1. Introduction and problem
In the 3-to-6-year span a package of three artefacts has been installed that claim to count as free play. The first is a “gamified” app: a path of levels, rewards and screens that the vendor describes as “playful learning” or “play-based.” The second is a social robot that directs turns, issues Simon Says prompts, asks the child to draw or asks “can you take turns?” The third is a model—generative or recommender—that suggests “playful activities” aligned to a goal, an indicator or a visible product. All three are visible, auditable and cheap in coordination time. They allow a setting to exhibit, before supervisors, networks or families, that it “already does free play with artificial intelligence.” The enunciative leap is large: one moves from sequencing, instructing or recommending to asserting that free play is present. That leap is authorised neither by evidence on the play spectrum nor by evidence on the right to play.
The thesis of this article is restrictive. A gamified app, a robot that directs turns or a model that suggests playful activities is not free play. Free play is the child’s initiative, open temporality and the absence of an imposed product; not an AI-optimised sequence. General Comment No. 17, reconstructed for early childhood services, fixes two indicia that children themselves recognise as play: choice and autonomy (Colliver and Doel-Mackaway, 2021). On Nesbitt et al.’s (2023) spectrum, free play is child-initiated and child-directed, without an explicit learning goal; guided play keeps the child’s agency, but the adult initiates with an objective; direct instruction is adult-initiated and adult-directed. An app level, a robot turn or a “playful activity” prompt often occupy the cell of instruction or of games with adult-set rules. They do not occupy the cell of free play.
The problem is sharpened by a professional reason that product sheets omit. The right to play—Article 31 of the Convention on the Rights of the Child—requires space, time, acceptance and a rights-informed approach (Lott, 2025). Starting Strong VI anchors ECEC quality in process quality: everyday interactions with adults, peers, materials and space are the most proximal engine of development (OECD, 2021). TALIS Starting Strong 2024 shows that facilitating play—responding to non-verbal invitations, letting a child play alone when deeply absorbed, allowing the child to take the lead—is a process practice; involving the child in plans for the day, which demands more agency, is less frequent (OECD, 2025). An app extracts a sequence, assigns an achievement and returns a report of “minutes of play.” It can produce a datum. It does not produce free play.
This paper does not recycle axes already treated in this series. The question is one of pedagogical category: what counts as free play when an early childhood setting “does AI.” The contributions are three: to reconstruct the state of the art that separates free play, guided play and instruction from the rhetoric of “playful learning”; to examine three families of cases—apps and tablets that shape play, robots that instruct and direct turns, and evidence of what the kindergarten does when free play is present; and to offer four tests for deciding when a setting may claim free play, and not merely a sequence, an instructor robot or a suggestion model.
2. State of the art: from the play spectrum to the optimised sequence
Four strata that the “AI for play” market tends to mix should be kept apart. The first is free play as activity initiated, controlled and structured by children, non-compulsory, intrinsically motivated and undertaken for its own sake. The second is the play spectrum: free play, guided play, games with rules and direct instruction, distinguished by who initiates, who directs and whether there is an explicit learning goal. The third is the artefact: “educational” apps, tablets, social robots and models that recommend or sequence. The fourth is the rights and AI-ethics frame, which treats play as a right and the 3–6-year-old as a subject who is not an autonomous user of a generative platform.
On the construct stratum, Colliver and Doel-Mackaway (2021) argue, on research with children themselves, that choice and autonomy are universal and essential indicia for an activity to be lived as play. Without them, the setting’s “programmed play” does not fulfil Article 31. Lott (2025) operationalises the right with four components—space, time, acceptance and a rights-informed approach—and recalls that the Committee on the Rights of the Child states that implementing the right to play is, by definition, the child’s best interests. Inference, marked as such: a setting that compresses play time into an app session with a visible product does not “digitise” Article 31. It replaces it.
