normative framework: what a framework, a guideline or a regulation prescribes. Pedagogical inference: the translation to kindergartens, preschools, CENDI and infant schools, marked as such. The limits are those of a narrative review (section 9).

4. Case 1. Reading the face is not educating: 4 Little Trees and the forbidden equivalence

Crawford (2021) documents, in Nature, the case that best illustrates the leap the thesis rejects. During the pandemic, technology companies pushed unproven emotion-recognition tools into workplaces and schools. A system called 4 Little Trees, developed in Hong Kong by Find Solution AI, claims to assess students’ emotions while they do classwork. It maps facial features to assign each pupil’s emotional state to a category — happiness, sadness, anger, disgust, surprise, fear — and, in addition, estimates “motivation” and forecasts grades. The CNN report Crawford anchors describes use at True Light College: while students work on tests and homework, the AI measures muscle points on the face via the computer or tablet camera. The founder claimed 85% accuracy in Hong Kong. The number of schools using the system in that jurisdiction rose from 34 to 83 in a year; prices ranged from USD 10 to 49 per student per course. Crawford herself recalls that a 2019 review found no reliable evidence that AI detects emotions, and that Rosalind Picard — a founding figure of affective computing — had called for regulation (Crawford, 2021).

Status of the evidence. Empirical finding of deployment, not of pedagogical efficacy: in 2021, dozens of Hong Kong schools bought a system that classifies faces, estimates motivation and forecasts grades, and presented it as classroom support — even as an improvement on the in-person classroom. It is not a finding from early childhood: True Light College is secondary, and the product was sold mainly in remote teaching. It is not a finding that classification produces co-regulation, empathy, relationship skills or responsible decision-making. It is not a finding of cross-cultural validity: the company itself warned that more ethnically mixed communities would be a greater challenge. Pedagogical inference, marked as such: the gesture a kindergarten copies when it “does SEL with a detector” is exactly this. A facial signal is taken, one of six labels is assigned, and socioemotional education is declared. What there is, is an affect file. CASEL SEL is not a grade forecast from a muscle. It is a process of acquiring and applying knowledge, skills and attitudes in real relationships (Chen et al., 2025). A “happiness” dashboard is not that process.

The scientific saturation is clear. Le Mau et al. (2021) show that even professional actors, given rich scenarios, do not produce the prototypical configurations the detector needs. Richard et al. (2025) show that, for 1,285 preschoolers, labelling facial expressions without context is the hardest of the three tasks they measured. Hayashi et al. (2022) recall that early-childhood SEL is embodied and situated. Inference: a system trained to force faces into six categories — and, in the case of 4 Little Trees, to forecast grades — does not “read” the SEL of a four-year-old. It reads a frame with the Ekman vocabulary the industry adopted because it fitted computer vision (Crawford, 2021). Calling that social and emotional learning is a category error.

5. Case 2. Identifying the preschooler’s emotional state is not pedagogical intervention

In 2025, the engineering literature repeats the 4 Little Trees gesture with the vocabulary of “intervention” and brings it, now, to preschool. Xin, Zhang and Zhang (2025) present, at the MVDL conference, a paper whose stated aim is to use deep learning to identify preschool students’ emotional states accurately and, on that basis, to develop intervention strategies. They build a database of facial expressions, speech and text; classify faces with a CNN, vocal emotional features with an RNN and textual tendencies with a sentiment model; and conclude that automatic recognition has “high accuracy and reliability.” Wang and Dong (2025), affiliated with a preschool-education college in Inner Mongolia, propose an AFCNN for children’s “psychological emotion recognition”: 86.5% accuracy on their CME set, 14.4 points above traditional CNNs, with claims of cross-age generalisation and real-time performance. A UBMK 2025 paper describes an emotion-recognition system “to monitor children’s emotions in preschool institutions”: CNN, training on FER-2013 — a seven-category set built largely on adult faces — 81.32% accuracy in training and validation, and dissemination of results to educators and families via a Django application and a Telegram bot. The abstract claims that the method “enhances continuous observation of psycho-emotional states in kindergartens,” “promotes timely intervention strategies” and “fosters communication” between families and institutions.

