Midjourney AI trained using a database of 16,000 artists

Many people use the new year as an opportunity to make resolutions and improve habits. However, for Midjourney, the beginning of 2024 brought about a different challenge – dealing with a list of artists whose work was used to train its generative AI program, – ArtNews reports.

During the New Year’s weekend, artists shared a Google Sheet on social media platforms X (previously known as Twitter) and Bluesky. They claimed that the sheet revealed how Midjourney had created a database of various time periods, styles, genres, movements, mediums, techniques, and thousands of artists to train its AI text-to-image generator. Jon Lam, a senior storyboard artist at Riot Games, also shared several screenshots showing Midjourney software developers discussing the development of an artist database to train its AI image generator to mimic.

Midjourney’s AI image generator was trained using a 24-page list of artists’ names, including prominent modern and contemporary artists, commercially successful illustrators for companies like Hasbro and Nintendo, and artists who contributed to the trading card game Magic the Gathering. The list, which was included in an amendment to a class-action lawsuit against Stability AI, Midjourney, and DeviantArt, also featured a six-year-old child who had contributed art for a charity event.

Artist Phil Foglio urged other artists to check the list and seek legal counsel if necessary. Although access to the original Google file was restricted, a version was uploaded to the Internet Archive. The lawsuit amendment and 455 pages of additional evidence were submitted on November 29, following the dismissal of several claims by a California federal court judge in October.

The class-action copyright lawsuit was initially filed almost a year ago in the Northern District of California. In September, the US Copyright Review Board ruled that an image created using Midjourney’s software could not be copyrighted due to its production method, causing concern among artists. This led to researchers from the University of Chicago developing a digital tool to help artists “poison” large image sets and destabilize text-to-image outputs.