Blind and Low-Vision Users Are Choosing When AI Can Help

For blind and low-vision people, a tool that reads a label or describes an image can do more than save time. It can change who has to be asked, what has to be shared and which decisions remain private.

A new qualitative study makes that shift visible without pretending to measure everyone. Researchers interviewed 19 U.S. adults who were blind or had low vision and already used generative-AI tools. Participants described using AI for smaller, frequent tasks that once required another person—while deliberately returning to people for medical, financial and navigation-critical questions.

That is a useful form of agency: not replacing human help, but having more choice about when to ask for it.

Research references: [1]

Another option for the everyday ask

The study recruited primarily through the National Federation of the Blind, with some snowball sampling. From 33 respondents, researchers selected 19 based on their experience and frequency of use. All primarily used screen readers; most were moderate or frequent generative-AI users, and seven reported using it several times a day.

In the interviews, participants described using general-purpose systems such as ChatGPT, Gemini, Claude, Copilot, Perplexity and Meta AI, alongside assistive tools such as Be My AI and Seeing AI. The uses included reading and describing images or documents, drafting messages and finding everyday information.

The important change was social as well as technical. Participants reported that AI could absorb some of the small requests previously directed to family or friends. That did not erase relationships. It let people reserve those relationships for situations where human attention, judgment or accountability mattered more.

Research references: [1]

Human help still matters

Participants did not describe trust as an on-off switch. They drew boundaries.

For lower-stakes work—drafting an email, organizing a presentation or getting a first description—an AI response could be a useful starting point. For higher-stakes questions involving medical instructions, financial documents or navigation, participants generally preferred a person.

That distinction matters because a fluent answer can still be wrong. A person can express doubt, look again or be asked how they reached a conclusion. An AI system may present an error with the same confidence as a correct response. The study’s participants were already doing the risk assessment that product design too often leaves to the user.

A good system should support that judgment rather than blur it. It should make uncertainty explicit, distinguish a stalled or failed process from a completed answer, and offer conservative modes for consequential tasks.

Research references: [1]

Accessibility is the whole interaction

Model capability is irrelevant if the interface cannot be used.

Participants reported barriers such as unlabeled prompt fields, unclear buttons and output that a screen reader could read but that was difficult to locate or copy. In those cases, the problem was not merely inconvenience. The conversation could not happen at all.

That makes accessibility a system-level requirement, not a final checklist item. Prompt entry, progress states, error messages, response navigation, copying and privacy notices all need to work with screen readers. Improvements in one part of the experience do not compensate for a broken path elsewhere.

The paper’s design recommendations follow from that point: communicate failure and uncertainty through nonvisual signals; preserve accessibility across the full product; and let people choose task-sensitive modes that foreground caution when the stakes rise.

Research references: [1]

Privacy before upload

Agency also depends on knowing what must be disclosed.

A person may upload a photo of a document, a room or a medical form to receive a description. Participants weighed that convenience against uncertainty about retention, reuse and company practices. Some accepted the trade-off because a tool was useful or accessible; others avoided companies they distrusted.

The design implication is straightforward: explain data handling before the image, document or recording is shared, in language and controls that work with screen readers. A privacy policy discovered afterward is too late to support an informed choice.

Research references: [1]

What this evidence can—and cannot—show

This was a semi-structured interview study, not a randomized trial, benchmark or population survey. The researchers selected 19 experienced users from 33 respondents. Participants lived in the United States, spoke English and primarily used screen readers. Most already used generative AI moderately or frequently. The study therefore shows how this particular group described trust, substitution and risk; it does not establish prevalence among blind and low-vision people or prove that any named product is accurate, safe or accessible.

The interviews were conducted remotely after university institutional-review-board approval. Participants received consent information, gave verbal consent and were paid a $30 gift card. Two researchers used reflexive thematic analysis and met regularly to refine their shared interpretations.

The paper was submitted to arXiv on September 14, 2026, as version 1 and is identified as a CSCW 2026 workshop proceedings paper. Its most durable contribution is not a score. It is a design principle: useful AI should widen a person’s options while preserving the right to choose human help.

That is the standard worth building toward—more ways to act, clearer signals about risk and no requirement to trade independence for safety.

*Disclosure: Next for Humanity independently reviewed the paper and its arXiv record. This article was prepared with AI assistance and human editorial review. It is not medical, financial or navigation advice.*

Research references: [1] [2]