Status of the evidence. Laboratory finding of classification: given certain images, voices or texts, a network reaches a hit rate against training labels. It is not an SEL finding. None of those papers observes co-regulation, joint play, process quality or CASEL’s five competencies in a 3–6 classroom. FER-2013 is not a kindergarten corpus. A Telegram bot that alerts the family that the model classified “sadness” is not emotion coaching. Pedagogical inference, marked as such: the word “intervention” in these abstracts performs the same leap the thesis forbids. One moves from a classification hit to a pedagogical strategy. Silkenbeumer et al. (2024) show what an intervention in this field is: a teacher who, after an expression of negative affect, mirrors, names, validates and asks. That happens in the second of the conflict, with the body present. A system that identifies and “disseminates results” produces a report. The report may arrive later, to someone else, detached from the situation. Confusing the report with the intervention is the error.

Araguas, Blanchard, Derégnaucourt, Chopin and Guellai (2025) offer a contrast and a limit. Sixty-two children aged 3–6, randomly assigned to a human demonstrator or a NAO robot, completed body-part, imitation and emotion-recognition-from-posture tasks. There were no main effects of demonstrator type on most tasks; age predicted performance between 3 and 6, with an age × demonstrator interaction on sequential motor imitation, stronger in the human condition. The authors suggest that well-designed social robots could be supplements to embodied learning under specific conditions (Araguas et al., 2025). Status: laboratory finding on labelling and imitation, not classroom SEL. Inference: identifying a “fear” posture on a robot is not co-regulation. A robot that demonstrates labels does not hold a child who is crying because a turn was taken away. UNESCO (2021) asks not to anthropomorphise technologies that recognise and mimic emotions, especially with children, and to allow refusal of the interaction. Scoring well on a task does not elude that warning.

6. Case 3. What the kindergarten does when there is SEL: situated teaching and co-regulation

Chen, Liang and Lin (2025) published in Early Childhood Education Journal (vol. 53, pp. 1521–1537) an observational study in four Hong Kong kindergarten classrooms. The team conducted 20 videotaped sessions of four teachers’ social and emotional teaching (SET) and 71 children’s SEL during whole-group instruction. They coded teacher and child utterances for CASEL’s five competencies — self-awareness, self-management, social awareness, relationship skills and responsible decision-making — and, for each competency, one of four strategies: telling/commanding/directing, explaining, asking questions and affirming/confirming. Both teachers and children evidenced responsible decision-making most frequently. Except for social awareness, teachers’ SET and children’s SEL in the other four competencies correlated significantly and strongly — for example, self-awareness r = 0.90, p < 0.001. Teachers mainly used asking questions; children mainly used telling (92% of their utterances). There were correlations between teachers’ asking and children’s telling, between teachers’ asking and children’s explaining, and between teachers’ explaining and children’s explaining, asking and affirming. The authors interpret that social learning was at work: children observed, modelled and imitated their teachers’ socioemotional teaching in the context of teaching, not in front of an app (Chen et al., 2025).

Status of the evidence. Empirical classroom finding: SEL, when observed in whole-group instruction, appears as a fabric of utterances between teachers and children, with strong concordance in four of five CASEL competencies. It is not an AI finding. It is not a finding that a detector or a chatbot reproduces those concordances. Pedagogical inference, marked as such: this is the object an early-childhood centre may properly call social and emotional learning. It is linguistic and social interaction, situated in a theme — in the qualitative excerpt, sharing — mediated by adult questions and by children’s telling. A face classifier does not ask. A chatbot that names “you are sad” does not wait for the 92% of child utterances Chen et al. record as “telling.” SEL is in the turn, not in the label.

Silkenbeumer et al. (2024) saturate the portrait at the moment of conflict. In 19 ECEC groups, with 48 teachers and 213 children aged 2–6, they analyse episodes in which the teacher intervenes after a negative emotion expression by one or two children. Multilevel results show that initial emotion coaching and co-regulation with metacognitive prompts are associated with independent self-regulation. Coaching and co-regulation are also systematically associated with episode characteristics, especially emotion quality and intensity. The impact statement is precise: ECEC teachers play a significant role in supporting self-regulation; mirroring, labelling and validating, and asking the child for ideas to regulate, promote self-regulation of negative emotions (Silkenbeumer et al., 2024). Mänty et al. (2022) translate that evidence into a 32-week programme for 450 professionals: the adult’s learning object is to offer conscious, consistent co-regulation in everyday interactions, not to operate a dashboard. Kostøl and Kovač (2024) recall that adult attunement is the element that defines the quality of initiative and negotiation. OECD (2021) situates those interactions as process quality. OECD (2025) situates joint play among the weekly practices of ECEC leadership.

