Can AI Spot Problems with Apps’ User Interface Design Better than Users?
July 06, 2026
Two researchers at LSU set out to answer this question: If you hand AI the raw code behind an app’s home screen, can it tell you what’s broken in the user interface design before a single user ever taps a button? The answer is (mostly) yes.
What’s in a login screen? ... A lot, as it turns out, from a user interface design perspective. The researchers built an AI tool that identified more than twice as many usability issues as human reviewers. In this example, both AI and human reviewers detected the absence of a progress indicator after a user presses the login button. Both also identified accessibility concerns, including the lack of support for screen reader devices. Beyond these shared findings, the AI tool detected the lack of real-time email validation, where users entering an invalid email address would not receive an error message until after entering their password and attempting to log in. It also uncovered a security issue in the screen’s source code, where password data was stored in a text editing controller longer than necessary.
The experiment
Shristi Shrestha, a doctoral student, and Anas “Nash” Mahmoud, an associate professor in the LSU Division of Computer Science and Engineering, focused on apps built with Flutter, a popular Google toolkit that lets developers write an app once and ship it to both iPhones and Android phones. They built a free web tool that takes a screen’s code, feeds it to OpenAI’s GPT-4o with a short, plain-English instruction, and gets back a list of problems along with suggested fixes.
Then came the real test. They ran their tool on 15 screens pulled from actual open-source apps—a crypto wallet, a nutrition tracker, a note-taking app, and others—and recruited 15 professional Flutter developers to judge whether the AI findings were actually correct.
What the AI caught
The researchers’ tool hunted for two kinds of trouble.
Usability problems make an app confusing or annoying: no loading spinner, so you can’t tell whether the app froze; error messages that don’t explain what went wrong; buttons that behave in unexpected ways.
Accessibility problems, meanwhile, lock people out: missing labels screen readers need for blind users, tap targets too small for people with limited dexterity, text colors too faint to read.
When the AI flagged a problem, the human developers agreed it was real about 82 percent of the time. Moreover, the AI consistently found more problems than the developers did—roughly twice as many usability issues per screen. And most of its suggested fixes were judged complete or at least a useful starting point. Almost none were flat-out wrong.
The twist
The team assumed giving AI a detailed, expert-style checklist would sharpen its eye. It did the opposite. The elaborate checklist made the AI better at routine box-checking but worse at catching the deeper, more creative problems. Once it had a list, it stopped exploring and just ticked boxes. The plain, simple request won.
The likely reason: a well-trained model already knows the best practices, so over-instructing it just boxes it in—a useful lesson for anyone who assumes more detailed prompts are always better.
The catches
The researchers’ tool wasn’t flawless. It sometimes overreacted, warning about risks developers considered trivial. And the people who tried it raised a practical worry: pasting your code into a web tool feels risky, especially when the code holds passwords or keys.
The takeaway
GPT-4o can act as a fast, cheap first-pass reviewer in app design—especially valuable for small startups. It spots real problems people miss and does its best work when you just ask.
Anas “Nash” Mahmoud
What the researchers are saying
“This study is part of a multi-stage research project that investigates how free-tier AI tools can be effectively leveraged within agile software development environments, particularly in settings constrained by limited manpower and resources,” Mahmoud said.