Ask Me Anything: Exploring Children's Attitudes Toward an Age-tailored AI-powered Chatbot

Synthesis: Ask Me Anything: Exploring Children's Attitudes Toward an Age-tailored AI-powered Chatbot — This exploratory study designed "Ask Me Anything" (AMA), a child-friendly, topic-bounded chatbot restricted to astronomy, sneakers and shoes, and dinosaurs, to investigate how 63 children (ages 6–14, first grade through middle school) in a U.S. public school form attitudes and trust toward AI-powered conversational agents. Thematic analysis of interactions and post-surveys revealed three patterns—expressing wonder and curiosity, testing trust and developing confidence, and building relationships through anthropomorphization—alongside a broad openness to and high trust in AI as an information source, which the authors link to gaps in critical engagement and AI Literacy. The study calls for age-sensitive, trust-aware design and explicitly teaching digital safety concepts to children engaging with AI systems.

Key Findings

• Designed and deployed AMA, a topic-constrained chatbot built on the ChatGPT API, using prompt engineering with topic and age variables (7–9, 9–11, 12–14) to tailor response complexity and tone; the authors switched from gpt-3.5-turbo to text-davinci-002 to maintain topic fidelity. This grounds the design in Prompt Engineering and Large Language Models (LLMs) behavior.
• 63 students from 1st grade and 6th–8th grade interacted with AMA in small groups; most (37/63) posed 1–3 questions, while a standout first-grader asked 21, revealing varied Student Engagement and curiosity. Dinosaurs and sneakers/shoes were the most popular topics.
• Identified three inductive themes in child–AI interactions: expressing wonder, surprise, and curiosity; testing trust and developing confidence; and building relationships and anthropomorphizing. These reveal how children form early relationships with AI across functional and relational dimensions.
• Children actively tested the chatbot's credibility by posing questions to which they already knew the answers (e.g., "how big is a t rex"), an expression of epistemic trust and self-agency that reinforces Trust in AI as a source of information.
• Survey results showed 52% of students generally trusted AMA's responses and 35% trusted it like a teacher or friend; about a third were willing to confide in it. Chi-square tests found no statistically significant grade-level differences in trust (χ²(6)=5.68, p=.459) or confiding (χ²(6)=3.05, p=.80), though qualitative data suggested developmental variation.
• Students' mental models of how AMA worked were assessed, with most selecting "a smart computer program that learns from questions and answers," pointing to nascent Machine Learning understanding; chi-square analysis found no significant grade differences (χ²(6)=5.95, p=.429).
• The findings highlight the need for age-sensitive Scaffolding and Trust Calibration, and reveal gaps in children's digital safety awareness—some children were willing to share secrets with the chatbot—underscoring the importance of teaching Privacy, data use, and social boundaries of AI in K-12 education.
• The study uses mixed methods, triangulating interaction logs, audio/screen recordings, behavioral observations, and post-surveys via inductive thematic analysis, grounded in developmental theories of epistemic trust.

What this means for practice

• Instructors. Teach Privacy, data use, and the social boundaries of AI explicitly and early: some children in this study were willing to share secrets with the chatbot, and the authors link broad trust to gaps in digital safety awareness.
• Instructors. Make credibility checking a routine rather than an improvised move. Children already tested AMA with questions to which they knew the answers, so name and rehearse that strategy as part of Critical Thinking with AI outputs.
• Instructors. Read high trust as a signal to intervene, not as success: 52% of students generally trusted AMA's responses and 35% trusted it like a teacher or friend, and the authors connect this openness to underdeveloped Trust Calibration.
• Designers. Keep topic boundaries and response limitations in place as deliberate safeguards; the authors argue these constraints shaped children's perception and engagement, and the design also had to switch models (from gpt-3.5-turbo to text-davinci-002) to hold topic fidelity.
• Designers. Design for the youngest users' input constraints, not just their reading level: first graders often could not type their questions and needed facilitator help, so plan voice input or adult assistance for early-elementary deployments.

Limitations

• One school and a self-selected group: 63 students took part from a single public K–8 school in the Northeastern United States, drawn from 125 children attending two days of 45-minute sessions, and they used AMA only when they chose that station and it was free; grades 1 and 6–8 were represented, with no grades 2–5.
• Exposure was very brief — three to ten minutes of interaction in a single session — with no pretest, comparison condition, or follow-up, so no claim about persistence or change over time is possible.
• The post-survey (four multiple-item items and one open-ended item) was not formally validated or piloted; language was simplified and read aloud for first graders, who answered verbally or by pointing, and the authors state it was refined collaboratively rather than field-tested.
• Trust and understanding are self-reported, and the grade-level statistical comparisons rest on small cells: chi-square tests found no significant grade differences in trust (χ²(6) = 5.68, p = .459), confiding (χ²(6) = 3.05, p = .80), or mental models of the bot (χ²(6) = 5.95, p = .429), even though the qualitative data suggested developmental variation.

Citation

Vahedian Movahed, S., & Martin, F. G. (2025). Ask Me Anything: Exploring Children's Attitudes Toward an Age-tailored AI-powered Chatbot. International Journal of Artificial Intelligence in Education, 35(4), 3979–4001.