Status of the evidence. Empirical finding: SEL from ages 2 to 6, when measured in the room, is co-regulation situated in the emotion episode, socioemotional teaching in classroom talk, and — as context, not as the object of this article — joint play. Framework: CASEL 5; process quality; co-regulation as a precursor of self-regulation. Inference: labelling a feeling can be a step of emotion coaching, but only when the educator does it in the episode, mirrors and validates, and opens a prompt. Extracted from that triad and handed to a model or an app, labelling ceases to be coaching and becomes classification. Cipriano et al. (2023) show that universal school SEL programmes — not classifiers — are associated with improvements. Confusing a CNN hit with those programmes is, again, the leap.

7. Inferential framework: four tests for claiming SEL, not a classifier

The framework that follows is this article’s pedagogical inference, 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 an emotions chatbot, an affect detector or an app that names feelings constitutes social and emotional learning.

7.1. Test of situated interaction, not of the frame. Chen et al. (2025) show SEL as concordance of utterances in teaching. OECD (2021) shows process quality as everyday interactions. Hayashi et al. (2022) recall that early-childhood SEL is embodied and situated. Inference: evidence of SEL is verified in the turn, the conflict, the shared material and joint play, not in a facial crop. If the “evidence” that the centre does SEL is a six-emotion dashboard or a chat log, the centre has done product management, not social and emotional education.

7.2. Test of co-regulation, not of the label. Silkenbeumer et al. (2024) decompose effective intervention: mirror, name, validate, ask. Kostøl and Kovač (2024) place attunement as the quality criterion. Mänty et al. (2022) train 450 professionals to offer that attunement consciously. Inference: naming a feeling is a means, not the end, and only counts inside a triad of adult presence. A detector that names without mirroring or validating, or a chatbot that names without being in the conflict, fails the test by construction. Decontextualised labelling is, moreover, the hardest route at ages 3–5 (Richard et al., 2025) and lacks the reliable mapping the product assumes (Le Mau et al., 2021).

7.3. Test of rights and of the age threshold, not of the platform user. Miao and Holmes (2023) set an age threshold for independent conversations with generative platforms. UNESCO (2021) requires not anthropomorphising technologies that recognise and mimic emotions, especially with children, and allowing one to identify and refuse interaction with a system that imitates the human. UNICEF (2021) requires the best interests of the child. Regulation (EU) 2024/1689 prohibits inferring emotions in educational institutions, with medical or safety exceptions that a class “well-being” dashboard does not meet (European Union, 2024; European Commission, 2025). Inference: an early-childhood centre cannot treat a 3–6-year-old as an autonomous interlocutor of an emotions chatbot or as the object of an affect detector. The thirteen-year threshold is not “met” by lowering the user’s age. It is respected by not deploying independent conversation or biometric inference. In European territory, the detector in the room is not a pedagogical innovation: it is, save a narrow exception, a prohibited practice.

7.4. Test of the SEL programme, not of the network’s hit rate. Cipriano et al. (2023) synthesise 252 universal school interventions. Xin et al. (2025) and Wang and Dong (2025) report classification accuracies. Crawford (2021) documents a product that also forecasts grades. Inference: evidence of SEL is not an F1 or an 86.5%. It is a process of teaching and of relationship, with quality implementation. A computer-vision paper that “develops interventions” from a CNN, or a bot that flags “sadness,” is not that process.

The framework admits naming feelings as a step of emotion coaching (Silkenbeumer et al., 2024); universal SEL programmes with implementation evidence (Cipriano et al., 2023); and, as a supplement and without anthropomorphising, a robot in embodied tasks (Araguas et al., 2025; UNESCO, 2021). It refuses to declare SEL on the basis of a chatbot, a detector in the room, a labelling app outside the episode, a classification hit on FER-2013 or another bank of adult faces, or treating the early-childhood child as a user of generative models (Crawford, 2021; European Union, 2024; Xin et al., 2025; Wang & Dong, 2025; Miao & Holmes, 2023).

8. Discussion

Three tensions organise the discussion. The first is between classifying and educating. It is a finding that schools bought a system that maps faces to six emotions, estimates motivation and forecasts grades (Crawford, 2021); that in 2025 systems are published that identify preschoolers’ emotional states and promise intervention (Xin et al., 2025; Wang & Dong, 2025); and that the science of affect does not authorise the reliable mapping those systems need (Le Mau et al., 2021). It is a framework that inferring emotions in educational institutions is prohibited in the European Union save for narrow exceptions (European Union, 2024). It is not a finding that classifying a face produces CASEL’s five competencies. The policy of the three artefacts — chatbot, detector, app — measures what engineering knows how to measure and declares what only situated SEL would authorise.

