Autistic Adults' Perspectives on Self-Expression and Dyadic vs. Group Interactions in AR Telepresence
Mahrukh Tauseef, Alexandra Watkins, Zalen Ingram, Rendong Zhang, Akshith Ullal, Ritam Ghosh, Amy Sue Weitlauf, Zachary Warren, Nilanjan Sarkar · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791141
Summary
This CHI 2026 paper investigates how autistic adults experience augmented reality (AR) telepresence as a medium for social interaction, a population whose needs are largely absent from existing avatar and social VR/AR research. Motivated by the high prevalence of social isolation and depression among autistic adults and by sensory barriers that make face-to-face interaction in unfamiliar environments stressful, the authors ask two research questions: how autistic individuals wish to be represented as avatars and which avatar components best support authentic self-expression (RQ1), and how they perceive social presence when comparing dyadic (one-on-one) and group AR interactions (RQ2). The study used a three-phase participatory design protocol with eight autistic adults (ages 22-40) recruited through an opt-in medical research registry. Phase 1 co-created photorealistic MetaHuman avatars over one to two Zoom sessions, starting from Polycam 3D face scans and iterating on hair, clothing, gender presentation and facial detail. Phase 2 used semi-structured interviews to elicit sensory and social preferences that shaped the AR environment and activities. Phase 3 evaluated social presence in a dyadic AR Checkers game and a four-person AR version of Apples to Apples, using HoloLens 2 headsets, Kinect-based body tracking and Dollars Mocap for facial blendshapes. Data included the Networked Minds Social Presence Inventory (rated on a 0-4 Likert scale, with Bonferroni-corrected Wilcoxon signed-rank tests) and reflexive thematic analysis of post-activity interviews.
Key findings
In avatar co-creation, self-expression consistently outranked physical realism: participants chose hair, clothing, body type and gender presentation that matched the 'vibes' they wanted to project rather than strict likeness, and they tailored representations to the social context (casual clothing was preferred; formal attire was quickly rejected). Three participants reported low self-perception of their own face and sought external validation from partners or friends to judge resemblance. Many experienced the uncanny valley yet still identified with their avatars and valued ownership of the representation. On social presence, dyadic Checkers scored significantly higher than group Apples to Apples on every subscale (all p < 0.001 after Bonferroni correction, effect sizes 0.884-0.908): co-presence 3.05 vs 1.95, attentional engagement 3.1 vs 1.8, comprehension 3.39 vs 2.625, and behavioural interdependence 3.65 vs 2.375. Participants described dyadic interaction as 'intimate' and 'personal' while group interaction felt 'segregated' and 'like a Discord call' because tracking multiple speakers, cards and avatars imposed cognitive and perceptual load that impeded connection. Six of eight participants reported sound sensitivity, driving a design decision to use only minimal functional audio. Participants also expressed distrust when avatar motion did not precisely match their partner's real movements.
Relevance
For accessibility practitioners, the study offers concrete design guidance for AR and social VR products aimed at neurodivergent users: provide deep avatar customisation across hair, clothing, accessories and gender presentation rather than optimising for one-shot photorealistic likeness; default to low-stimulation sensory settings (minimal audio, soft lighting, no flashing) with user-adjustable controls; and recognise that group configurations beyond two or three people may impose cognitive load that degrades social presence even when each individual interaction works. The finding that shared turn-taking alone does not deliver behavioural interdependence is a useful caution for designers who assume activity structure will carry the social experience. Limitations include a very small (n=8), racially homogeneous and female-skewed sample recruited from a single medical registry; findings should be read as design considerations, not generalisable requirements. The work complements broader literature on autistic self-representation in social VR and extends participatory design practice into AR telepresence.
Tags: autism · augmented reality · avatars · social presence · telepresence · participatory design · self-expression · neurodiversity · sensory sensitivity · co-design