1. Introduction and problem

In early childhood and basic education schools a promise circulates that is symmetrical to that of “personalization”: artificial intelligence “frees” the teacher, “automates” the administrative and “returns time” to the relationship with young children. That promise usually silences its condition of possibility. To automate, a workflow that can be extracted is needed, a prompt that can be written, an output that can be corrected, and an institution that can demand the same product—or more products—in less time. Time is not an empty vessel that the machine fills. It is the substance of the craft: planning, assessing, communicating with families, sustaining the simultaneity of a classroom of four-year-olds. If AI reorganizes those tasks without reducing institutional demand, the risk is not idleness: it is intensification.

The problem is not hostility to the tool. It is confusion among three objects. The first is an empirical finding: what is observed when teachers are asked about stress, time pressure or genAI use; what planning minutes a workload diary records; what primary teachers who plan with ChatGPT perceive. The second is a normative framework: what UNESCO, the OECD or a ministry prescribe when they ask for human agency, training or “workload reduction.” The third is a pedagogical inference: what should not be sold as well-being even if marketing says “ally against burnout.” Mixing them produces a pedagogy of unmeasured liberation: the school “uses AI” without being able to say whether the teacher sleeps more, whether the classroom gains presence, or whether only the type of after-hours work has changed.

The thesis of this article is restrictive. There is not sufficient evidence to claim that AI reduces burnout or “returns time to the classroom” in early childhood and basic education. It can reorganize teachers’ work and, at times, intensify it. Where the source does not measure burnout, this article does not assert it. Where it does not measure time recovered for presence with children, it does not promise liberation. The gap that organizes the work is intensification: Apple (1986) conceptualized how an economistic rationality compresses the craft; Ballet and Kelchtermans (2009) showed, in Flemish primary schools, that that experience is mediated by professional identity and by school organization. GenAI enters that history as an apparent resource and as a new demand: one must know how to use it, verify it, justify it and, often, produce more with it.

This article does not treat the UNESCO teacher competency framework (19 August morning), nor classroom privacy or datafication (21 August, 1:02), nor literacy for students, nor inclusion by disability, nor parental mediation, nor play or PopBots, nor adaptive tutors as an axis of pupil learning, nor gaps in “learning AI,” nor multilingualism (21 August, 5:02). The object is teachers’ work: load, automation of planning, assessment and communication, and the risk that AI does not free time but increases demands.

2. State of the art: intensification, demands-resources and the promise of automating the craft

It is useful to separate three strata. The first is conceptual: what it means to intensify teachers’ work. The second is normative: which frameworks fix well-being, agency and “ethical use” of AI. The third is empirical: what has been documented on load, stress and genAI among early childhood and basic-education teachers, not on pupil learning as the axis.

In the conceptual stratum, Apple (1986) formulated the intensification thesis: the economistic perspective on education compresses time, multiplies tasks and threatens to deskill professional judgment. Hargreaves (1992) analyzed that thesis from teachers’ time and warned that intensification is not a simple clock. Ballet, Kelchtermans and Loughran (2006) and Ballet and Kelchtermans (2008, 2009) refined it with qualitative studies in primary schools: there are multiple sources (macro, meso and self-imposed); the impact is not linear; it is mediated by school organization and by the teacher’s interpretative framework; not everyone experiences the same reform equally. Finding from those Flemish cases: not every change “adds hours” automatically; the experience of intensification passes through professional sense-making and through collective norms—including that of “being willing to innovate”—which can support and, at the same time, overload. That framework does not measure ChatGPT; it is retained because it describes the mechanism that genAI can reactivate: an innovation presented as relief and lived as a new productivity norm.

In the same stratum, Bakker and Demerouti (2017) consolidate job demands-resources (JD-R) theory: demands erode health if there are no buffering resources. Collie, Martin and Gasevic (2024) apply that grammar to genAI (school support as resource; time pressure as demand). The model does not guarantee that software is a resource; it allows one to ask whether genAI reduces the demand or becomes a new demand (learning it, verifying it, being accountable with it).