The second is between the label and the coaching triad. Silkenbeumer et al. (2024) show that labelling counts when it goes with mirroring, validation and a prompt, in the episode, with the teacher present. Chen et al. (2025) show that kindergarten SEL lives in adult asking and in children’s telling. Richard et al. (2025) show that labelling faces without context is the hardest task at ages 3–5. Extracting the label, automating it and returning it by Telegram or chat inverts the sequence: it produces the name of the affect without the relationship that makes it educational. In early childhood, the name without the relationship is not emotional literacy. It is a detached vocabulary.

The third is between the artefact that imitates emotions and the adult who co-regulates. Araguas et al. (2025) find that, in the laboratory, demonstrator type does not change most scores; age does. UNESCO (2021) asks not to anthropomorphise and to allow refusal of the interaction. Miao and Holmes (2023) ask for an age threshold. An emotions chatbot performs, by design, that anthropomorphisation: it speaks as if it felt and offers itself as a regulation partner to someone who still needs an adult body. Mänty et al. (2022) and Kostøl and Kovač (2024) locate that partner in human attunement. OECD (2021, 2025) locates development in interactions and in joint play. A model does not play. A detector does not attune.

9. Limits

This review is narrative. It does not apply PRISMA or estimate pooled effects. 4 Little Trees is a deployment in secondary and remote teaching, not a kindergarten trial; transfer to ages 3–6 is inference. Xin et al. (2025) and Wang and Dong (2025) are classification papers; the UBMK 2025 paper is cited by verified DOI and abstract, without inflating authors not recovered on the publisher page. Araguas et al. (2025) is a NAO laboratory study, not curricular SEL. Chen et al. (2025) observes four classrooms in whole-group instruction. Silkenbeumer et al. (2024) and Mänty et al. (2022) are co-regulation without AI. Cipriano et al. (2023) cover K–12. Le Mau et al. (2021) use professional actors. Richard et al. (2025) measure emotion comprehension, not AI products. UNESCO, UNICEF, the European Union and the 2022–2023 guidelines are prescriptive. No Latin American trials of SEL and AI in early childhood education were located with the same degree of DOI. The inferences in section 7 are hypotheses of pedagogical category, not implementation evidence.

10. Conclusions

An emotions chatbot, an affect detector or an application that names feelings does not constitute social and emotional learning in an early-childhood centre. The verified evidence does not authorise that declaration. Schools that bought an emotion-recognition system classified faces, estimated motivation and forecast grades; that is affective surveillance, not SEL (Crawford, 2021). 2025 papers reach classification accuracies and call sending a report an “intervention”; that is engineering, not co-regulation (Xin et al., 2025; Wang & Dong, 2025). In four kindergarten classrooms, SEL appears as concordance of utterances between teachers who ask and children who tell (Chen et al., 2025). In 19 ECEC groups, emotion coaching and metacognitive prompts are associated with independent self-regulation (Silkenbeumer et al., 2024). Co-regulation is formed in 32 weeks with 450 professionals, not on a dashboard (Mänty et al., 2022). Emotion comprehension at ages 3–5 stumbles on the labelling of faces without context (Richard et al., 2025). Professional actors do not produce the prototypical faces the detector assumes (Le Mau et al., 2021). Current law requires an age threshold, not anthropomorphising technologies that recognise and mimic emotions, the best interests of the child and, in the European Union, a prohibition on inferring emotions in educational institutions (Miao & Holmes, 2023; UNESCO, 2021; UNICEF, 2021; European Union, 2024).

Where the sources do not measure a kindergarten, this article does not assert one. Where they measure classification or the laboratory, it does not translate them into SEL. Educating three- to six-year-olds socioemotionally is sustaining situated interaction, co-regulation and play. The rest is a file of labels. It is not SEL, and it must not be presented as what it is not.

Laboratorio Editorial de NEXTECH.IA / Ingeniero Mitre.

