New study reveals copyrighted images can be reproduced.
(Image credit: Eric Wallace on Twitter)
Another day, another AI art controversy. Concerns over ethical issues surrounding originality and copyright have followed the tech around since day one, and are showing no signs of abating. And, perhaps unsurprisingly, new research has just proven that, yes, AI art generators reproduce copyrighted imagery.
Image diffusion models such as DALL-E 2, Imagen, and Stable Diffusion have hogged headlines over the last few months thanks to their scarily detailed text-to-image generation. But while AI fans argue that everything the tools create is technically new, research has revealed that not only do they 'memorise' trademarked and sensitive images, but they're also not particularly private. (Not sure what we're talking about? Check out the weirdest art created with DALL-E 2 if you're feeling brave.)
A screenshot from the study (Image credit: Eric Wallace on Twitter)
A study by various tech experts headed up by Google Brain researcher Nicholas Carlini has discovered that Google’s Imagen tool and Stable Diffusion are both capable of exactly reproducing trademarked images from a mere text prompt. In a Twitter thread (below), researcher Eric Wallace explains how "many of these images are copyright or licensed, and some are photos of individuals."
Models such as Stable Diffusion are trained on copyrighted, trademarked, private, and sensitive images.Yet, our new paper shows that diffusion models memorize images from their training data and emit them at generation time.Paper: https://t.co/LQuTtAskJ9 👇[1/9] pic.twitter.com/ieVqkOnnoXJanuary 31, 2023
Indeed, it's clear from Wallace's examples that the tools have spat out almost exact replicas of existing images, albeit with the addition of some digital noise and distortion. While the team only found a .03% memorisation rate, as researcher Wallace told Gizmodo, "Maybe in the next year, whatever new model comes out that’s a lot bigger and a lot more powerful, then potentially these kinds of memorisation risks would be a lot higher than they are now."
A screenshot from the study (Image credit: Eric Wallace on Twitter)
Daniel John is Design Editor at Creative Bloq. He reports on the worlds of design, branding and lifestyle tech, and has covered several industry events including Milan Design Week, OFFF Barcelona and Adobe Max in Los Angeles.
