Researchers tested a virtual AI guide with blind and low‑vision users in a VR setting, finding that participants used it purely as a tool when alone but began to treat it as a companion—giving it nicknames and feeling friendly—when other avatars were present. This demonstrates that the social context of virtual reality reshapes how visually impaired users perceive and interact with AI aids, suggesting designers can create guides that shift between functional and social roles to improve accessibility and reduce isolation.
When alone, blind and low‑vision users treat a large‑language‑model‑powered guide as a tool, but in social settings they anthropomorphise it and engage it as a companion.
Collins, Lin, Liu and colleagues at arXiv investigated this dynamic by recruiting sixteen blind or low‑vision participants to navigate social virtual‑reality environments. In each session, the participants interacted with a virtual guide powered by a state‑of‑the‑art large‑language model. Confederates—actors trained to behave as other VR users—were present in half of the sessions to simulate a typical social experience. The researchers recorded audio, gaze, and interaction logs while participants asked the guide questions about navigation, object identification, and social etiquette.
The study revealed a striking shift in user perception. When the participants were alone, they approached the guide as a functional tool: they asked for directions, requested descriptions of objects, and relied on the guide’s factual responses. In contrast, when confederates were present, participants began to attribute a more human‑like identity to the guide. They gave it nicknames such as “Buddy” or “Guide‑Pal,” rationalised occasional inaccuracies by citing the guide’s “virtual appearance,” and even encouraged the confederates to converse with the guide. One participant remarked that the guide “felt like a friend” when other avatars were nearby, a sentiment that did not surface in solitary trials.
These findings matter because they provide the first empirical evidence that the social context of virtual reality reshapes how blind and low‑vision users interact with AI‑powered aids. Prior proposals for “sighted guides” existed, but without user studies they remained theoretical. By demonstrating that BLV users are willing to anthropomorphise a purely digital assistant, the work signals that designers can treat such guides as social agents rather than mere tools. This has concrete implications for accessibility: a guide that can seamlessly switch between providing reliable information and offering companionship may reduce isolation and increase engagement in VR experiences for visually impaired users.
The authors also distilled several design recommendations. First, guides should support user‑defined nicknames and consistent vocal cues to reinforce a stable identity. Second, error handling should be transparent; when the guide cannot answer, it should explain its limitation rather than offer false certainty. Third, the guide’s behavior should adapt to the presence of other avatars, perhaps by modulating its level of assertiveness or offering to facilitate introductions. Finally, designers should provide options for users to toggle between a “tool mode” and a “companion mode,” allowing them to calibrate the guide’s social presence to their comfort level.
Future research must explore how to balance the guide’s reliability as an informational tool with its warmth as a social companion, ensuring that BLV users receive accurate assistance without compromising the supportive experience that social interaction can afford.