Author – Gökçe Şahin

Over the last decade, children have increasingly interacted with artificial intelligence (AI) through AI-embedded educational tools, apps, and toys. The latter, in particular, has accelerated the onset of the age at which children begin interacting with artificial agents, such as large language models (LLMs). As LLM-powered generative and conversational AI tools gradually integrate into multiple aspects of everyday lives, this raises concerns about their ethical and safe use by children, as they represent a vulnerable position in their interactions with AI, due to their sensitive profiles, as well as constraints of AI design (Jiao et al., 2025; Kurian, 2024). Even though implementing safe frameworks for LLMs provides user protection to a significant extent, child populations require fine-grained designs tailored to their developmental needs (Kurian, 2025a; Wang et al., 2022). In this respect, child-centred LLM designs and governance require further thinking to comprehend the requirements from a developmental perspective.

There is a common phenomenon prevalent in early childhood called “animism” (Piaget, 1973). Even in the absence of agency in toys (or other physical objects, abstract concepts), children can attribute psychological and biological capacities to these non-living entities. Considering the advancing capacity of GenAI-embedded tools to produce rich, human-like outputs in interactions, this highlights a critical factor for understanding children’s cognitive representations of the artificial agents. The developing cognitive skills that also underlie animism may lead to more anthropomorphic attributions to LLM-driven tools among young end-users (Okanda et al., 2021; Xu & Warschauer, 2020). In addition, as robotics and AI research advance, the embodiment of AI agents is expected to improve (e.g., social humanoid robots). This includes a critical risk of blurring the line between human and non-human partners as everyday interlocutors in children’s social environments (or even earlier developmental stages, for infants see Meltzoff et al., 2010). As a result, this can also cause distorted attributions to AIs in children’s mental representations (Salles et al., 2020). To mitigate such consequences, one solution could be to implement instructions adapted for children and to include explicit disclaimers reminding (on a reasonable basis) young users that they are interacting with a “tool” (Kurian, 2025b; but see Laestadius & Campos-Castillo, 2026). Additionally, children may be exposed to emotionally charged content through interactions with chatbots, as these models could also be trained on datasets containing emotionally expressive language or narratives of human emotional experiences (Ben-Zion, 2025; Stein & Shollo, 2025). It may consequently foster a false sense of connection with emotionally responsive AI companions (Ben-Zion, 2025), which, ultimately, could shape children’s interaction patterns with LLM-driven technologies (e.g., building friendships). On the other hand, it could also be an opportunity to exercise social skills for children whose personality traits are relatively shy or introverted, or who have developmental abnormalities that may limit their interactions with others. As a result, they might become more confident and open to building connections with their peers and be better at navigating their social environment.

Furthermore, LLMs should be fine-tuned to individual differences in children, such as age, language abilities (e.g., reading comprehension), and related cognitive skills (e.g., working memory, question-asking abilities, see Abdelghani et al., 2025), as well as to the child’s preferences in interaction (e.g., direct or gradual provision of information). As information about the end user accumulates across dialogues, the tool should learn to comprehend these variables and generate user-centred profiles for effective communication. In parallel with this, having an accurate understanding of the child’s prior knowledge will help to gauge the information gap and provide an appropriate level of information (i.e., contingent shifting; Wood, 1980; Wood et al., 2016; Zhang & Whitebread, 2017), which would allow children to learn progressively. Correspondingly, we should have a deeper understanding of the developmental needs that display significant differences across age groups and of individual differences. There is, in turn, a fundamental need to develop comprehensive theoretical frameworks for designing AI tools grounded in children’s cognitive skills to maximize the benefits of LLMs as complementary learning tools.

Another important aspect of child-centred LLMs is the design to exercise children’s metacognitive skills during interactions. This can help create a space for children to reflect on their own thoughts, mitigating the risk of over-reliance and over-trust in these LLM-powered learning partners. These learning processes, in turn, can culminate in advanced critical thinking, rigorous (re)evaluations of the presented information, improving the practice of selective trust (Geng et al., 2025; see also Koenig et al., 2019; 2022), and children’s awareness of privacy that can offer them protection from possible threats (Ferrarello et al., 2022).

The design of LLMs should encourage and provide collaborative learning experiences for children (Tomasello, 2016). We need to develop systems that facilitate learning experiences by engaging children as active learners, rather than passive consumers of information generated within an instructive teaching model. For instance, scaffolding the child’s learning (which ultimately leads to the child’s competence in self-scaffolding; Bickhard, 2005): providing hints rather than providing complete answers upfront should be a prominent consideration in LLM designs. As the child progressively gains competence in the task, the artificial learning tool should gradually fade support (i.e., transfer of responsibility, van de Pol et al., 2010). In this regard, developmental science could be an engine for better understanding which variables make the learning-teaching processes more egalitarian and collaborative (Rogoff, 2003). Accordingly, not dominating the learning environment, stating the knowledge gap, and indicating potential biases in the information can guide children to think further before accepting the generated information as the truth. This dynamic learning process would enable the (co-)construction of knowledge by two active agents, human and artificial, without LLM-dominated knowledge asymmetry. This theoretical framework could then inform the development of advanced user-centred and child-safe design approaches for LLMs.

