Bridging Visual Asymmetry: Exploring AI-Mediated Communication Support for Parents with Visual Impairments and Their Sighted Children in Outdoor Informal Learning
Yutong Jiang, Zixuan Zhang, Jiaying Xu, Qingyun Zheng, Qian Guo, Qinyang Wang, Qi Wang, Guanhong Liu · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3790604
Summary
This CHI 2026 paper investigates how AI-mediated communication can support parents with visual impairments (PVI) in engaging their sighted children during outdoor informal learning - settings like parks, insect houses, and museums where unplanned, visually-driven discovery is central. The authors run a three-phase study. Study 1 is semi-structured interviews with 9 PVI plus field observations with 2 mixed-ability families, surfacing the perceptual, knowledge, and interactional challenges that shape parent-child engagement. Study 2 is a design consultation with three expert PVI that translates those findings into concrete interaction modalities, establishing AI's role as a communication mediator rather than an information provider. Study 3 is a Wizard-of-Oz deployment of Bond, a distributed prototype in which the child wears a chest-mounted GoPro camera while the parent uses a smartphone app that delivers interest detection (first-person scene descriptions of what the child is looking at) and two parent-triggered prompt generators - Learning Guidance (open-ended questions, progressive questioning, moderate prompts, fading guidance) and Knowledge Expansion (analogies, essences, functional reasoning). A within-subject field study with 14 families across three outdoor venues compares parent-child interactions with and without Bond. The authors propose Symbiotic Learning as a conceptual lens that reframes AI as a relational mediator bridging perceptual asymmetry and supporting co-constructed learning in mixed-ability families.
Key findings
The interview study identifies three polarized PVI interaction patterns - hands-off (withdrawal from educational engagement), over-involvement (authoritarian overcompensation), and limited support (willing but perceptually blocked) - plus two structural barriers: physical environment perception (PVI avoid visually dense settings like museums; one parent had never taken their child to one) and knowledge-reserve gaps amplified by inaccessible developmental resources. In the Bond field deployment, paired-samples t-tests across 14 families showed significant improvements in parental response rate (M=2.03 vs 1.58, p<.001) and effective guidance instances (M=12.29 vs 5.00, p<.001) with system support. Dialogue shifted from quick browsing (median 21.5 turns, shallow) to in-depth exploration (15.2 turns but 4.6 guidance points vs 1.1). BPSES parental self-efficacy scores improved significantly across all five dimensions (Wilcoxon signed-rank, p<.01). Conversation initiation rebalanced between parent and child, activating non-dominant participants - children in over-involved families started asking more; parents in hands-off families began responding meaningfully ("Before, I often said 'I don't know'"). TAM scores showed high perceived usefulness (M=4.59), ease of use (M=4.62), and trust (M=4.51), though perceived accuracy was lower (M=4.0). Participants described the system as an "invisible assistant" and reported collaborative correction when it misrecognized objects.
Relevance
This paper matters to accessibility practice because it moves AI assistive technology for blind users beyond one-way information delivery (describing the world to the PVI) toward relational, communication-mediating design that preserves the parent's caregiving role. For anyone designing AI tools for BLV users in family, educational, or social contexts, the design principles translate directly: start from the sighted partner's attention rather than the PVI's query, provide context-sensitive scaffolding rather than raw descriptions, and minimize intrusiveness so technology supplements rather than replaces human interaction. The Symbiotic Learning framing also offers a vocabulary for evaluating mixed-ability collaboration tools beyond simple task-completion metrics. Limitations include short-term deployment (no longitudinal data on sustained use), a Chinese urban sample that may not generalize culturally, reliance on Wizard-of-Oz simulation that sidesteps real-world latency and recognition-error handling, and an ethically complex camera-on-child design that the authors partially mitigate through child-led activation. The wearable-camera-on-child approach will also raise privacy questions in Western deployment contexts.
Tags: parents with visual impairments · mixed-ability collaboration · family informal learning · large language model · conversational guidance · joint attention · wizard of oz · symbiotic learning · assistive technology · blind parenting