In the normative stratum, UNESCO and the International Task Force on Teachers for Education 2030 (2024) estimate that 44 million additional primary and secondary teachers are needed by 2030 and group well-being and working conditions as push factors toward leaving the profession. The report mentions digital evolution, including AI, as part of the shortage context; it cites that, in a UNESCO survey of more than 450 schools and universities, fewer than 10% had developed institutional policies or formal guidance on generative AI applications. It is not a burnout trial with ChatGPT: it is a framework of shortage and conditions. Miao and Holmes (2023) orient a human-centred use of genAI in education and research. Miao and Cukurova (2024) define fifteen teacher AI competencies—a framework this article cites and does not reopen as its axis—and declare principles of protecting teachers’ rights, human agency and sustainability. UNESCO’s (2021) Recommendation on the ethics of AI elevates proportionality and human oversight. The United Kingdom Department for Education (2023/2025), in its position on genAI, and the report by The Open Innovation Team and Department for Education (2024) formulate that genAI could reduce administrative load and “free” time to teach. Status: prescription or policy hypothesis. It is not a finding that that liberation has been measured in early childhood education.

In the empirical stratum, the gap is precisely the one that organizes this article. There is an international survey of stress and AI use (OECD, 2025a, 2025b). There is a study of motivation and genAI integration under demands and resources (Collie et al., 2024). There are qualitative primary-school cases (Calleja and Camilleri, 2025; Özsongür and Çeliktürk Sezgin, 2026). There is a randomized trial of planning minutes in secondary science (Roy, Poet, Staunton, Aston, and Thomas, 2024). There are analyses of ChatGPT outputs for lesson plans, including first grade (van den Berg and du Plessis, 2023; Cooper, 2023; Powell and Courchesne, 2024). There is an article whose title promises burnout relief and whose method tests ChatGPT’s capabilities, not exhaustion scales in teachers (Hashem, Ali, El Zein, Fidalgo, and Abu Khurma, 2024). None of those objects, by itself, authorizes the sentence “AI reduces burnout in early childhood education.”

3. Review method

A critical narrative review was conducted, not a meta-analysis. The purpose was not to estimate a homogeneous effect size of “AI and teacher well-being”—non-existent in a comparable way among an international survey, a structural equation model, a three-teacher case study and a minutes-diary trial—but to articulate the intensification thesis with verified sources. Inclusion criteria were: (a) focus on teachers’ work in early childhood education, primary, or, as explicit saturation, basic/lower secondary when the study measures teachers’ time or stress and not pupil learning as the axis; (b) preferred publication 2021–2026, with a justified exception for still-current frameworks (Apple, 1986; Hargreaves, 1992; Ballet et al., 2006; Ballet and Kelchtermans, 2008, 2009; Bakker and Demerouti, 2017); (c) documentary type of journal or proceedings with DOI, OECD or UNESCO report or registered trial by an education agency, or ministerial position on an official portal; (d) access to a DOI, editorial page or official URL confirming authors, year, title, sample and scope. Excluded as nuclear cases were the UNESCO teacher competency framework (cited as a framework), classroom privacy/datafication, literacy for students, inclusion by disability, parental mediation, play/PopBots, adaptive tutors as a learning axis, gaps in “learning AI,” and multilingualism.

The search was executed on 21 August 2026 (around 09:05, America/Mexico_City) on DOI pages, ScienceDirect, Emerald, PLOS, MDPI, Springer, EEF/NFER, OECD iLibrary, UNESDOC, GOV.UK, DergiPark and institutional repositories. Each cited source was verified against at least one of those pages. Discarded, for not being confirmed as burnout evidence among early childhood/basic-education teachers, were non-scientific higher-education surveys, tertiary literature reviews without well-being measurement, and a preprint or conference communication alleging percentage reductions in planning not contrasted here with a journal DOI. The corpus was organized into three nuclear cases: the international structure of stress and AI use; time pressure as a demand that predicts valuing genAI; and planning with genAI among primary teachers. The EEF/NFER trial and the first-grade plan analyses are used as saturation of measured time and of verification work, not as proof of well-being in early childhood education.

