Anyone who has spent enough time on social networks over the last few days has seen something like this: a collection of images that make a mash-up of two or more elements or situations, such as Greek philosophers playing Jenga, Sonic as a special guest on Friends, or Internet Explorer ascending to browser heaven. Behind these works of art is DALL-E, a neural network trained to create (the craziest) images with just a few given words.
Since 2015 it was already possible for these models to generate text from images, but their creators thought of doing it the other way around: an Artificial Intelligence capable of creating images from text. In 2022, all that is needed is to indicate DALL-E, its more powerful version DALL-E 2, or DALL-E Mini… and wait.
“We found that it has a diverse set of capabilities,” its creators, Open AI, explain on their website. “Such as creating anthropomorphized versions of animals and objects, plausibly combining unrelated concepts, interpreting text, and applying transformations to existing images.”
Curiously, like many of his creations, the name Open AI gave to DALL-E is a combination of words: “DALL” for Salvador Dalí and “-E” for WALL-E, the robot from the Pixar movie. DALL-E is a surrealist and technological invention that materializes in images anything that people can think of. DALL-E Mini is responsible for most of the strange images you’ve probably seen in your feed: they are not photorealistic, but they certainly deliver when it comes to representing what they are told.
Ever wondered what SpongeBob would look like at Coachella? You got it. Walter White from Breaking Bad in Animal Crossing? It’s a done deal. A Demogorgon as a guest on a TV show? No problem.
Dall-E and Dall-E 2 can indeed create images from text that look almost real. That’s where the ethical implications of artificial intelligence with the power to create “fake” images come in.
So far, Open AI has limited access for those who want to try DALL-E, but its “mini” version is accessible to almost anyone, through models created from open source code that is fed by alternative text from thousands of websites. This also means that the images DALL-E’s models throw up may be biased, and reproduce harmful stereotypes or violent contexts: the AI does not filter the information; it simply reproduces it.
Another problem lies in the authorship of the images: Who owns an illustration made by DALL-E of the pyramids of Egypt but based on the paintings of Vincent Van Gogh? What about those based on the work of living artists? Should artists decide whether they want the AI to make use of their style, colors, and strokes for their creations? These and other questions will come as the use of DALL-E and similar technologies become more widespread and increasingly sophisticated.
Story originally published in Cultura Colectiva
