A peer-reviewed ASSETS 2026 paper describes a small, exploratory study in which five tech-savvy blind or low-vision co-designers used an AI coding helper to build personal camera tools. Over two months they created 37 camera-based assistants, including tools for spotting a possible Uber and for describing hand gestures.
This is a proof of concept, not a product already in wide use. The researchers built ProgramAT, a mobile interface that lets someone describe a needed tool in everyday language, then uses an agentic coding system, including GitHub Copilot, to create and revise a camera app that can talk, buzz, or display text.
Some of the tools addressed needs that the co-designers said were unmet by existing commercial assistants, such as identifying Uber rides or interpreting hand gestures.
The value, in other words, was not a single killer app. It was the chance to make a helper that fit a personal, real-world gap.
What the study looked like
The authors include Ellie Seehorn, Gene S-H Kim, Aziz Zeidieh, Ather Jammoa, Kun Lee, Aditi Shah, Jaylin Herskovitz, Anhong Guo, and Venkatesh Potluri, writing from the University of Michigan and partner institutions. They worked with five co-designers over two phases. Across the project there were 68 iterations of the tools, plus diary entries and more than 24 hours of interviews.
The co-designers were not a random sample of all blind or low-vision people. They were recruited as tech-savvy collaborators, which helped the team learn what is possible when the technical bar is lower, but it also means the results may not extend to people with less coding or AI experience.
The process was not always smooth. The paper reports challenges such as limits in the underlying models, conflicting instructions, and trouble getting a tool to work as intended. People found ways around some of those problems, including careful testing and stepwise requests, but the authors are clear that better conversational support, ways to share tools, and more specialized models would still be needed.
Why this matters for people
Many camera apps for blind users are general-purpose. They can describe a scene, but they may not do the one odd, important task you need today, such as recognizing the car you booked or following a gesture in a conversation. If people can describe that need and get a working helper, they do not have to wait for a company to build it. That kind of personal control is rare in assistive technology, where people too often have to adapt themselves to a general tool that almost, but not quite, fits.
Limits to keep in view
This was a very small exploratory study, with five co-designers. It is the right scale for co-design, and the wrong scale for claiming that anyone can now build a reliable camera tool. The helpers depended on imperfect models, and the authors note that people who cannot see the camera view have a harder time knowing when a tool is wrong.
The work does not show long-term safety, accuracy in every setting, or ease of use for people who are new to this kind of software. It is a carefully observed prototype study, rich in detail and still early in what it can promise to a wider group of users.
Looking ahead for camera tools that people can make themselves
For blind and low-vision users, the people-first promise is independence: a way to make a camera helper that fits your life, not a generic description of the world. In this small study, that promise started to look real, at least for experienced users working closely with researchers over many weeks of making, testing, and revising.
The best next step is to test the same idea with a wider group of users, not only the most technically confident ones.
You can read the paper through ACM. The authors also argue that communities will need ways to share, test, and label these homemade tools, so that one person’s useful Uber finder does not remain a one-off experiment.
If later tools make this kind of building safer and more widely usable, more people may be able to invent the small, personal assistants they have been waiting for.