The analysis distinguished three enunciative statuses. Empirical finding: what was observed in the sample or the questionnaire. Normative framework: what an organization prescribes. Pedagogical inference: the translation to the early childhood and basic classroom, marked as such. The limits are those of any narrative review (section 9).

4. Case 1. The load already there: TALIS 2024 and the early childhood workforce

Before asking whether AI “frees,” one must fix from what it would free. TALIS 2024 (OECD, 2025a) is the international survey of teachers and principals; the comparable core is, habitually, lower secondary education, not kindergarten. The report Results from TALIS 2024: The State of Teaching states that around a third of teachers use artificial intelligence in their work, and that seven in ten worry that it facilitates plagiarism and cheating. Nine in ten declare job satisfaction. The Insights booklet (OECD, 2025a) specifies that, on average, only around one in five teachers say they experience stress “a lot” in their work, and that around half experience little or no stress; in Alberta (Canada), Australia, Bahrain, Costa Rica, Malta and New Zealand more than 30% report stress “a lot.” Among those who use AI, most employ it to plan lessons and to learn about or summarize topics; a quarter employ it to assess or mark. Among users, on average, 73% use it to learn about and summarize topics efficiently and 69% to generate lesson plans. Training on AI is highest in Singapore (76%) and lowest in France (9%). Around half of teachers believe AI should not be used in teaching; one in ten report school policies that ban it. On the OECD average, 40% report that too much marking is a source of stress. The Sweden note (OECD, 2025a) places the OECD average of AI use at 36% and, among non-users, 75% without knowledge or skills to teach with it; 10% say the job impacts their mental health “a lot” and 8% their physical health.

Status. Empirical finding: there is heterogeneous AI adoption, concern about plagiarism, stress sources linked to marking and to the accumulation of responsibilities (discipline, curriculum, administration), and a massive training deficit. Finding that TALIS does not produce: that those who use AI report less burnout, fewer hours or more classroom time. The Insights text formulates the attraction (“the software claims it can assess hundreds of tests in seconds”) and, in the next breath, the absence of robust large-scale evidence that AI improves student learning. To infer from adoption a relief of the craft is, therefore, a leap: it is the argument of “supporters,” not a questionnaire result.

For early childhood education, the relevant instrument is TALIS Starting Strong 2024 (OECD, 2025b), administered to staff and leaders in early childhood education and care in 17 countries or subnational entities (15 in pre-primary ISCED 02; 8 in settings for children under three). The report documents high satisfaction and, at the same time, persistent stress. The percentage reporting work-related stress ranges from under 20% in pre-primary in Colombia and Israel to over 75% in Germany at both levels. Frequent sources: having too many tasks at the same time and performing extra duties because of staff absences—the latter especially in settings for children under three. Staff with more stress are more likely to consider leaving because of mental or physical health or for another sector. Leaders identify absences and shortages as the main barrier to quality environments. Status. Empirical finding of load in ECEC: simultaneity, substitution, shortage. Finding that Starting Strong 2024 does not produce in the pages verified here: that AI reduces those sources. Pedagogical inference, marked as such: the core of work in early childhood education is not drafting a 5E plan on a keyboard; it is sustaining bodies, play, care and multiple demands at once. Automating a planning text does not, by itself, touch the most reported stress source in several systems: doing too many things at the same time in the presence of young children.

UNESCO and the International Task Force (2024) anchor the problem in shortage: 44 million primary and secondary teachers toward 2030; well-being and conditions are push factors. It is a macro diagnostic framework, not a ChatGPT measurement. Marked inference: introducing AI into a system that already loses teachers to load, without evidence of relief, is an institutional wager, not a treatment of burnout.

5. Case 2. Time pressure and valuing genAI: 368 Australian teachers

Collie, Martin and Gasevic (2024) published in Computers and Education: Artificial Intelligence a longitudinal study over one school term with 368 teachers in Australian schools (56% in primary, 40% in secondary, 4% at both levels). The framework is JD-R: resource (in-school support to apply genAI) and demands (time pressure, disruptive behavior) predict motivation (self-efficacy and valuing of genAI) and, in turn, integration (in teaching-related work and in student learning activities). At Time 2, 241 teachers remained (35% loss). Mean time pressure was high (5.84 on a scale whose wording includes “I feel pressed for time in my job”); mean school support for genAI was comparatively low (2.62).

