In theory turning everything into an image then compressing using lossy formats aiming at very low size should mask subtle visual artifacts, but yeah, steganogr

In theory turning everything into an image then compressing using lossy formats aiming at very low size should mask subtle visual artifacts, but yeah, steganography could be implemented in many other ways, such as slightly different fonts (sizes, line spacing, orientation, etc). I wonder if AI, could be used to analyze and find those details by comparing potentially watermarked identical documents.