Investigating the Role of Agentic AI in Facilitating Travel Planning for People with Low Vision
Ranran Ding, Maryam Bandukda · 2026 · Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26) · doi:10.1145/3772363.3798364
Summary
This CHI 2026 Extended Abstract examines a stage of accessible travel that most assistive-technology research has overlooked: the pre-trip planning work people with low vision (PLV) do before ever leaving the house. The authors argue that most existing tools — navigation apps, obstacle-detection aids, computer-vision wearables — are reactive and on-route, offering little support for the anticipatory labour of judging whether a trip to an open space such as a park or public plaza is even feasible. Open spaces are particularly demanding: irregular layouts, ambiguous boundaries, dynamic conditions, and visually-oriented signage that offers poor contrast or legibility for residual vision. The study used a two-stage qualitative design. First, ten semi-structured online interviews with adults with moderate or severe low vision and varying assistive-technology proficiency explored planning practices, accessibility barriers, and experiences with existing tools. Second, a two-hour co-design workshop with two of those participants used a voice-based conversational prototype (a customised GPT) to surface expectations about tone, information integration, proactivity, and control. Data were thematically analysed in NVivo following Braun and Clarke. The authors position agentic AI — proactive, contextual, voice-first, LLM-backed — not as an information-delivery tool but as a judgement-support system that can integrate fragmented accessibility information, adopt a companion-like tone, and give users layered control over how much guidance they receive.
Key findings
The analysis produced two overarching themes. Under "PLV Needs and Barriers," participants described fragmented information access across mainstream mapping tools (distance-based routes that hide steep terrain, unclear or missing accessibility signage, accessibility details scattered across platforms) and an ongoing negotiation between dependence and autonomy — sighted companions provide safety but constrain self-determination, and participants strongly preferred remaining the decision-maker even when planning was effortful. Under "Role of Technology," participants described assistive-technology reliability (network drops, battery, ~20-metre location errors) as a gating condition for trust: if a tool might fail, they simply would not travel. Expectations for agentic AI clustered into three design patterns: (1) real-time, transparently time-stamped information, so users can judge whether an update reflects current conditions such as a closed park entrance; (2) companion-like communication that augments directions with contextual landmarks ("at the traffic light, next to the crossing") rather than terse commands; and (3) layered interaction with explicit user control over verbosity — asking whether the user wants step-by-step guidance or a summary. The authors reframe autonomy as relational and information-dependent rather than an individual trait, and trust as a precondition for early-stage reliance on assistive systems.
Relevance
For practitioners building AI assistants for disabled users, this paper is a useful corrective to the assumption that more information is always better. The participants consistently asked for less but more trustworthy information, delivered with provenance and time-stamping, and wrapped in a tone that supports rather than directs. The pre-trip framing is also transferable beyond low vision: anyone managing a disability-related access cost — chronic illness, fatigue, anxiety, mobility — does substantial anticipatory work that current apps ignore. The study is small (n=10 interviews, n=2 workshop) and the prototype is a wizard-of-oz-style customised ChatGPT rather than a deployed system, so claims about long-term trust and real-world feasibility are untested. The authors appropriately call for in-situ longitudinal deployments. Practitioners should also treat the "companion-like" tone recommendation carefully: it raises well-known risks around over-trust in LLMs that hallucinate, especially for safety-critical navigation decisions.
Tags: low vision · wayfinding · agentic AI · large language models · conversational agents · voice interface · co-design · autonomy · trust · orientation and mobility