References

  1. Araguas, A., Blanchard, A., Derégnaucourt, S., Chopin, A., & Guellai, B. (2025). Body knowledge and emotion recognition in preschool children: A comparative study of human versus robot tutors. Behavioral Sciences, 16(1), 29. https://doi.org/10.3390/bs16010029
  2. Chen, J. J., Liang, X., & Lin, J. C. (2025). Weaving the fabric of social and emotional learning in the context of teaching: A study in Hong Kong kindergarten classrooms. Early Childhood Education Journal, 53(5), 1521–1537. https://doi.org/10.1007/s10643-024-01682-0
  3. Cipriano, C., Strambler, M. J., Naples, L. H., Ha, C., Kirk, M., Wood, M., Sehgal, K., Zieher, A. K., Eveleigh, A., McCarthy, M., Funaro, M., Ponnock, A., Chow, J. C., & Durlak, J. (2023). The state of evidence for social and emotional learning: A contemporary meta-analysis of universal school-based SEL interventions. Child Development, 94(5), 1181–1204. https://doi.org/10.1111/cdev.13968
  4. Crawford, K. (2021). Time to regulate AI that interprets human emotions. Nature, 592, 167. https://doi.org/10.1038/d41586-021-00868-5
  5. Development of an emotion recognition system to monitor children’s emotions in preschool institutions. (2025). In 2025 10th International Conference on Computer Science and Engineering (UBMK). IEEE. https://doi.org/10.1109/UBMK67458.2025.11206893
  6. 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
  7. European Commission. (2025). Commission guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act). https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-prohibited-artificial-intelligence-ai-practices-defined-ai-act
  8. Hayashi, A., Liew, J., Aguilar, S. D., Nyanamba, J. M., & Zhao, Y. (2022). Embodied and social-emotional learning (SEL) in early childhood: Situating culturally relevant SEL in Asian, African, and North American contexts. Early Education and Development, 33(5), 746–763. https://doi.org/10.1080/10409289.2021.2024062
  9. Kostøl, E. M. F., & Kovač, V. B. (2024). Co-regulation in dyadic parent–child relationships: A video analysis of well-functioning interactions. Early Child Development and Care, 194(1), 72–86. https://doi.org/10.1080/03004430.2023.2282357
  10. Le Mau, T., Hoemann, K., Lyons, S. H., Fugate, J. M. B., Brown, E. N., Gendron, M., & Barrett, L. F. (2021). Professional actors demonstrate variability, not stereotypical expressions, when portraying emotional states in photographs. Nature Communications, 12, 5037. https://doi.org/10.1038/s41467-021-25352-6
  11. Mänty, K., Kinnunen, S., Rinta-Homi, O., & Koivuniemi, M. (2022). Enhancing early childhood educators’ skills in co-regulating children’s emotions: A collaborative learning program. Frontiers in Education, 7, 865161. https://doi.org/10.3389/feduc.2022.865161
  12. Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535
  13. OECD. (2021). Starting Strong VI: Supporting meaningful interactions in early childhood education and care. OECD Publishing. https://doi.org/10.1787/f47a06ae-en
  14. OECD. (2025). Results from TALIS Starting Strong 2024: Strengthening early childhood education and care. OECD Publishing. https://doi.org/10.1787/20af08c0-en
  15. Richard, S., Cavadini, T., Dalla-Libera, N., Angonin, S., Alaria, L., Lafay, A., Berger, C., & Gentaz, E. (2025). The development of specific emotion comprehension components in 1285 preschool children. Scientific Reports, 15, 8562. https://doi.org/10.1038/s41598-025-90613-z
  16. Silkenbeumer, J., Lüken, L. M., Holodynski, M., & Kärtner, J. (2024). Emotion socialization in early childhood education and care – How preschool teachers support children’s emotion regulation. Social and Emotional Learning: Research, Practice, and Policy, 4, 100057. https://doi.org/10.1016/j.sel.2024.100057
  17. UNESCO. (2021). Recommendation on the ethics of artificial intelligence. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000381137
  18. UNICEF. (2021). Policy guidance on AI for children 2.0. UNICEF Innocenti. https://www.unicef.org/innocenti/reports/policy-guidance-ai-children
  19. European Union. (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence. Official Journal of the European Union, L 2024/1689. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  20. U.S. Department of Education, Office of Educational Technology. (2023). Artificial intelligence and the future of teaching and learning: Insights and recommendations. https://www.ed.gov/sites/ed/files/documents/ai-report/ai-report.pdf
  21. Wang, F., & Dong, J. (2025). The application of improved AFCNN model for children’s psychological emotion recognition. Scientific Reports, 15, 24138. https://doi.org/10.1038/s41598-025-10269-7
  22. Xin, S., Zhang, W., & Zhang, T. (2025). Research on emotional state identification and intervention strategies for preschool students using deep learning techniques. In Proceedings of SPIE 13689, International Conference on Machine Vision and Deep Learning (MVDL 2025), 136890N. https://doi.org/10.1117/12.3071992