AI text in images: why signs, labels and logos give fakes away
Background text that almost reads as words is a strong sign of a generated image.
Language and pixels are learned differently. Recent models can write a short headline correctly when the prompt asks for it, but the small text nobody asked for, like shop signs, price tags, book spines and license plates, often comes out as letter-shaped noise.
Look past the main subject. A street scene with a dozen signs gives a model a dozen chances to slip. Real photos of the same street have text that is boring, consistent and spelled right, including the small print.
Logos are a related tell. Generated brands look familiar but belong to no company: a coffee cup with a mermaid-ish shape, a car badge that is almost a known brand.
Checklist
- Zoom into the smallest text you can find, not the largest.
- Check whether the same word is spelled the same way twice.
- Look at license plates, menus, keyboards and clock faces.
- Ask whether a logo belongs to a brand you could name.
False alarms
Foreign scripts, stylised fonts and motion blur can make real text unreadable. Gibberish in a language you can read is the stronger signal.
Test yourself
Put it into practice in the daily AI or Real game or the endless quiz.