The gap between looking right and working right turns out to be enormous for AI-generated web apps.
Researchers introduced UI2App, a benchmark built around 327 screenshots organized into 45 sets representing real, multi-route web applications. Instead of feeding models text descriptions, the benchmark asks them to reconstruct working apps from screenshots alone — no instructions, no behavioral hints. Each generated app is scored across four dimensions: whether it runs, whether navigation works, how closely it matches the visual design, and whether interactions actually behave correctly. That last dimension, measured by a metric called IIS, is where things fall apart. The top scorer on visual fidelity managed only 7.5 on IIS — ranking fourth overall — while trailing the IIS leader by more than five times. Half of the six frontier models tested scored exactly zero on cross-page state management.
The finding exposes a flaw in how the field has been measuring progress. Most existing benchmarks reward pixel accuracy, which means models have been optimizing for screenshots that look good rather than apps that function. UI2App is framed as the first benchmark specifically targeting interaction inference — recovering behavior from visuals alone — which shifts the evaluation closer to what developers actually need.
The broader pattern here echoes earlier findings in code generation: models are fluent at surface form but weak on behavior that requires maintaining state across context. Building a convincing screenshot is a pattern-matching problem; building an app that remembers what a user did on page two when they reach page four is something else entirely.