Further improvements are required to enable terminating the conversation, redirecting the flow towards an age-appropriate subject, and explaining why the tool cannot answer the current prompt, which could be informative and helpful for children to reframe subsequent interactions accordingly, while ensuring the child’s safety (i.e., age-based guardrails, Kurian, 2025). Beyond the features and alignment of AI tools with children’s needs, adult-led monitoring serves as another prominent domain in children’s learning with AI. Given this, the corresponding implication is to provide AI literacy training for caregivers and other adults (e.g., teachers) involved in children’s formal and informal learning experiences, thereby enhancing their competence to monitor children’s activities.

Contributing to the development of accessible and equitable LLM tools represents a crucial scientific and societal responsibility. These tools should be accessible to everyone in the community, including children living in poverty (Bassignana et al., 2025; Katona & Gyonyoru, 2025) or children with disabilities whose needs require the utilisation of different modalities to facilitate learning in these digital environments. In addition, fairness requires comprehensive approaches (including appropriate training data) during LLM development to represent and ensure more inclusive language in dialogues that embrace differences and underrepresented minorities in children and adolescent populations (e.g., related to gender identification, religion, sexual orientation, or any other reasons).

Finally, LLM-driven AI system designs should be informed by children’s hallmark socio-cognitive capacities, such as curiosity, flexibility, and selectivity, as well as their exploratory strategies during learning experiences. Thus, these technological tools, as learning partners, ought to emulate, rather than inhibit, such competencies and align with them. Consequently, we need to formulate comprehensive and secure guidelines for child-LLM interactions. This effort will not only address age-appropriate, safe LLM designs but also enable children to benefit from LLMs as tools, thereby enhancing their learning experiences in both formal and informal settings.

References

Abdelghani, R., Murayama, K., Kidd, C., Sauzéon, H. & Oudeyer, P. (2025). Investigating middle school students’ question-asking and answer-evaluation skills when using ChatGPT for science investigation. 10.48550/arXiv.2505.01106.

Bassignana, E., Curry, A. C., & Hovy, D. (2025). The AI gap: How socioeconomic status affects language technology interactions. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1) (pp. 18647-18664).

Ben-Zion, Z. (2025). Why we need mandatory safeguards for emotionally responsive AI. Nature, 643(8070), 9. https://doi.org/10.1038/d41586-025-02031-w.

Bickhard, M. H. (2005). Functional scaffolding and self-scaffolding. New Ideas in Psychology, 23, 166–173.

Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., … & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489-530.

Ferrarello, L., Fiadeiro, R., Mazzon, R., & Cavallaro, A. (2022). Reframing the narrative of privacy through system-thinking design. DRS2022: Bilbao, 2022, 1-18.

Geng, Z., Zeng, B., & Huang, J. (2025). Living in a digital ecology: Children’s selective trust in technological informants. Journal of Applied Developmental Psychology, 101, 101872.

Jiao, J., Afroogh, S., Chen, K., Murali, A., Atkinson, D., & Dhurandhar, A. (2025). LLMs and Childhood Safety: Identifying Risks and Proposing a Protection Framework for Safe Child-LLM Interaction. ArXiv, abs/2502.11242.

Katona, J. & Gyonyoru, K.I.K. (2025). AI-based adaptive programming education for socially disadvantaged students: bridging the digital divide. TechTrends 69, 925–942. https://doi.org/10.1007/s11528-025-01088-8.

Kidd, C. & Birhane, A. (2023). How AI can distort human beliefs. Science, 380(6651), 1222–1223.Another important aspect of child-centered LLMs.

Koenig, M. A., Li, P. H., & McMyler, B. (2022). Interpersonal trust in children’s testimonial learning. Mind & Language, 37, 955– 974. https://doi.org/10.1111/mila.12361.

Koenig, M. A., Tiberius, V., & Hamlin, J. K. (2019). Children’s judgments of epistemic and moral agents: From situations to intentions. Perspectives on Psychological Science, 14(3), 344-360.

Kurian, N. (2024). ‘No, Alexa, no!’: designing child-safe AI and protecting children from the risks of the ‘empathy gap’ in large language models. Learning, Media and Technology, 1–14. https://doi.org/10.1080/17439884.2024.2367052.

Kurian, N. (2025a). Designing Child-Safe Conversational AI: Three Dilemmas for Responsible Design. In Proceedings of the 7th ACM Conference on Conversational User Interfaces (pp. 1-5).