On the spectrum stratum, Nesbitt et al. (2023) place free play at one end: the child initiates and directs, with no explicit goal, as in building a pillow fort. Guided play occurs when the adult has a clear objective and supports the child to reach it, keeping child agency. Playful instruction and direct instruction close the continuum. Skene et al. (2022) meta-analyse 39 interventions (17 in meta-analysis; N = 3,893; ages 1–8) and find that guided play outperforms direct instruction on early maths (g = 0.24), shape knowledge (g = 0.63) and task switching (g = 0.40), and free play on spatial vocabulary (g = 0.93; three studies, combined n = 137). Differences were not identified for other key outcomes. Inference: the finding does not authorise calling guided play “free play,” nor treating an AI sequence as guided play. Guided play requires the child’s agency. An app that constrains the path and a robot that issues the prompt do not conserve that agency by being “fun.”
On the artefact stratum, Meyer et al. (2021) show that the most downloaded “educational” apps score low on the four pillars of learning. Samuelsson, Price and Jewitt (2022) show that iPad play is less ludic than play with non-digital artefacts. Neumann et al. (2023) show a NAO robot as instructor of Simon Says and drawing. Kim et al. (2024) show a robot that asks whether there is agreement, whether turns can be taken and what comes next. Flatebø, Tran, Wang and Bongo (2024) map 29 studies of robots with infants and toddlers: most are laboratory experiments in a dyad. Inference: the “AI for play” literature produces, above all, controlled tasks. Confusing a task with free play is the category error this article names.
On the rights and ethics stratum, the Recommendation on the Ethics of AI requires human oversight, proportionality and particular attention when children are involved (UNESCO, 2021). UNICEF (2021) requires prioritising the best interests of the child and supporting development. The 2022 Ethical Guidelines and the U.S. Department of Education report agree on not replacing professional judgement or the teacher (European Commission, 2022; U.S. Department of Education, 2023). Miao and Holmes (2023) set an age threshold for independent conversations with generative platforms—thirteen—and require pedagogical validation. The global report on early childhood care and education locates the right to a strong foundation in environments, educators and families, not in a product indicator (UNESCO and UNICEF, 2024). Inference: a four-year-old is not the user of a model that “suggests the play.” The child is the subject of a right that the model does not sustain.
3. Review method
A critical narrative review was conducted, not a meta-analysis. The purpose was not to estimate a homogeneous effect size among a store app, an instructor robot and a practice of letting children play, but to articulate an argument of pedagogical category with verified sources. Inclusion criteria: (a) 2021–2026; (b) free or unstructured play, guided play, the play spectrum, or the right to play in early childhood; (c) apps, tablets, social robots or recommender models with relevance to ages 3–6, kindergarten, preschool or CENDI; (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 were excluded, including PopBots and genAI scaffolding as object, SEL/emotion chatbots, adaptive tutors, parental mediation, inclusion/disability, privacy/datafication as axis, and family–school communication.
The search was run on 25 August 2026 on DOI pages, Springer, Elsevier, PLOS, Frontiers, Taylor & Francis, Wiley, Oxford Academic, OECD iLibrary, UNESDOC, UNICEF, ERIC and PubMed/PMC. Each source was verified against at least one of those pages. The corpus was organised into apps and tablets; robots that instruct or direct turns; and evidence of free play as practice and as right.
The analysis distinguished three enunciative statuses. Empirical finding: what was observed or measured in the sample. Conceptual or normative framework: what a framework, guide or regulation prescribes. Pedagogical inference: the translation to kindergartens, preschools, CENDI and infant schools, marked as such. Limits are those of a narrative review (section 9).
4. Case 1. The “educational” app and the tablet that shape play are not free play
Meyer, Zosh, McLaren, Robb, McCaffery, Golinkoff, Hirsh-Pasek and Radesky (2021) publish in Journal of Children and Media the study that best illustrates the first artefact. They operationalise the four pillars of learning—active learning, engagement in the process, meaningful learning and social interaction—on a 0-to-3 scheme per pillar. They analyse 100 children’s “educational” apps with the most downloads on Google Play and the App Store, plus 24 apps most used by preschoolers in a longitudinal cohort (N = 124). They define as lower-quality those summing to 4 or less. Overall scores are low across all pillars. A score of 1 is most frequent on pillar 1: 81 of 124 apps (65%) show structural constraints and a low degree of cognitive effort. Seventy-two apps (58%) fall into the lower-quality category. Free apps score worse on engagement (M = 1.05 vs 1.70; p < .0001) and on the total (M = 4.31 vs 5.32; p < .0047), because of distracting enhancements. Pillar 4—social interaction—is the weakest.
