Because you know, AI never gets an image wrong, right? https://arxiv.org/pdf/2401.15817 https://arxiv.org/pdf/2401.15817 Quite long paper, but rich in example

Because you know, AI never gets an image wrong, right? https://arxiv.org/pdf/2401.15817 https://arxiv.org/pdf/2401.15817 Quite long paper, but rich in examples that speak for themselves, like AI mistaking a Starbucks logo for a Ferrari one, or an airplane for a baby stroller, etc. CSAM producers will always find a way to poison images or use other tricks to conceal them. There was one in fact years ago that involved appending a .rar archive to a .jpg image without renaming the image (including the .jpg suffix): an image viewer would correctly show the image ignoring the following data, while a .rar unpacker would ignore anything preceding the .rar header, thus the image, and would extract the archived file anyway without errors. One wonders if the capability was intentional and how many times this trick has been used to exfiltrate from somewhere sensitive data disguised as kitten pictures.