Text is still a tell in AI images
Signs, shirts, book spines, and UI chrome fall apart when a generator has to commit to letters.
If a picture contains writing, read it. Not a glance: read it. Generators are better at short common words than they were, and still bad at a full sentence, a second language, or a word they have to repeat consistently.
Check three things. The letters have to be letters, not shapes that almost are. The same word repeated on a sign and a shirt should match. Straight lines of type should stay straight in perspective instead of bending to follow a texture.
Small UI is especially useful: a phone screen, a laptop, a scoreboard. Those regions are dense and the model often turns them into a plausible blur. A real photo of a screen can be blurry too, so blur alone is not the tell. Illegible glyphs that are trying to be a headline are.
This will keep changing. Treat melted text as a reason to look harder, and combine it with hands and light rather than using it as a single rule.