Empirical finding: school support was associated with greater genAI self-efficacy and greater genAI valuing. Time pressure was associated with greater genAI valuing, not with lower subsequent pressure. Disruptive behavior was not significantly associated with genAI motivation or integration. Self-efficacy predicted both types of integration; valuing predicted only integration in the teacher’s own work (planning, preparing), not in student activities. The authors interpret that time pressure may function as a “lifeline”: facing untenable loads, a relatively easy and often free tool is valued. Limitations the article itself declares: it did not examine the effectiveness of genAI tools; it did not nest schools; one term does not cover the year. Status. It is a finding of prediction of use and of valuing under demand. It is not a finding of reduced burnout, of hours worked, or of classroom time. On the contrary: the empirical mechanism compatible with the data is that demand (pressure) drives adoption, not that adoption dissolves demand.

Pedagogical inference, marked as such: if in a primary and secondary system time pressure is already high and institutional support for genAI is low, the injunction “use it to free yourselves” reproduces the innovation norm that Ballet and Kelchtermans (2008) saw as paradoxical: it relieves external pressure and, at the same time, adds a productivity standard. Collie et al. (2024) do not measure that paradox as burnout; they make it thinkable. Where the early childhood teacher must, in addition, care, watch and sustain play, genAI enters as extra-school cognitive work—prompt, trimming, adaptation to ages the model did not see—on a day already simultaneous (OECD, 2025b).

6. Case 3. Primary teachers who plan with genAI: perceptions, not burnout scales

Calleja and Camilleri (2025) published in the International Journal for Lesson and Learning Studies a multiple-case study with three lesson-study teams in two primary schools in Malta. The pedagogical focus was Maltese (sentence structure, vocabulary, written advertisement) in Years 4 and 5. The teams integrated tools—ChatGPT, image and character generation, Canva, CapCut, word banks—into the cycle of planning, observation and discussion. Nuclear data are email interviews, at the end of the school year (June–July 2024), with the three teachers who taught the lessons, plus each team’s reports. The lens is the Technology Acceptance Model (perceived usefulness, ease of use, intention). Empirical finding: situated use in lesson study improved perceptions and attitudes toward AI and the agency to keep learning. The teachers reported initial nervousness, anxiety about licence limitations, a login described as time-consuming, and the conviction that benefits outweighed challenges. One teacher (Helen, Year 5) was the only person on team A with prior AI experience in planning; for the rest, the tool was new. Limitation the authors declare: small sample, not generalizable. Status. It is a finding of perception and acceptance in a professional-development device. It is not a finding of minutes saved, nor of burnout, nor of early childhood education (the ages are 8–10 years). Lesson study, moreover, adds meetings: the “innovation that relieves” is inscribed in a format that demands collective time.

Özsongür and Çeliktürk Sezgin (2026) published in Educational Academic Research a basic qualitative study with 27 primary teachers (sınıf öğretmenleri) in Istanbul. Face-to-face semi-structured interviews, between 11 December 2023 and 19 January 2024 (average duration close to fourteen minutes), analysed by content. Empirical finding of discourse: teachers report that genAI tools facilitate access to resources, content creation and personalized materials; that they reduce administrative work and, therefore, “allow” more time for teaching; that they support classroom management and speed up feedback; and that training and technical support are needed. Several extracts insisted on “saving time” and “less effort.” Status. It is a finding of self-reported perception, not of a workload diary or a burnout inventory. The authors recommend applied longitudinal studies precisely because they do not claim to have measured impact. The sample is young (nine teachers aged 20–25; ten aged 26–30; eight aged 30 or more) and voluntary. Inference, marked: the narrative of relief is a sense-making datum, not proof that the working day has shortened. In Ballet and Kelchtermans’s (2009) thesis, professional sense-making mediates intensification: what is lived as help can, institutionally, become a new norm of materials production.

