Generating images from text might seem like magic, and until recently, even machine learning (ML) experts would agree. Two obstacles are immediately apparent: How can a computer understand language, and how can it use that knowledge to generate a new image? At the core of both these obstacles in particular and machine learning at large is prediction: Training a model to predict what the output of a function should be by giving it examples of the correct outputs for various inputs.
At the forefront of text-to-image machine learning models, the three most well-known models are DALLE-2 (named after WALL-E and Salvador Dali), Midjourney, and Stable Diffusion. DALLE-2 is developed by OpenAI, a large corporation with ties to Microsoft, while Midjourney and Stable Diffusion are developed by small research companies. These models have all been released in the last two years, but they already have wide-reaching significance. These models have inspired precedent-setting copyright lawsuits and have ignited controversy over what it means for something to be art.
A critical question that’s arisen from the usage of text-to-image models is who owns the copyright to generated images. The question was first attempted to be answered by Pamela Samuelson in 1985, who argued that “ allocating rights in computer-generated output
to the user of the generator program is the soundest solution to the dilemma.” However, many artists in the status quo feel that this solution doesn’t account for their own art being trained on. In September 2022, Reema Selhi, head of policy at the Design and Artists Copyright Society, stated that "there are no safeguards for artists to be able to identify works in databases that are being used and opt out." In January 2023, three artists issued a lawsuit against Midjourney and Stable Diffusion, stating that “these organizations have infringed the rights of millions of artists by training their AI tools on five billion images scraped from the web without the consent of the original artists.” The suit was dismissed in July 2023, but the plaintiffs were allowed to file a new suit. The current copyright status of AI-generated art is that AI-generated art isn’t copyrightable. This was upheld at a federal level by the U.S. Copyright Office on August 18, 2023.
Copyright aside, questions have been raised over whether art generated by text-to-image models can be considered art at all. This debate was amplified when Jason Allen won the Colorado State Fair’s annual art competition by submitting an AI-generated image. Allen was transparent about his artistic process, submitting under the name “Jason M. Allen via Midjourney”, but critics feel his submission was invalid. Rob Biddulph, an author and illustrator, says that AI-generated art “is the exact opposite of what I believe art to be. Harry Woodgate, a picture book author and illustrator, said: “These programs rely entirely on the pirated intellectual property of countless working artists, photographers, illustrators, and other rights holders.”
Largely, artists are calling for compensation when images are created using their art as a reference. However, doing so is not so simple. Stable Diffusion was trained on 2.3 billion text-image pairs. Once a model is trained, it never sees the images that it was trained on again; its internal weights and architecture are set. Given the size of the dataset and the training process, it is not currently possible to see exactly what images a generated image is derived from, and filtering a dataset to remove work from living artists would be incredibly difficult.
“This isn’t going to stop,” Mr. Allen said. “Art is dead, dude. It’s over. A.I. won. Humans lost.” On the other hand, Eric Groza, an award-winning creative director, said, “AI is not going to replace human creativity, It's going to enhance it.”