This is utterly absurd, but also given that all of this is based on preserving the "semantic integrity" of the content (or some similar nonsense) then why are the raw bits relevant anyway? Surely the decoded data is what matters? e.g. a signature over the encoded data doesn't give you anything a signature over the actual output.
This ignores the obvious idiocy of having some AI slop machine determine the semantics of the image: what is the semantic meaning of which people are in a picture or video? Well, AI has a well established model of saying random black people are in fact a different black person, with a warrant, or is a shoplifter, etc. Those AIs would not interpret swapping a black person's face as a semantic change.
e.g. a signature over the encoded data doesn't give you anything a signature over the actual output.
I guess they don't believe in bit-reproducible JPEG decoding with hardware trigonometry, and don't want to pay the cost of software trigonometry on their own verification servers.
olliej | 23 hours ago
This is utterly absurd, but also given that all of this is based on preserving the "semantic integrity" of the content (or some similar nonsense) then why are the raw bits relevant anyway? Surely the decoded data is what matters? e.g. a signature over the encoded data doesn't give you anything a signature over the actual output.
This ignores the obvious idiocy of having some AI slop machine determine the semantics of the image: what is the semantic meaning of which people are in a picture or video? Well, AI has a well established model of saying random black people are in fact a different black person, with a warrant, or is a shoplifter, etc. Those AIs would not interpret swapping a black person's face as a semantic change.
Hooray!
k749gtnc9l3w | 21 hours ago
I guess they don't believe in bit-reproducible JPEG decoding with hardware trigonometry, and don't want to pay the cost of software trigonometry on their own verification servers.