AutiHero: Engaging Parents in Creating Personalized, Multi-path~Social Narratives for Autistic Children

1. Problem Statement

Autistic children often struggle with interpreting implicit social cues and understanding the consequences of their actions. Social narratives (structured stories explaining social situations) are a proven therapeutic intervention, but their practical application at home faces significant barriers:

• High Parental Burden: Creating effective narratives requires significant time, effort, and expertise to tailor stories to a child's specific interests, cognitive level, and daily context.
• Lack of Personalization: Existing tools often generate generic stories that fail to ground the narrative in the child's real-world environment, reducing engagement and efficacy.
• Limited Behavioral Exploration: Traditional stories often present a single linear path, missing the opportunity for children to rehearse different behavioral choices and observe their causal consequences.
• GenAI Limitations: While Generative AI (GenAI) can create stories, it often fails to maintain cross-scene consistency (characters/objects changing appearance) and causal structure (clear links between actions and outcomes), which are critical for autistic audiences who rely on literal and stable representations.

2. Methodology

A. System Design: AutiHero

AutiHero is a GenAI-based system comprising two applications: a Creator (web-based for parents) and a Reader (tablet-based for joint parent-child reading). The system is built on a multi-level narrative construction pipeline informed by formative interviews with 10 autism experts.

Key Design Principles:

1. Multi-Path Structure: Stories are not linear. They branch at a "Challenge" point into Desirable and Undesirable paths.
   • Desirable Path: Leads to positive social outcomes.
   • Undesirable Path: Shows negative consequences but includes a "Repair" phase where the situation is resolved, teaching children how to fix mistakes.
   • Both paths converge to a shared positive ending.
2. Context Grounding: The system uses parent-provided text and images (child's photo, friends, favorite toys, locations) to generate content.
3. Visual Consistency: To address GenAI hallucinations, the system uses a colored-pencil illustration style (less distracting than photorealism) and enforces strict entity consistency across pages.

Technical Architecture & Generative Pipeline:
The system utilizes a multi-stage pipeline involving Large Language Models (LLMs) and Text-to-Image models (OpenAI GPT-4o, GPT-4.1, GPT-Image-1):

1. Input: Parents register a child profile (photos, interests, familiar people/places) and a target behavior.
2. Narrative Generation: An LLM generates a draft multi-path story based on a template specific to the behavior type (Relationship, Social Rules, or Healthy Habits).
3. Validation & Refinement:
   • Content Validator: Checks for metaphorical language (must be literal) and context alignment.
   • Text Refiner: Adjusts reading level (targeting US Grade 5 or lower) and simplifies vocabulary.
   • Translation: Translates English to Korean using few-shot learning with manually refined pairs.
4. Scene & Entity Management:
   • A Scene Description Generator creates detailed prompts for every page.
   • A Scene-Entity Matcher ensures characters and objects from previous scenes appear consistently in new scenes.
   • An Entity Visual Describer generates visual descriptions for registered entities (e.g., "fire truck toy") and converts parent-uploaded photos into stylized portraits to protect privacy while maintaining recognizability.
5. Image Generation: A multimodal LLM generates illustrations for each page using the scene description and entity visual descriptions as inputs.
6. Privacy Safeguards: All facial images are converted to stylized portraits on the backend before processing; original images are deleted. OpenAI Enterprise is used to ensure data is not used for training.

B. Deployment Study

• Participants: 16 parent-child dyads (children aged 7–12, diagnosed with Level 1 or 2 Autism Spectrum Disorder).
• Duration: 2 weeks.
• Procedure: Parents used the Creator app to generate stories; families read them together on tablets. Data was collected via system logs, daily surveys, and post-study semi-structured interviews.
• Evaluation Metrics: Usage frequency, story creation/editing patterns, child engagement (parent-reported), and Technology Acceptance Model (TAM) scores.

3. Key Contributions

1. System Design: The implementation of AutiHero, the first GenAI system specifically designed to support parents in creating personalized, multi-path social narratives with strict context grounding and visual consistency.
2. Technical Pipeline: A novel context-grounded text+image generation pipeline that solves the consistency problem in GenAI by decoupling entity definition from scene generation and enforcing causal logic in narrative branching.
3. Empirical Evidence: A two-week field study demonstrating that parents can effectively use GenAI to create high volumes of personalized content (218 stories) and that these stories positively influence parental confidence and child engagement.

4. Results

• Usage & Creation: Parents created an average of 13.63 stories each (218 total). They actively edited 27% of stories, refining text for tone and regenerating images for consistency.
• Content Diversity: Stories covered 18 behavior categories, with the highest focus on social norms (35%) and self-care/daily living (25%). Parents successfully integrated specific child interests (e.g., dinosaurs, firefighters) into diverse scenarios.
• Engagement: Children showed high engagement (avg. 4.02/5.0). They actively participated in choosing story paths, often reacting emotionally to the consequences of their choices (e.g., smiling at positive outcomes, showing reluctance at negative ones).
• Parental Impact:
  • 82% of parents reported their own responses to challenging behaviors became more positive (less nagging, more reflective).
  • 76% of parents observed perceived improvements in their child's target behaviors (e.g., better turn-taking, trying new foods).
  • High TAM scores indicated high perceived usefulness and ease of use.

5. Significance and Implications

• Democratizing Intervention: AutiHero lowers the barrier for parents to create clinical-grade social narratives, shifting the role from "passive consumer" to "active co-creator."
• AI as a Reflective Aid: The study suggests GenAI should not just generate content but act as a reflective tool that helps parents analyze their child's needs and refine their communication strategies.
• Ethical AI in Sensitive Contexts: The paper provides a framework for using GenAI with children's data, emphasizing privacy-preserving stylization and the need for safeguards against parents targeting inappropriate behaviors (e.g., suppressing autism-specific traits like stimming).
• Future Directions: Highlights the need for on-device processing to further enhance privacy and the importance of designing systems that balance parental autonomy with evidence-based therapeutic guidance.

In summary, AutiHero demonstrates that with careful architectural design (multi-path structures, entity consistency, and privacy safeguards), GenAI can effectively support the complex, personalized, and sensitive task of behavioral guidance for autistic children.