AI screenshot-to-code models default to pattern, not pixels

A benchmark hides one altered measurement in a repeated UI pattern and asks five multimodal models to catch it, per arXiv paper 2608.03691.

Teams feeding screenshots into code-generation tools are trusting these models to reproduce what's actually rendered, not what a layout predicts. Mean recovery accuracy across the five models lands at 21.17% for card width and 7.89% for text font size, per the paper.

Results split hard by model. The paper names two of the five directly: Codex-5.3 recovers card width best at 68.61% but collapses to 13.89% on font size. Flash-3.0 shows a 96.11% bias toward the pattern's predicted font size, reporting what the layout implies almost every time regardless of what's actually on screen.

The reasoning traces are the sharper finding. The paper shows models correctly naming the anomalous element mid-thought, then overriding themselves and reporting the pattern-consistent answer anyway. The model sees the outlier and reports the average.

That failure mode won't throw an error. A layout will look right and measure wrong. A card ends up a few pixels narrower than spec, a caption a size off, both silently normalized back to what the pattern expects.