As saturation of the object “first-grade plan,” not as a teacher sample, Powell and Courchesne (2024) examined in PLOS ONE an exploratory case study: a series of prompts to ChatGPT for a first-grade science plan on heredity, aligned to the Massachusetts curriculum and the 5E model. Empirical finding: within about thirty minutes a formally aligned plan could be generated and refined; iterations included questionable components, missing details and a fake resource. Van den Berg and du Plessis (2023) analysed ChatGPT plans for a grade-6 prepositions lesson and concluded that the tool is a “starting point, not a final product.” Cooper (2023), in a science-education self-study, warned of the risk of situating ChatGPT as an “ultimate epistemic authority.” Status. It is a finding of quality and of verification work on model outputs, not of teacher well-being. Pedagogical inference: each fake resource, each bias and each missing detail is minutes that do not appear in the story of a “plan in thirty minutes.” That invisible work—checking, adapting to age, translating to an early childhood classroom, discarding the unsuitable—is a cognitive demand. It is not measured in Özsongür et al. (2026) or Calleja and Camilleri (2025); Powell and Courchesne (2024) make it visible in the artefact.

As saturation of measured time—not of early childhood education—Roy et al. (2024) evaluated for the Education Endowment Foundation a school-randomized trial with 259 Year 7 and 8 science teachers in 68 schools in England (summer 2024; ten weeks). The ChatGPT group received an implementation guide; the comparison group was not to use genAI for those lessons. Primary result: 56.2 weekly minutes of preparation versus 81.5 in the non-genAI group (mean saving of 25.3 minutes; 31%; high security). An expert panel, blind to origin, found no quality difference in a sample of resources. Teachers in the ChatGPT group used it modestly (a third of lessons; one or two activities, mainly questions, quizzes and ideas). Crucial finding for this article: “Teachers reported typically using the time saved to complete other lesson and resource planning (LRP) or teaching tasks, or to reduce overall workload.” Status. It is a finding of minutes in a fragment of the secondary timetable, not of burnout, not of early childhood education, not of “time returned to the classroom” as presence with young children. Part of the saving is reallocated to more planning. The report itself warns that most applied the approach only to part of their timetable.

7. Inferential framework: five tests before speaking of well-being

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 five tests. If a proposal of “AI for teacher well-being” in early childhood or basic education does not pass them, it is not announced as relief.

7.1. Test of the measured variable. Collie et al. (2024) measure self-efficacy, valuing and integration. Calleja and Camilleri (2025) measure TAM perception. Özsongür and Çeliktürk Sezgin (2026) measure interview themes. Roy et al. (2024) measure minutes of a planning block. TALIS (OECD, 2025a, 2025b) measures stress, sources and AI use separately. Hashem et al. (2024) measure ChatGPT outputs evaluated with the 5E model, not exhausted teachers. Inference: reduction of burnout is not asserted if there is no inventory (Maslach or other), health follow-up or, at least, total hours. The title of Hashem et al. (2024)—“teacher ally for workload relief and burnout prevention”—is a rhetorical frame; the method is a test of the model’s capacity. Confusing them is the error this article refuses to repeat.

7.2. Test of reallocation. Roy et al. (2024) document that “saved” time is typically used on more planning or other teaching tasks, or to reduce overall load: all three at once, without the trial isolating the third. Inference: a local saving (Year 7 science quizzes) is not liberation of the craft. In early childhood education, if twenty minutes of drafting a plan are saved and twenty are added to verify age-appropriateness, communicate with families or produce variants, the day did not open: it was redistributed. Apple (1986) and Ballet and Kelchtermans (2009) describe precisely that qualitative compression: not only more hours, but more density.

7.3. Test of verification work. Powell and Courchesne (2024) find a fake resource in a first-grade plan. Van den Berg and du Plessis (2023) demand critical examination. Cooper (2023) names the risk of epistemic authority. Inference: automating the first page does not automate judgment. That judgment is time. In early childhood education, judgment includes safety, play, situated language and what a model trained on adult corpora does not see. Miao and Holmes (2023) and Miao and Cukurova (2024) prescribe human agency: as a framework, not as evidence that that agency is cheap in minutes.

