Adaptive technology researchers at the University of Maryland found that artificial intelligence photo-description tools often overwhelm blind users with lengthy, unfiltered audio responses rather than delivering the specific answers they actually need. The study examines interactions with assistive apps like Be My AI, highlighting design challenges for visual accessibility.
University of Maryland Researchers Evaluate Be My AI Assistive App
When a blind man aimed his smartphone camera at a frying pan and asked Are the sausages brown yet?
he did not expect an essay. The app he relied on answered that a few spots still showed a pink tint. Four repeated snapshots and questions later, his meal was finally done. What he wanted was practical guidance rather than a tedious loop of visual verification.
That conversational friction sits at the heart of findings presented on October 20, 2026, by a research team at the University of Maryland. Published as a study investigating how much control visually impaired users retain over AI descriptions, the project analyzed interactions with Be My AI, a feature integrated into the Be My Eyes platform in 2023. Be My Eyes itself has connected blind and low-vision individuals with sighted volunteers through phone cameras since 2015, but the AI addition provides instant automated answers without waiting for a human helper.
Researchers recruited 11 participants, all identifying as blind, who tested the technology in a laboratory setting before using it in their daily routines for 10 days. The study team reviewed 418 distinct conversational exchanges between users and the software to understand where current generative tools succeed and where they stumble.
Long Responses Hinder Accessibility for Blind Users
The quantitative data gathered during the evaluation revealed a ratio mismatch. While participants typically posed concise questions ranging from 10 to 20 words, the software routinely fired back with responses more than 10 times that length. Some interactive sessions ballooned to 21 back-and-forth exchanges.
Because blind users must listen to these auditory descriptions read aloud, lengthy replies force them to sit through excessive background details just to extract a single relevant fact. Part of the difficulty stems from default software behavior. Be My AI opens every new conversation by reciting a comprehensive description of the entire captured scene rather than focusing on the specific object or question the user brought to the interaction.
More than half of the study participants flagged this default behavior as a major frustration. One user searching strictly for house keys complained that the system flooded her with unwanted data. Yuehan Qiao, an INFO Ph.D. student involved in the research, noted that forcing users to sift through unnecessary speech creates real-world friction.
We are still asking blind users to do much of the work of figuring out what information is actually relevant.
Yuehan Qiao, INFO Ph.D. student via Umd
In time-sensitive scenarios, that mental and auditory burden can cause operational delays and generate safety concerns.
Users Employ Prompt Engineering to Narrow AI Focus
Faced with long-winded answers, study participants experimented with prompt engineering techniques borrowed from standard AI interfaces. When basic queries failed to hit the mark, they adopted creative workarounds to force the system to narrow its focus.
One participant explicitly explained her underlying motivation to the app, while another photographed her kitchen and simply typed, I want to remodel my kitchen, I need suggestions.
Co-author Hernisa Kacorri, an associate professor in the College of Information with an appointment in the University of Maryland Institute for Advanced Computer Studies, pointed out that systems are generating richer text but failing to honor user intent.
We find that these systems are getting better at generating descriptions, but without effectively capturing user intent, that capability often translates into lengthy responses rather than useful answers.
Hernisa Kacorri, associate professor via Umd
Additional researchers contributing to the work include Hal Daumé III, professor of computer science and director of the NSF Institute for Trustworthy AI in Law and Society; Farnaz Zamiri Zeraati, a CS Ph.D. student; and Yang Trista Cao, a former UMD CS Ph.D. student now at Amazon. Financial backing for the project came from the National Institute on Disability, Independent Living, and Rehabilitation Research alongside TRAILS.
Designing Smarter Adaptive Technology
While the participants demonstrated remarkable resourcefulness through tinkering and problem-solving, the study authors emphasize that users should not have to master complex prompting strategies just to receive a straightforward answer.
We find that these systems are getting better at generating descriptions, but without effectively capturing user intent, that capability often translates into lengthy responses rather than useful answers.
Hernisa Kacorri, associate professor via Umd