Status of the evidence. Empirical finding of design quality: apps the market calls educational do not, in the main, sustain active learning, undistracted engagement, meaning or social interaction. It is not a finding of free play. The scheme does not measure child initiative, open temporality or absence of product; it measures learning pillars in an artefact. Pedagogical inference, marked as such: the gesture a kindergarten copies when it “does free play with a gamified app” is exactly this. A path of levels is taken, an achievement is assigned, and play is declared. What is present is a sequence with a product. Sixty-five percent of the apps constrain the path. That constraint is the opposite of the choice Colliver and Doel-Mackaway (2021) locate as an indicium of play. A child who taps the next menu item has not initiated the context or directed the course. The child has executed the design.
Samuelsson, Price and Jewitt (2022) saturate the portrait in the classroom. In two preschools, with groups of 2-year-olds and 4–5-year-olds, they index 98 play activities with iPads and with non-digital artefacts (toys, dress-up) using Bird and Edwards’s digital play framework. iPad play is characterised as less ludic than play with other artefacts and diverges from age-typical play norms. With the tablet, epistemic play predominates—exploring menus, functions, the camera, interface problem-solving—and pretend and fantasy recede. Samuelsson (2023), in a seven-month ethnography with a robot in two groups (ages 1–2 and 3–5; n = 38), shows the useful contrast: when children recruit the artefact into their own play types—Hughes’s sixteen—the object enters the child’s play; when the artefact imposes the path, play folds to the artefact. Inference: a store app is not play because the child touches it. It is play if the child initiates, directs and can leave without having “finished the level.” Meyer et al. (2021) show that commercial design is not built for that.
5. Case 2. A robot that instructs Simon Says or directs turns is not free play
Neumann, Koch, Zagami, Reilly and Neumann (2023) publish in Early Childhood Research Quarterly the case that best illustrates the second artefact. They observe 35 English-speaking children (M = 4.60 years) in two tasks—Simon Says and iPad drawing—under the guidance of a NAO social robot or a human instructor. They measure behavioural, emotional and verbal engagement. Behavioural engagement and positive emotional engagement are higher with the human than with the robot across both tasks. Children utter more words with the human in drawing; there is no difference in Simon Says. Engagement and utterances correlate positively in both conditions. In a parallel observational study, Neumann, Neumann and Koch (2023) record 40 preschoolers (M = 4.58 years) with NAO in a drawing activity: 83% take part, 60% talk to the robot, 10% neither talk nor attempt the task. The robot succeeds in instructing and engaging most children. Not all.
Status of the evidence. Empirical finding of engagement in a directed task: preschoolers can follow prompts from an instructor robot; engagement is, overall, higher with a human. It is not a finding of free play. Simon Says is a game with rules given by the adult or the robot. Drawing, in these protocols, is a prompt (“draw / write your name / follow the instruction”). The child does not initiate the context or direct the course. Pedagogical inference, marked as such: the gesture a setting copies when it “puts a robot in so the children play” is often this. A humanoid is taken, a turn sequence is programmed, and play is declared. What is present is instruction mediated by a plastic body. That the child smiles does not convert the prompt into free play. Colliver and Doel-Mackaway (2021) require choice and autonomy. A NAO Simon Says does not offer them.
Kim, Hwang, Lim, Cho and Lee (2024) show the robot as director of collaboration. In four triadic sessions—two children and the robot Skusie, presented as a friend from another planet—the robot mediates “playful learning” interactions. There are conversational and tablet-assisted activities. Skusie reminds: “Do you agree? Can you take turns? What’s next?” They observe five friendship categories (liking, togetherness, parity, agreement, co-construction). Both mediations contribute, complementarily, to friendship behaviours. Status: finding of robotic mediation of collaboration in a session design. It is not a finding of free play. The robot initiates, regulates the turn and asks what comes next. That “what’s next” is the seal of the sequence. Inference: directing the turn may be a collaboration intervention. It is not thereby unstructured play. Free play does not need an agent to ask what comes next. The child decides, changes or leaves.
