Experiments with Midjourney and DALL-E 3 show a copyright minefield

Article by IEEE Spectrum: The writers tested the two AI image generators MidJourney and Stable Diffusion, testing their abilities to generate imagery that closely resembled copyrighted material, which proves that the training data of the image generators had to contain copyrighted material. Implemented safeguards were largely unsuccessful to curb the output of potentially infringing images.

  • mindbleach
    2·
    11 months ago

    “The machine did what we asked!!!” is an increasingly common and consistently stupid headline.

    Absolutely nobody promised the draw-anything machine would invent perfectly unique images, unburdened by any prior intellectual property concerns. How the fuck would it. We fed it every image on the internet, with labels; no kidding it knows what Darth Vader looks like. And it’ll emit images of him in a Star Wars-y context, with other related elements. That’s what it’s for.

    Ask for a deer and you get a forest.

    Ask for trees and it should pick a particular kind of tree, rather than making up some whole new species that’s never been seen. Ask for images of a cartoon family, and yeah, sometimes that will be the Simpsons. What else was it supposed to show you? A lot of example trees are oak. A lot of example cartoons are the Bart.

    When you explicitly name a character and a movie, you have negative room to complain that the machine did its fucking job. Cutesy indirect descriptions are no better. You know damn well which cartoon families have yellow skin, and which sci-fi series have laser swords. Don’t type “tree with needles instead of leaves” and go “Ah-HA! Spruce!”