Raw capture — raw/anthropomorphism-children-llm-chatbots-review.md
url: https://arxiv.org/html/2607.18250v1 title: "Anthropomorphism in Children's Interactions with LLM Chatbots: A Systematic Review of Drivers and Outcomes" author: Hansinie Madushika Jayathilake, Renkai Ma date_fetched: 2026-07-25 date_published: 2026-05-09
Authors: Hansinie Madushika Jayathilake and Renkai Ma, University of Cincinnati
arXiv: 2607.18250v1, submitted 9 May 2026 (License: arXiv.org perpetual non-exclusive license)
The authors systematically review 35 empirical studies (2022–2025) examining how children anthropomorphize LLM-based chatbots. They identify four drivers of anthropomorphism and five outcome themes, revealing both benefits and risks for children's development.
LLM chatbots now support children across education, mental health, storytelling, family therapy, and environmental awareness. Children naturally anthropomorphize non-human entities as part of normal development. The authors position their work within Epley et al.'s (2007) three-factor SEEK theory (Elicited Agent Knowledge, Effectance Motivation, Sociality Motivation) and pose two research questions: (RQ1) What drivers trigger or facilitate anthropomorphism in child-LLM interactions? (RQ2) What are the anthropomorphic outcomes of those interactions?
The paper makes four contributions: (1) a descriptive landscape of 35 empirical studies; (2) a taxonomy of anthropomorphic drivers extending SEEK theory; (3) five inductively derived outcome themes moving beyond a simple benefits/risks binary; and (4) a stakeholder-oriented future research agenda.
The SEEK framework proposes that anthropomorphism arises from three determinants. Elicited Agent Knowledge involves using self-knowledge to interpret unknown agents. Effectance Motivation drives people to attribute human traits to reduce uncertainty. Sociality Motivation reflects the need for social connection. Prior work has applied this to robotics and voice assistants, but the authors note that "a systematic synthesis of recent empirical work on the drivers of anthropomorphism in the context of LLM chatbot interactions remains absent."
Existing reviews focus on functional efficacy in education and therapy or on safety assessments. Reviews have examined chatbot design for learning, mental health outcomes, bullying prevention, and risk frameworks. However, these overlook "how children's cognitive maturity fundamentally constrains their understanding of these 'living' or social entities."
Piaget's four stages (sensorimotor, preoperational, concrete operational, formal operational) shape how children perceive AI. Hong et al. (2025) proposed a Neo-Piagetian framework where users progress through parallel stages of AI understanding. The authors argue that "no systematic review explicitly connects these developmental stages to the anthropomorphic drivers and outcomes."
The review followed PRISMA guidelines across ACM Digital Library, Scopus, and Web of Science. Inclusion criteria required English-language empirical studies published between 2022 and August 2025, focusing on children under 18, with chatbots explicitly based on LLMs or generative AI. The search string combined terms for conversational agents, LLMs/generative AI, and child/youth populations.
After screening 178 full-text articles, 35 studies met all criteria. Exclusion reasons included: no LLM (n=22), population/interaction mismatch (n=118), and non-empirical nature (n=3). The authors extracted bibliographic metadata, design features (for RQ1), and qualitative evidence of anthropomorphic outcomes (for RQ2).
Analysis used a three-step combined inductive-deductive approach: (1) preparatory classification (mapping ages to Piagetian stages and use cases); (2) inductive thematic coding without pre-existing frameworks; (3) deductive theoretical organization mapping RQ1 drivers to SEEK theory and RQ2 outcomes to Huppert's wellbeing model.
Research grew from 1 paper in 2023 to 15 in 2024 and 19 in early 2025. Venues are concentrated in CHI (n=17) and IDC (n=7). The US leads geographically (n=18), followed by China (n=5) and South Korea (n=4). Technologically, 28 of 35 studies used OpenAI's GPT models. Functional applications include Self-Directed Information Seeking (n=6), Co-Authoring Stories (n=5), Teaching Science through Dialogue (n=4), and Eliciting Emotional Expression (n=4). Age distribution peaks at age 12 (concrete-to-formal transition). Methodologically, 19 studies used mixed methods, 13 qualitative, and 3 quantitative. Most were short-term, single-session deployments.
Human-Like Persona Construction (n=19) — activates Elicited Agent Knowledge through human-like roleplay characters with backstories, voice and auditory cues (gendered, accented), visual expression cues, communicative competence (first-person pronouns, sustained dialogue), and demonstrable cognitive performance.
Supportive Companionship (n=16) — satisfies Sociality Motivation through safe, non-judgmental peer positioning, emotional responsiveness (reflective empathy, encouragement), and entertaining role adoption (humor, flexible interaction modes, casual conversation).
Adaptive Scaffolding (n=14) — triggers Effectance Motivation by reducing uncertainty through cognitive scaffolding (strategic questions, comprehension support, personalization) and affective scaffolding (emotional support, adaptive responsiveness to sustain engagement).
Non-human Embodiment Design (n=10) — extends beyond SEEK by using fantasy archetypes (whimsical animals, nature embodiments, cartoon visuals) with synchronized sensory coordination (speech-text coordination, physical-digital integration). The authors note this shows "coherent cross-modal responsiveness may override surface appearance in activating anthropomorphic responses."
Social Ties Construction (n=20) — children form relationships ranging from Strong Emotional Ties (parental figures, romantic partners—the latter appearing only at formal operational stage, close friends) to Weak Instructional Ties (teacher/mentor) and Weak Entertainment Ties (playful co-participants). The authors observe that "anthropomorphism is not a monolith; children tailor the social tie intimate or instrumental based on the perceived affordances."
Social Boundary Exploration — children test the chatbot's capacities and limitations through their anthropomorphic framing, exploring what the "social other" can and cannot do.
Human Narrative Attribution — when conversations break down, children attribute human-like intentions or explanations to the chatbot's errors rather than recognizing them as technical failures.
Paradoxical Social-Moral Responses — children exhibit conflicting ethical stances, simultaneously treating the chatbot as socially real (deserving politeness, capable of being hurt) while also recognizing it as a machine, creating moral ambiguity.
Dual Consciousness Formation — children develop a cognitive dissonance where they know the chatbot is not human yet respond to it as if it were, holding both frameworks simultaneously rather than resolving the contradiction.
The fourth driver (non-human embodiment) demonstrates that "LLM conversational coherence alone is sufficient to trigger anthropomorphism independent of visual human-likeness." This challenges assumptions that anthropomorphism requires human-like appearance.
All five outcome themes carry both beneficial and concerning expressions. For example, strong emotional ties can support vulnerable children but also create substitute attachments that may displace human relationships.
Children at different Piagetian stages show qualitatively different patterns. The romantic partner sub-theme emerged exclusively at the formal operational stage, suggesting "the most concerning form of substitute attachment...requires the abstract social cognition characteristic of adolescence."
The evidence base is predominantly short-term and educationally oriented, so findings "primarily reflect children's immediate interactional responses rather than sustained developmental trajectories." Geographic and technological monoculture (US-centric, GPT-dominated) also limit generalizability.
The authors identify three pressing directions: longitudinal investigation of sustained anthropomorphic outcomes, stage-differentiated design approaches, and extension of child AI governance frameworks beyond safety thresholds.
The review establishes a taxonomy of anthropomorphic drivers extending SEEK theory and synthesizes five outcome themes revealing how children's anthropomorphic engagement simultaneously implicates multiple dimensions of developmental wellbeing. The findings can inform design of LLM chatbots focused on children's wellbeing and sustainable interactions meeting developmental needs.