A blind man points his phone at a frying pan and asks: "Are the sausages brown yet?" The app answers aloud that in places they are still pinkish. He takes a new photo and asks again - about four times in all, until the food is ready. But what he actually needed was not to repeat the question but to get advice: "Maybe they are still pink - stir them and wait about five minutes."
This man was one of 11 participants in a University of Maryland (UMD) study presented at the ACM Conference on Human Factors in Computing Systems. All participants identify as blind. The researchers wanted to understand how well they can control the responses of an artificial intelligence (AI) app that describes photos, and what they do when the answer does not fit.
"We see that these systems are getting better and better at generating descriptions, but without effectively accounting for the user's intent, this capability often turns into long answers instead of useful ones," says co-author Hernisa Kacorri, an associate professor in the UMD College of Information.
The study also involved Hal Daume III, a professor of computer science and director of the Institute for Trustworthy AI in Law and Society (TRAILS), Farnaz Zamiri Zeraati, Yuehan Qiao and Yang Trista Cao.
The Be My AI app is part of the Be My Eyes service, which since 2015 has connected blind and low-vision people with sighted volunteers through a phone camera. Be My AI, which appeared in 2023, answers questions about photos instantly, without human involvement.
Participants learned the app in the lab, then used it in everyday life for 10 days and afterwards discussed their experience with the researchers. The team analyzed 418 conversations with the app. It turned out that participants usually asked questions 10-20 words long. The app's answers contained more than 10 times as many words, and conversations sometimes stretched to 21 exchanges. For blind people this is tiring, because the answers are read aloud: they have to listen to everything to find one useful detail, and ask again.
Part of the problem is built into the system itself. Be My AI, like many similar assistive apps, begins each dialogue with a full description of the photo rather than the part of the scene the user asked about. More than half of the participants called this a serious problem. "I was given a bunch of information that I did not need," said one user who simply wanted to find his keys.
"We are still asking blind users to do most of the work of figuring out what information really matters," says Qiao. "In situations where time is limited, this can not only delay action but also create a safety risk."
People adapt to the app's limitations. When simple questions were not enough, they showed ingenuity: one explained to the app why he was asking the question. Another photographed the kitchen and wrote: "I want to remodel the kitchen, I need suggestions." According to one participant, if you state the goal, it "sharply narrows the corridor... and you get to the finish faster."
They also broke big tasks into small questions. One woman, checking bread dough, asked about its size, surface, air bubbles and flour separately. It took more time but was "the only way to make sure everything is 100% right."
The researchers emphasize: although participants showed skill in solving problems, wrestling with technologies and adapting them, users should not have to rely on these skills to get a usable answer. A complex querying strategy is "a workaround for an existing limitation of the technology."