7.4. Test of the new demand. Collie et al. (2024) show low institutional support and high pressure. TALIS (OECD, 2025a) shows that three in four non-users lack knowledge or skill, that training ranges from 9% to 76% by country, and that half reject use in teaching. Calleja and Camilleri (2025) document nervousness, licences and logins. Özsongür and Çeliktürk Sezgin (2026) ask for training. UNESCO and the International Task Force (2024) record that fewer than 10% of more than 450 institutions had a genAI policy. Inference: “using AI” is an additional competency (Miao and Cukurova, 2024), a source of role anxiety and a possible criterion of teacher evaluation. That is a JD-R demand, not an automatic resource (Bakker and Demerouti, 2017).

7.5. Test of presence in early childhood education. TALIS Starting Strong (OECD, 2025b) locates stress in simultaneity and in substitution, not in a shortage of planning text. Inference: a tool that writes plans, rubrics or messages to families does not replace the teacher’s body in the classroom, nor the ratio, nor the missing substitute. Promising that genAI “returns time to the classroom” when the classroom already requires being in several places at once is a desktop metaphor. What is non-delegable in early childhood teacher well-being is the density of presence, not the speed of the first draft.

Operationally, the framework admits: (a) workload diaries and burnout measures before celebrating relief; (b) local tools with real support (the resource in Collie et al., 2024); (c) protected time to verify outputs; (d) not turning adoption into new performativity. It rejects: (e) selling ChatGPT as burnout prevention without measuring burnout; (f) extrapolating secondary science minutes to kindergarten; (g) taking perceptions of “less effort” as real hours; (h) adding automated family communication without counting the extra message channel that automation enables.

8. Discussion

Three tensions organize the discussion. The first is between declared relief and measured demand. The Open Innovation Team and Department for Education (2024) collect interviews and grey literature according to which genAI could transform education, save time and personalize; the DfE position (2023/2025) states that, used appropriately, it can reduce load and free time for teaching. It is a policy framework. Roy et al. (2024) offer the cleanest finding of minutes, and even so they bound it: secondary, a fragment of the timetable, frequent reallocation to more planning, no burnout measure. Collie et al. (2024) show the inverse mechanism to the ministerial narrative: the tool is valued because there is pressure; it is not documented that pressure falls. Hashem et al. (2024) illustrate the field’s rhetorical slip: an explanatory design on English, science and mathematics plans for ages 12–18 is titled as burnout prevention. Discussing that article as well-being evidence would falsify its method.

The second tension is between automating an artefact and sustaining a craft. Van den Berg and du Plessis (2023), Cooper (2023) and Powell and Courchesne (2024) agree that the generated plan requires critique. Calleja and Camilleri (2025) show teachers who become excited about planning and, at the same time, uneasy about licences. Özsongür and Çeliktürk Sezgin (2026) collect the hope of “more time to teach” without timing it. In early childhood education, the artefact (plan, worksheet, message to the family) is a fraction of the work. Contemporary intensification may not resemble Apple’s (1986) on the surface—it is no longer only ministerial paperwork—and yet does in its logic: more visible products (variants, differentiation, immediate communication) with the same body. Ballet and Kelchtermans (2009) warned that the collective norm of innovating intensifies. GenAI is, today, that norm with a conversational interface.

The third tension is between a use survey and a health survey. TALIS (OECD, 2025a) can say, at once, that a third use AI and that a fifth have a lot of stress, without articulating both items in a causal model. Starting Strong (OECD, 2025b) can say that ECEC staff are stressed by doing too many things at once, without there being, in the pages verified, a module that attributes relief to a chatbot. UNESCO and the International Task Force (2024) can ask that the profession be valorized and, at the same time, record the void of genAI policies. Miao and Cukurova (2024) can require human agency. None of that is evidence that automation has returned presence to a classroom of three-year-olds. The leap from “it is used for planning” to “one is better” is the one this article cuts.

