Guideline 2.3.3: why AI-faked screenshots get apps rejected

July 7, 2026 · 5 min read

There is a tempting shortcut in 2026: feed an image model your app name and let it dream up five gorgeous screenshots. The problem is that the model does not know your app, so it invents one. Invented UI in your store listing is exactly what Apple's guideline 2.3.3 exists to reject.

What the guideline requires

Guideline 2.3.3 sits in the "Accurate Metadata" section of the App Store Review Guidelines. The core requirement: screenshots should show the app in use, reflecting what users will actually experience, not concept art, not marketing renders of features that do not exist, and not somebody else's UI.

Reviewers compare your screenshots against the binary they are reviewing. If your listing shows a beautiful dashboard that is not in the app, that is a metadata rejection, and repeat metadata problems put your account on a shorter leash.

Why generated UI drifts

Image models are trained on the general shape of mobile apps, so their inventions look plausible: rounded cards, a tab bar, a chart. But they will not match your actual fonts, your actual data model, or your actual features. They also garble small text, and store reviewers read small text.

The compliant way to use AI anyway

The trick is to split the job. Let the AI do what it is good at, which is painting an attractive marketing background and headline. Then composite your real screenshot into the device frame deterministically, in code, where the model cannot touch it. The result looks designed, and every pixel of UI in the frame is genuinely your app.

This is how Storeboard renders every screenshot when you upload your app screens: the model paints the stage, your UI is placed on it pixel-perfect, and a vision model then reads the render back to verify the headline letter for letter before you ever see it. Compliance is not a review-day gamble; it is built into the pipeline.