For blind AI users, independence can mean choosing who to ask

Sana was looking for a coat. She wanted to know whether it had a belt and what colour its zips and buttons were. The shopping website’s description left those details out. In an account published by Be My Eyes in April, she described using Be My AI to fill the gap.[1]

Blind since birth, she had become a regular user after a friend demonstrated a lipstick check with the app. Her examples were ordinary: an unfamiliar hotel display, a card carrying the number for reception, the small visual choices involved in shopping. Those details explain the attraction more clearly than a promise of universal assistance.[1]

A September research paper examines the social change behind such requests. Protik Dey, Mohd Saifuzzaman and Taslima Akter interviewed 19 US adults who were blind or had low vision, primarily used screen readers and already had experience with generative AI. Selected from 33 respondents, they described their use in remote interviews lasting roughly 45–60 minutes. The workshop paper’s qualitative findings describe these participants’ experiences.[2]

Participants described moving smaller requests away from family and friends while retaining human help for consequential situations. Some felt this supported their relationships: routine assistance no longer had to occupy every exchange. The researchers found people making these decisions actively, weighing usefulness, privacy and reliability rather than handing over judgement wholesale.[2]

An answer has to be reachable

One participant could hear an AI answer through JAWS but could not locate and copy it. Another caught a measurement error through existing knowledge. These examples expose different barriers: reaching the response and judging its content.[2]

Web accessibility guidance already addresses part of the first problem. W3C explains how an application’s status messages can be announced by assistive technology without moving the user’s focus. A visible notice that something is waiting, complete or has failed should be available through the interface’s programmatic structure. Its guidance concerns those status messages, rather than requiring every newly generated word to interrupt the listener.[3]

Applied to an AI conversation, that suggests a useful division of work: announce the state of the task, then let the person explore its result. A system that repeatedly speaks over someone reading the previous answer can make access harder, even while producing more information.[3]

Be My AI’s documentation offers a concrete example of that trade-off. It says the service presents a completed reply instead of streaming words as they appear, to support the screen-reader experience. The provider acknowledges that this makes responses feel slower. Waiting time and usable listening time are different parts of the same interaction.[4]

A human connection remains available

The mobile service also keeps a route from an AI description to a sighted volunteer. A user can ask follow-up questions, add another image or choose a call. The product therefore accommodates a change of approach within the same task. Its guidance separately excludes medicine-reading and other safety-critical uses.[4]

Choosing an automated description also changes who receives the image. Be My Eyes’ September privacy policy says it and its AI providers do not train models on Be My AI submissions, questions or responses; it may analyse them for other service purposes. Recorded volunteer calls have separate provisions, including possible sharing for research and development; users can opt out of third-party video sharing without losing service access. Privacy from another person and privacy from a service are different decisions.[5]

For Sana’s shopping question, the missing information was a belt and the colour of the fastenings. Obtaining those details left the next decision where it belonged: which coat she wanted.[1]

Report an error or suggest a correction