A note remains on communication with families, announced in this work’s title. TALIS (OECD, 2025a) includes, among uses, generating text for feedback or communication with guardians, at lower frequencies than planning in several systems. No nuclear source measures whether that automation reduces or increases message volume, the expectation of immediate reply, or the emotional labour of the school-home relationship. Inference, marked: opening a cheaper channel for producing text may increase the demand for interaction, not decrease it. To claim the contrary would be to invent a result.

9. Limits

This review is narrative. It does not apply a full PRISMA protocol nor estimate combined effects on burnout. TALIS 2024 is not, in its core, early childhood education; Starting Strong 2024 is not, in the pages used, an AI study. Collie et al. (2024) mix primary and secondary and do not measure effectiveness or burnout. Calleja and Camilleri (2025) interview three Year 4–5 teachers. Özsongür and Çeliktürk Sezgin (2026) offer perceptions of 27 volunteers, with brief interviews, in a Turkish university journal with a verified DOI: the status is local qualitative, not a trial. Roy et al. (2024) is the most robust time design and is secondary. Powell and Courchesne (2024), van den Berg and du Plessis (2023), Cooper (2023) and Hashem et al. (2024) analyse model outputs or capabilities, not kindergarten teacher cohorts. Apple (1986), Hargreaves (1992) and Ballet and Kelchtermans (2009) are intensification frameworks, not ChatGPT evidence. UNESCO and DfE texts are prescriptive or policy. The geography is biased toward the OECD, Australia, England, Malta, Türkiye and the Emirates in Hashem’s title; a Latin American trial of burnout and genAI in early childhood education that met the criteria was not located with the same degree of verifiable openness. That absence is a gap, not proof of non-existence. The inferences of section 7 are hypotheses of an ethical-labour threshold, not evidence of national implementation. This article does not claim that genAI cannot, in some future design, reduce hours; it claims that, in the verified corpus, neither reduction of burnout nor return of time to the early childhood and basic classroom was demonstrated.

10. Conclusions

Artificial intelligence in early childhood and basic education does not, today, have sufficient evidence to be sold as a teacher well-being policy. Three families of verified sources sustain the argument. The TALIS and TALIS Starting Strong surveys document stress, load sources and AI adoption without articulating a relief effect (OECD, 2025a, 2025b). Collie, Martin and Gasevic’s (2024) study shows that time pressure predicts valuing genAI among primary and secondary teachers, not that genAI dissolves that pressure. The primary-school cases in Malta and Istanbul (Calleja and Camilleri, 2025; Özsongür and Çeliktürk Sezgin, 2026) collect perceptions of planning support and requests for training, not burnout scales; Powell and Courchesne (2024) make visible the work of verifying a first-grade plan; Roy et al. (2024) measure secondary-school minutes that, in part, are reallocated to more planning.

The frameworks are not mute. UNESCO and the International Task Force (2024) locate well-being in the retention of a profession in shortage. Miao and Holmes (2023) and Miao and Cukurova (2024) mandate human agency and competencies, which are themselves work. Bakker and Demerouti (2017) recall that a “resource” that does not buffer demand is not a resource. Apple (1986) and Ballet and Kelchtermans (2009) recall that innovation can densify the craft. The DfE (2023/2025) may expect relief: that expectation is not a finding. Where the sources do not measure burnout or classroom time, this article does not invent it. The risk is not that the machine writes a plan. The risk is that the institution reads that plan as proof that the teacher has already been freed—and then asks for more.