Flatebø et al. (2024) saturate the methodological portrait: of 29 studies with robots and children aged 2 to 35 months, most are quantitative, experimental, laboratory, one-to-one, with physically present or virtual robots. The concepts are animacy, action understanding, imitation and early conversational skills. Young children can learn from contingent, interactive robots in some situations, not always. Inference: the laboratory that places a robot in front of a toddler does not observe free play. It observes a task. Torpegaard, Knudsen, Linnet, Skov and Merritt (2022) offer the contrast: 214 preschoolers in free play with a cardboard robot that facilitates play with DUPLO. Children treat it as a social agent and as a material object; four social-interaction themes and five play patterns emerge. When the robot does not direct the turn and the child can use it as a thing, the artefact enters the child’s play. That resembles the recruitment Samuelsson (2023) describes more than the instructor NAO of Neumann et al. (2023). Inference: the robot is not, in itself, free play or its negation. Who initiates, who directs and whether a product is imposed decide.
6. Case 3. What the kindergarten does when free play is present: initiative, time and no product
OECD (2025) publishes, with TALIS Starting Strong 2024, the system portrait that best anchors, in this corpus, what a play practice in early education actually is. Staff report how often, in the last complete week, they facilitate play and plan in a child-centred way. Facilitating play is more common than planning that demands agency. Responding positively to non-verbal invitations to play is among the three most frequent practices in almost all systems, especially with children under three. Letting a child play alone if deeply absorbed is common in most, though it ranks among the three least common in pre-primary in Chile, Colombia, Morocco, Norway, Spain and Türkiye. Allowing children to take the lead when the adult plays with them reaches, in pre-primary, 90% daily in Ireland and 54% in Finland. Involving children in plans for the day is, systematically, less frequent. OECD (2021) had already placed those interactions—with adults, peers, materials and space—as process quality, the most proximal mechanism of development.
Status of the evidence. Finding of ECEC staff self-reported practices: the craft of the kindergarten, when play is taken seriously, appears as responding, not interrupting absorption and, less often, yielding initiative over the day’s plan. It is not an AI finding. It is not a finding that an app reproduces those practices. Pedagogical inference, marked as such: this is the object an early childhood setting may properly call a condition of free play. It is the child’s initiative, a temporality that is not cut because “the level is unfinished,” and the absence of a product the adult or the model demand. An app classifier does not ask whether the child is absorbed. A robot that says “what’s next” interrupts absorption. A model that suggests the day’s activity does not involve the child in the plan: it replaces the child.
Veraksa, Veresov, Sukhikh, Gavrilova and Plotnikova (2024) close the experimental contrast. One hundred and thirty-six older preschoolers (5–7 years), equalised on baseline executive functions, attend 14 sessions of 20–30 minutes by play type: role play (free, adult-directed, child-directed), games with rules, digital play and control. Post-tests are given at the end and at four months. Role play and games with rules show a sustainable positive effect on executive functions; digital play, a lasting result only for inhibition. Long-term effects of role play and games with rules exceed the control. Status: finding that not all “play” produces the same development, and that digital play does not replicate the effect of role play. It is not a finding of generative AI. Inference: if the setting “plays” by substituting role play with an app, it is not offering an equivalent. It is changing category.
Martín-García and Rico-González (2024) compel an honesty the thesis does not evade. In a review of 16 randomised trials, free play by itself is not supported as superior to other methods for physical activity, cognitive competence and socioemotional competence; combined with others, it may foster motor competence. Skene et al. (2022) had already shown that guided play beats free play on spatial vocabulary, not on the remaining key outcomes. Status: finding that free play is not, in RCTs, a wildcard of academic superiority. Inference, marked as such: the thesis of this article is not that free play always wins a test. It is that an app, an instructor robot or a suggestion model are not free play. One may value guided play—with an adult present, a clear goal and child agency—and still refuse to call a sequence “free play.” Confusing the cells of the spectrum is the error. OECD (2025), Colliver and Doel-Mackaway (2021) and Lott (2025) locate the evidence of play in absorption, choice, time and space, not in an app’s AUC.
