ChatGPT sees something else in this image

A blogger named Gavin describes a way to make ChatGPT see a different picture from the one people see, using animated PNG files. The post relies on two images that an animated PNG can carry: the animation, which can be a single frame set to loop indefinitely, and a static fallback image. People viewing the file see the looping frame, while ChatGPT, in the author's tests, described the fallback. The author went looking for this out of worry about AI tools altering the author's own likeness, after earlier attempts that altered pixels did not work well.

The method is fragile. Copying and pasting the image or taking a screenshot removes the hidden image, so the original file has to be uploaded, and the author notes that file size changed what ChatGPT saw, without pinning down a threshold. Hidden instructions placed in an image worked inconsistently: in one test, the model followed the hidden prompt and wrote a script instead of the todo app the visible prompt asked for. The author concludes that the trick is an interesting find but does not yet solve the problem of people using someone's likeness in AI-generated images. These are informal results from one person's tests, and the example images themselves were not reviewed here.