After some controversy stemming from Reddit-based reports that it is not accepting games with AI art into the Steam store, Valve has clarified that it only intends to block those submissions where the artwork may infringe copyright.
"While developers can use these AI technologies in their work... they can not infringe on existing copyrights,” Valve representative Kaci Boyle told the Verge in a statement.
A recent development on Reddit revealed that Valve is willing to ban Steam games that feature questionable, AI-generated images. According to the Reddit post, user potterharry97 was trying to publish a game to Steam with 2 or 3 AI-generated assets that he thought would pass Valve's inspection.
However, Valve contacted him, saying it could not ship the game due to copyright issues with the AI-generated art in the game, saying the art assets appeared to be "relying on copyrighted material owned by third parties." Valve added, "As the legal ownership of such AI-generated art is unclear, we cannot ship your game while it contains these AI-generated assets..."
For the game to pass Valve's inspection, the developer must "own the rights to all of the IP used in the data set that trained the AI to create the assets..." according to Valve's response to the original poster. Even after the developer went back and modified his artwork by hand, Valve still banned his game due to the same copyright issues.
Valve's stance on AI-generated content will prevent a lot of developers from utilizing AI-generated content in their games (unless they want to jump ship). Based on Valve's response, developers must effectively own the rights to ALL source material used to train the AI. This is a blocker since most AI networks are trained on millions of images/assets across the web to create compelling content.
The right of AI art-generation services like Midjourney and Stable Diffusion to use copyrighted training data without permission is very much in question. Some claim that grabbing any image or text that's available on the open Internet and ingesting it for training is "fair use," while many artists say otherwise and courts have yet to decide.
Getty Images is currently suing Stability AI for using its images as training data without permission. A group of independent artists is also suing over unauthorized use of their images. The problem has caused a group of developers to build Glaze, a tool that's designed to make it harder for image generators to train on pilfered images.
While we don't know yet how courts will rule on the issue of using copyrighted images as training data, the outcome could be costly not only for the AI-generators themselves but also for anyone who relies on them to create game art. If you build a successful game that uses AI-generated art based on copyrighted images, you could find yourself on the receiving end of a future lawsuit as could a company like Valve which would distribute and profit from the game.
To calm liability fears, Adobe recently promised to indemnify enterprise customers of its Firefly image-generation service against potential copyright claims. However, Adobe's training data comes from tis own collection of images that has already secured permission to use.