7. Inferential frame: four tests for claiming free play, not a sequence
The frame that follows is pedagogical inference of this article, anchored in the cases and in 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 gamified app, a robot that directs turns or a model that suggests playful activities constitute free play.
7.1. Test of the child’s initiative, not the model’s. Nesbitt et al. (2023) place free play as child-initiated and child-directed. Colliver and Doel-Mackaway (2021) require choice and autonomy. OECD (2025) measures whether staff let children take the lead and involve them in the day’s plan. Meyer et al. (2021) show apps that constrain the path. Kim et al. (2024) show a robot that asks what comes next. Inference: evidence of free play is verified in who opens the context and who can change it. If the “evidence” that the setting offers free play is a log of levels or a prompt of activities, the setting has done sequence, not play.
7.2. Test of open temporality, not the product session. Lott (2025) places time—not only space—as a component of the right to play. OECD (2025) records the practice of not interrupting a deeply absorbed child. Samuelsson et al. (2022) show that the tablet folds play to menus and functions. Neumann et al. (2023) time Simon Says and drawing tasks. Inference: free play does not end because the level was completed or because the robot exhausted the script. If the child’s time is cropped to the app’s, the setting has done session management, not open temporality.
7.3. Test of the absence of an imposed product, not the gamified achievement. Nesbitt et al. (2023) reserve free play for the cell without an explicit learning goal. Skene et al. (2022) show that guided play—with a goal—is another cell, useful for some learning, not identical to free play. Meyer et al. (2021) score apps whose design pushes the next item. Veraksa et al. (2024) show that digital play does not replicate the long-term effect of role play. Inference: a sticker, a star, a “level completed” or a drawing demanded by NAO are products. Free play may produce objects; it must not owe them. If the setting evaluates play by the app’s product, it has inverted the category.
7.4. Test of the right and of age, not of the platform user. Lott (2025) requires space, time, acceptance and a rights approach. UNESCO (2021) and UNICEF (2021) require human oversight and the best interests of the child. European Commission (2022) and U.S. Department of Education (2023) require not replacing the teacher. Miao and Holmes (2023) place at thirteen the threshold for independent conversation with generative AI and ask for pedagogical validation. UNESCO and UNICEF (2024) locate the strong foundation in environments and educators. Inference: an early childhood setting cannot treat the 3–6-year-old as the user of a model that “proposes the play.” The age threshold is not “met” by lowering the age of the classified child. It is respected by not using the generative suggestion as if it were the child’s initiative. An instructor robot and a levels app do not become free play because the vendor calls them playful.
The frame admits guided play as a distinct cell, with adult, goal and agency (Skene et al., 2022; Nesbitt et al., 2023); the robot or digital object when the child recruits them into the child’s own play (Samuelsson, 2023; Torpegaard et al., 2022); practices of not interrupting absorption and of yielding the lead (OECD, 2025); and role play and games with rules as forms with their own effects, not interchangeable with digital play (Veraksa et al., 2024). It refuses to declare free play on the basis of a constrained gamified app, a robot that instructs Simon Says or directs the turn, a model that suggests the day’s playful activity, or treating the early-years child as the user of an optimised sequence (Meyer et al., 2021; Neumann et al., 2023; Kim et al., 2024; Miao and Holmes, 2023).
8. Discussion
Three tensions organise the discussion. The first is between sequencing and playing. It is a finding that 58% of 124 “educational” apps fall into lower quality and that 65% constrain the path (Meyer et al., 2021); that iPad play is less ludic than play with artefacts (Samuelsson et al., 2022); that engagement with a human instructor exceeds that with NAO in Simon Says and drawing (Neumann et al., 2023); and that a robot can direct turns and ask what comes next (Kim et al., 2024). It is framework that free play is initiative, time and absence of product, and that the right requires choice, autonomy, space and time (Colliver and Doel-Mackaway, 2021; Nesbitt et al., 2023; Lott, 2025). It is not a finding that a sequence produces the process OECD (2025) describes as letting children play. The policy of the three artefacts—app, robot, model—measures what engineering knows how to measure and declares what only the child’s play would authorise.
