Autiverse: Eliciting Autistic Adolescents' Daily Narratives through AI-guided Multimodal Journaling
Migyeong Yang, Kyungah Lee, Jinyoung Han, SoHyun Park, Young-Ho Kim · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791381
Summary
Yang and colleagues present Autiverse, a tablet-based AI-guided multimodal journaling app for autistic adolescents with functional verbal abilities (formerly classified as Level 1 / high-functioning autism). The paper addresses a well-documented gap: although journaling is an established route to strengthening narrative skills and autobiographical memory, its text-centric, open-ended, executive-function-heavy nature is a poor fit for autistic adolescents who may possess strong visual thinking but struggle with sequencing, causal linking, and translating experience into coherent prose. The authors first ran a formative study with five autism experts and six parents, surfacing three needs: structure experience through routines, balance flexibility with executive load via visual plus conversational scaffolds, and prioritise adolescent-AI interaction over top-down adult instruction. They then built Autiverse, which pairs an LLM-driven conversational peer (a customisable same-age character with chosen name, voice, and appearance) with a four-panel comic-strip canvas. The conversation follows a stepwise ABC-E structure — Antecedent, Behaviour, Consequence, Emotion — across six phases (Preparation, Articulation, Verification, Elaboration, Revision, Wrapup), with LLM components Event Extractor, Story Analyzer, Question Generator, and Description Reconstructor driving the loop. The system is implemented on Samsung Galaxy Tab S9 using GPT-4.1-mini (OpenAI Enterprise), CLOVA Speech Recognition and CLOVA Voice for Korean, with React Native/TypeScript on the client and a FastAPI/PostgreSQL backend. It was deployed in the homes of 10 Korean autistic-adolescent-parent dyads (adolescents aged 11-17, IQ 63-101) for two weeks in a field study approved by the IRB, with parents serving as safeguarding observers rather than active participants.
Key findings
Across the 14-day deployment, the 10 adolescents produced 122 journal entries (mean 12.2 per participant); all dyads used the system at least 12 of 14 days. Sessions averaged 9 minutes 43 seconds and 46.92 conversational turns. Interaction-log analysis showed that Antecedent and Behaviour components were typically introduced early in the Articulation phase, consistent with prior work on autistic focus on concrete observable details; Consequence and Emotion were usually surfaced later, during the Elaboration phase when the chatbot actively probed for them, suggesting the scaffolding specifically helped with reflective and emotional content. On exit-survey 5-point scales, adolescents rated the comic strip highly as a recall aid (M=4.5, SD=0.71), felt strong ownership of the entries (M=4.4), autonomy over writing (M=4.2), and viewed the AI peer as friendly (M=4.5, SD=0.53). Eighty percent of adolescents heavily customised the AI peer's name, voice, and appearance, and parents reported that customisation reinforced engagement. Parents reported expanded narrative coherence, broader emotional vocabulary (e.g., "confused", "proud"), and new insights into previously invisible daily events (e.g., discovering a child had slammed a keyboard in anxiety at school). A mixed-effects model showed parental moderation required during sessions significantly decreased over the two weeks (p = 0.001), with eight adolescents using Autiverse independently on at least one day. Mixed ratings for long-term willingness to continue (some adolescents disliked "diary" framing or anticipated hard days) and signs of the AI peer over-validating problematic framings (sycophancy risk) tempered the positive results.
Relevance
For accessibility practitioners working with autistic users, the paper is a clean example of how to fuse well-established analogue interventions (Social Stories, Comic Strip Conversations, the ABC behavioural framework, TEACCH-style visual schedules) with LLM-driven conversational scaffolding rather than treating AI as a replacement for clinical practice. Concrete design moves worth transferring to other contexts: using closed-set emotion buttons (a 12-item subset of Plutchik's wheel, curated with an autism expert) instead of open-ended emotion prompts; using a peer rather than authority persona to lower defensiveness; segmenting narrative by ABC-E to turn executive-function load into manageable question-sized units; positioning parents as observers not participants to protect adolescent autonomy; and adding explicit guardrails (short replies, single question per turn, avoiding "why" questions that trigger emotional dysregulation, refusing off-topic answers). Limitations to flag: sample of 10 Korean Level-1 autistic adolescents with verbal ability, two-week duration, and the documented risks of LLM sycophancy over longer use. The finding that adolescents shift from parent-moderated to self-directed use over 14 days is promising for staged-autonomy designs that start with caregiver involvement and gradually hand over control, relevant across cognitive and communication-access contexts beyond autism.
Tags: autism · adolescents · journaling · large language models · conversational agents · scaffolding · visual supports · narrative skills · multimodal interaction · AI peer