Ingeniero Mitre / Laboratorio Editorial de NEXTECH.IA

References

  1. Apple, M. W. (1986). Teachers and texts: A political economy of class and gender relations in education. Routledge.
  2. Bakker, A. B., and Demerouti, E. (2017). Job demands-resources theory: Taking stock and looking forward. Journal of Occupational Health Psychology, 22(3), 273–285. https://doi.org/10.1037/ocp0000056
  3. Ballet, K., and Kelchtermans, G. (2008). Workload and willingness to change: Disentangling the experience of intensification. Journal of Curriculum Studies, 40(1), 47–67. https://doi.org/10.1080/00220270701516463
  4. Ballet, K., and Kelchtermans, G. (2009). Struggling with workload: Primary teachers’ experience of intensification. Teaching and Teacher Education, 25(8), 1150–1157. https://doi.org/10.1016/j.tate.2009.02.012
  5. Ballet, K., Kelchtermans, G., and Loughran, J. (2006). Beyond intensification towards a scholarship of practice: Analysing changes in teachers’ work lives. Teachers and Teaching, 12(2), 209–229. https://doi.org/10.1080/13450600500467415
  6. Calleja, J., and Camilleri, P. (2025). Primary school teachers’ perceptions towards the use of generative AI in teaching using lesson study. International Journal for Lesson and Learning Studies, 14(3), 237–252. https://doi.org/10.1108/IJLLS-11-2024-0268
  7. Collie, R. J., Martin, A. J., and Gasevic, D. (2024). Teachers’ generative AI self-efficacy, valuing, and integration at work: Examining job resources and demands. Computers and Education: Artificial Intelligence, 7, 100333. https://doi.org/10.1016/j.caeai.2024.100333
  8. Cooper, G. (2023). Examining science education in ChatGPT: An exploratory study of generative artificial intelligence. Journal of Science Education and Technology, 32, 444–452. https://doi.org/10.1007/s10956-023-10039-y
  9. Department for Education. (2023/2025). Generative artificial intelligence (AI) in education. GOV.UK. https://www.gov.uk/government/publications/generative-artificial-intelligence-in-education
  10. Hargreaves, A. (1992). Time and teachers’ work: An analysis of the intensification thesis. Teachers College Record, 94(1), 87–108. https://doi.org/10.1177/016146819209400110
  11. Hashem, R., Ali, N., El Zein, F., Fidalgo, P., and Abu Khurma, O. (2024). AI to the rescue: Exploring the potential of ChatGPT as a teacher ally for workload relief and burnout prevention. Research and Practice in Technology Enhanced Learning, 19, Article 23. https://doi.org/10.58459/rptel.2024.19023
  12. Miao, F., and Cukurova, M. (2024). AI competency framework for teachers. UNESCO. https://doi.org/10.54675/ZJTE2084
  13. Miao, F., and Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535
  14. OECD. (2025a). Results from TALIS 2024: The State of Teaching. OECD Publishing. https://doi.org/10.1787/90df6235-en
  15. OECD. (2025b). Results from TALIS Starting Strong 2024: Strengthening Early Childhood Education and Care. OECD Publishing. https://doi.org/10.1787/20af08c0-en
  16. Özsongür, M. Z., and Çeliktürk Sezgin, Z. (2026). Primary school teachers’ experiences with generative artificial intelligence in education: Opportunities and limitations. Educational Academic Research, 61, 11–20. https://doi.org/10.33418/education.1714789
  17. Powell, W., and Courchesne, S. (2024). Opportunities and risks involved in using ChatGPT to create first grade science lesson plans. PLOS ONE, 19(6), e0305337. https://doi.org/10.1371/journal.pone.0305337
  18. Roy, P., Poet, H., Staunton, R., Aston, K., and Thomas, D. (2024). ChatGPT in lesson preparation: A Teacher Choices trial. Evaluation report. Education Endowment Foundation and National Foundation for Educational Research. https://educationendowmentfoundation.org.uk/projects-and-evaluation/projects/choices-in-edtech-using-generative-ai-chatgpt-for-ks3-science-lesson-preparation-2024-teacher-choices-trial
  19. The Open Innovation Team and Department for Education. (2024). Generative AI in education: Educator and expert views (RR1402). Department for Education. https://www.gov.uk/government/publications/generative-ai-in-education-educator-and-expert-views
  20. UNESCO. (2021). Recommendation on the ethics of artificial intelligence. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000381137
  21. UNESCO and International Task Force on Teachers for Education 2030. (2024). Global report on teachers: Addressing teacher shortages and transforming the profession. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000388832
  22. van den Berg, G., and du Plessis, E. (2023). ChatGPT and generative AI: Possibilities for its contribution to lesson planning, critical thinking and openness in teacher education. Education Sciences, 13(10), 998. https://doi.org/10.3390/educsci13100998