The LLM is reading the bytes of the file, not looking at a picture of its rendering. File metadata exists as well, and change history. Tons of places to hide text.

Even if you processed it via a screenshot, image files are processed byte by byte as well and can contain textual metadata.