Technical Summary: Tinker Tales

Problem Statement

While conversational AI agents are increasingly utilized as creative partners in adult contexts, the design of child–AI dialogue in co-creative settings remains underexplored. Existing research on children's interactions with AI predominantly focuses on AI-led, instructional learning activities rather than collaborative creation. Furthermore, current child–AI storytelling systems often rely on screen-based interfaces, potentially neglecting the benefits of tangible and multisensory modalities for younger children's cognitive development. A critical gap exists in understanding how conversation design—specifically how an agent structures questions and elaboration—shapes the nature of children's contributions and their ability to engage in iterative, shared meaning-making.

Methodology

System Design: Tinker Tales

The authors developed Tinker Tales, a tangible dialogue system designed for child–AI collaborative storytelling. The system integrates three core components:

1. Physical Layer: A four-sided game board representing four narrative stages (Start, Journey, Climax, End) and NFC-embedded toys (characters, places, items, emotions).
2. Mobile Application: Mediates interaction by sensing NFC tags to identify story elements and facilitating voice-based dialogue (Speech-to-Text and Text-to-Speech).
3. Conversational AI Backend: Powered by AWS services (Lambda, Cognito, DynamoDB) and AI models (Amazon Bedrock/Claude Haiku 4.5, OpenAI TTS).

Interaction Flow: Children physically place and move story elements on the board. The system scans these elements via NFC, and the AI agent generates story drafts based on these inputs. The agent then poses follow-up questions to encourage elaboration. Children respond via voice, and the agent updates the story, creating a recurring co-creative loop.

Conversation Design & Scaffolding:
The AI's behavior is governed by a scripted prompting strategy organized into a dialogue graph with eight phases. Crucially, the agent's responses are grounded in two educational frameworks:

• Applebee's Narrative Development Model: To scaffold narrative thinking (moving from isolated events to coherent narratives).
• CASEL's Social-Emotional Learning (SEL) Framework: To guide the elaboration of characters' emotions and perspectives.

User Study

• Participants: 10 children (ages 6–8, 5 girls, 5 boys) with spoken English proficiency.
• Setting: Home-based user study to capture naturalistic interaction.
• Conditions: The study employed two conversation design conditions, counterbalanced across participants:
  1. Structured Condition: The agent posed elaboration questions explicitly aligned with the narrative development and SEL frameworks (e.g., asking about causal links or character emotions).
  2. Generic Condition: The agent posed general elaboration questions without targeted pedagogical guidance (e.g., "Would you like to add something?").
• Data Collection: Turn-by-turn dialogue transcripts, session logs, pre-study AI literacy questionnaires, and optional post-study surveys.
• Analysis: An iterative, inductive qualitative coding approach was used to analyze child–AI dialogues at the turn-level (utterances) and stage-level (narrative phases). Coding focused on:
  • Agent question framing.
  • Narrative functions of child input (e.g., adding events, causality, emotion).
  • Uptake of child contributions (how the AI incorporated inputs).

Key Contributions

1. System Development: The creation of Tinker Tales, a tangible dialogue system that instantiates educational frameworks (Applebee and CASEL) directly into conversation design for child–AI co-creative storytelling.
2. Empirical Study: A home-based study with 10 children examining how different forms of conversational scaffolding influence child–AI collaboration.
3. Dialogue Analysis: A characterization of how prompt framing shapes the form and consistency of children's narrative contributions and how children navigate interactional breakdowns through conversational repair.

Results

Impact of Prompt Framing on Narrative Contributions

The study found that the framing of the agent's questions significantly influenced the form and consistency of children's contributions:

• Consistency: In the Structured condition, scaffolded questions elicited narrative contributions in 90% of responses for "Primitive narrative" scaffolds and 100% for "Chain narrative" scaffolds. In contrast, the Generic condition elicited contributions in only 37% of responses, with high variability (some children refused or gave brief answers).
• Form:
  • Primitive narrative scaffolds (e.g., "What else might they do?") successfully directed children toward event elaboration (adding new events).
  • Chain narrative scaffolds (e.g., "Why did they do that?") elicited causal reasoning linking events to character motivations.
  • Social awareness scaffolds (e.g., "How do they feel?") supported children in articulating emotional states and perspectives (62% response rate), enriching the narrative's emotional structure.
  • Generic questions resulted in less consistent emotional elaboration (only 12% of responses).

Interactional Dynamics and Repair

The study observed that collaborative continuity relied on conversational repair mechanisms when breakdowns occurred (due to STT errors, noise, or scanning mistakes):

• Agent-initiated repair: The AI sought clarification when ambiguity was detected.
• Child-initiated repair: Children actively corrected misinterpretations (e.g., correcting "Lion is a liar" to "Lion is friendly").
• Collaborative grounding: Children monitored the interaction state beyond narrative content, intervening when system behavior conflicted with their intentions (e.g., correcting accidental token scans).

Narrative Agency

Children maintained primary control over the narrative. They initiated story ideas, selected characters, and defined story elements. The AI acted as a support mechanism, transforming inputs into narrative content that remained open to further shaping by the child.

Significance and Claims

The paper claims that conversation design functions as a primary mechanism for shaping child–AI collaborative dialogue. Specifically:

• Prompt Framing as Interaction Design: The way an agent frames questions is not merely a content generation tool but an interactional move that structures turn-taking dynamics. Pedagogically grounded questions provide a principled mechanism to structure participation while preserving children's agency over narrative content.
• Structured yet Flexible Design: Effective child–AI collaboration requires a system that is structured enough to guide development (via scaffolding) but flexible enough to incorporate unscripted moments and repair breakdowns.
• Tangible Interaction: The integration of physical manipulation (NFC toys) with voice dialogue lowers barriers to participation, allowing children to express ideas fluidly without the constraints of reading and writing.

The authors conclude that these findings offer practical insights for designing conversational AI applications in education, emphasizing that the design of the dialogue itself is critical to the success of co-creative learning experiences.

Limitations

The authors note several limitations:

• Sample Size: The study involved a small number of participants (N=10), limiting generalizability.
• Data Modality: Conducted in homes without audio/video recording to protect privacy, meaning non-verbal behaviors were not captured.
• Demographics: Participants were English-speaking children aged 6–8; findings may not extend to other linguistic, cultural, or developmental contexts.
• Duration: Interactions were limited to a few sessions, preventing observation of how collaboration patterns evolve over extended use.