The second is between guided play and free play. Skene et al. (2022) and Martín-García and Rico-González (2024) prevent a naïve apology: free play does not, by itself, win every RCT of learning or physical activity. Guided play wins on some academic outcomes. Inference: that evidence serves to design guided play with adults present, not to relabel an app as free play. Guided play conserves the child’s agency. Meyer et al.’s (2021) app and Neumann et al.’s (2023) NAO do not demonstrate that agency; they demonstrate a path and a prompt. OECD (2021) had warned that the lever is process quality. The “playful activities” model does the inverse operation: it extracts play from the process and returns it as a list.
The third is between the artefact as thing and the artefact as director. Samuelsson (2023) and Torpegaard et al. (2022) show what happens when the child recruits the robot: the object enters the child’s play types, used as material and as character. Neumann et al. (2023) and Kim et al. (2024) show what happens when the robot directs: there is a task, a turn and “what’s next.” Inference: the useful question for a kindergarten is not “do we have a robot or an app?” It is “who initiates, who directs, is there a product, is the time the child’s?” Veraksa et al. (2024) add that changing play type changes development. Substituting role play with digital play is not an update. It is a category change the setting must name as such, not as “free play with AI.”
9. Limits
This review is narrative. It does not apply PRISMA or estimate combined effects. Meyer et al. (2021) evaluate app design, not Latin American 3–6 classrooms; transfer to kindergarten is inference. Samuelsson et al. (2022) compare 98 activities in two preschools; the size is small. Neumann et al. (2023) and Kim et al. (2024) observe mediated tasks and sessions, not free play; N = 35 and few-session designs limit generalisation. Flatebø et al. (2024) cover 2–35 months, not the 3–6 span as sole focus. Torpegaard et al. (2022) use a cardboard robot, not a generative model. Samuelsson (2023) is ethnography of a robot in two Swedish groups. Veraksa et al. (2024) include ages 5–7 and a digital-play arm that is not generative AI. Skene et al. (2022) meta-analyse interventions 1977–2020; several base studies predate 2021. Martín-García and Rico-González (2024) review RCTs of physical and competence free play, not of AI. OECD (2025) is staff self-report, not observation of AI. Colliver and Doel-Mackaway (2021) and Lott (2025) are rights frameworks. UNESCO, UNICEF, the European Commission and the 2022–2023 guides are prescriptive. Miao and Holmes (2023) do not evaluate kindergartens. Latin American trials of generative models that sequence free play at ages 3–6 were not located at the same DOI standard. Section 7 inferences are hypotheses of pedagogical category, not implementation evidence.
10. Conclusions
A gamified app, a robot that directs turns or a model that suggests playful activities do not constitute free play in an early childhood setting. Verified evidence does not authorise that declaration. One hundred and twenty-four “educational” apps score low and, in the main, constrain the path; that is product design, not play (Meyer et al., 2021). iPad play is less ludic than play with artefacts (Samuelsson et al., 2022). A NAO that instructs Simon Says and drawing obtains less engagement than a human; that is a directed task, not the child’s initiative (Neumann et al., 2023). A robot that asks whether turns can be taken and what comes next mediates collaboration; it does not open a temporality without product (Kim et al., 2024). Digital play does not replicate, at four months, the effect of role play on executive functions (Veraksa et al., 2024). By contrast, when free play is present, there is choice and autonomy (Colliver and Doel-Mackaway, 2021); there is space, time and acceptance (Lott, 2025); there is an adult who does not interrupt absorption and sometimes yields the lead of the plan (OECD, 2025); there is a spectrum that does not allow calling guided play free, nor instruction guided (Nesbitt et al., 2023; Skene et al., 2022). Current law requires human oversight, the best interests of the child, and not treating the early-years child as an autonomous user of a generative platform (UNESCO, 2021; UNICEF, 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 sequence, prompt or turn mediation, it does not translate them into free play. Accompanying three- to six-year-olds in play is sustaining initiative, open time and the absence of an imposed product. The rest is a sequence. It is not free play, and it must not be presented as what it is not.
Editorial Laboratory of NEXTECH.IA / Ingeniero Mitre.