Kurian, N. (2025b). Developmentally aligned AI: A framework for translating the science of child development into AI design. AI, Brain and Child 1, 9. https://doi.org/10.1007/s44436-025-00009-z.

Laestadius, L. I., & Campos-Castillo, C. (2026). Reminders that chatbots are not human can be risky. Trends in Cognitive Sciences, 30(3), 187–189. https://doi.org/10.1016/j.tics.2025.12.007.

Meltzoff, A. N., Brooks, R., Shon, A. P., & Rao, R. P. (2010). “Social” robots are psychological agents for infants: a test of gaze following. Neural networks: the official journal of the International Neural Network Society, 23(8-9), 966–972. https://doi.org/10.1016/j.neunet.2010.09.005.

Okanda, M., Taniguchi, K., Wang, Y., & Itakura, S. (2021). Preschoolers’ and adults’ animism tendencies toward a humanoid robot. Computers in Human Behavior, 118, 106688.

Piaget, J. (1973). The child’s conception of the world. Transl. by Joan and Andrew Tomlinson. Paladin.

Rogoff, B. (2003). The cultural nature of human development. Oxford University Press.

Salles, A., Evers, K., & Farisco, M. (2020). Anthropomorphism in AI. AJOB Neuroscience, 11(2), 88–95. https://doi.org/10.1080/21507740.2020.1740350.

Stein, M.-K., & Shollo, A. (2025). Microfoundations of rationality in the age of AI: On emotions, bodies and intelligence. Information and Organization, 35(3), 100583. https://doi.org/10.1016/j.infoandorg.2025.100583.

Tomasello, M. (2016), Cultural Learning Redux. Child Development, 87, 643-653. https://doi.org/10.1111/cdev.12499.

van de Pol, J., Volman, M., & Beishuizen, J. (2010). Scaffolding in teacher–student interaction: A decade of research. Educational Psychology Review, 22(3), 271–296. https://doi.org/10.1007/s10648-010-9127-6.

Wood, D. J. (1980). Teaching the young child: Some relationships between social interaction, language, and thought. In D. R. Olson (Ed.) The social foundations of language and thought (pp. 280–296). W. W. Norton & Company.

Wood, E., Petkovski, M., De Pasquale, D., Gottardo, A., Evans, M. A., & Savage, R. S. (2016). Parent scaffolding of young children when engaged with mobile technology. Frontiers in Psychology, 7, 690. https://doi.org/10.3389/fpsyg.2016.00690.

Xu, Y., & Warschauer, M. (2020). What are you talking to?: Understanding children’s perceptions of conversational agents. In Proceedings of the (CHI) Conference on Human Factors in Computing Systems (pp. 1-13).

Zhang, H., & Whitebread, D. (2017). Linking parental scaffolding with self-regulated learning in Chinese kindergarten children. Learning and Instruction, 49, 121- 130.

Zhao, J., Milosevic, T., Livingstone, S., & James, C. (2024). Designing for children’s autonomy in the age of AI (Part II). Oxford Child-Centred AI (OxfordCCAI) Design Lab, University of Oxford.

Further Reading/Watching/Listening:

Readings:

Binz, M., & Schulz, E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences, 120(6), e2218523120.

Booth, R. (2025, December 9). ‘I feel it’s a friend’: Quarter of teenagers turn to AI chatbots for mental health support. Guardian. https://www.theguardian.com/technology/2025/dec/09/teenagers-ai-chatbots-mental-health-support.

Goodacre, E., & Gibson, J. (2026). AI in the Early Years: Examining the implications of GenAI toys for young children. Apollo – University of Cambridge Repository. https://doi.org/10.17863/CAM.126270.

Waytz, A., Cacioppo, J., & Epley, N. (2010). Who Sees Human?: The Stability and Importance of Individual Differences in Anthropomorphism. Perspectives on Psychological Science, 5(3), 219-232. https://doi.org/10.1177/1745691610369336.

Videos:

BrainMind Summit (2022, January 16). How the child’s mind informs AI research – Alison Gopnik at BrainMind. [Video]. Youtube. https://www.youtube.com/watch?v=3_mDloRjiHQ.

Sean Carroll (2025, March 17). Mindscape 308 | Alison Gopnik on children, AI, and modes of thinking. [Video]. Youtube. https://www.youtube.com/watch?v=WHbMIpNrY64.

Image Attribution

Generated by: Nadia Piet  & Archival Images of AI + AIxDESIGN

Date: 2026

Prompt: “Diptych contrasting a whimsical pastel scene with large brown rabbits, a rainbow and a girl in a red dress on the left, and a grid of numbered superpixels on the right-emphasising the difference between emotive seeing and analytical interpretation.”