It sure looks like OpenAI trained Sora on game content — and legal experts say that could be a problem

OpenAI's Sora, a video-generation AI launched recently, appears to have been trained on video game content, raising significant legal concerns. The AI can generate 20-second videos that remarkably resemble gameplay from titles like Super Mario Bros. and first-person shooters, even capturing the likeness of popular Twitch streamers like Auronplay and Pokimane. Its ability to recreate game-like scenes suggests extensive training on video game playthroughs and streaming content.

Legal experts Joshua Weigensberg and Evan Everist warn of complex copyright risks inherent in training AI on game content. They highlight that video game footage involves multiple layers of copyright protection, including rights of game developers, video creators, and potentially user-generated content creators. The probabilistic nature of generative AI means these models learn patterns by observing data, which can inadvertently reproduce copyrighted material and potentially infringe on intellectual property rights.

OpenAI remains deliberately vague about Sora's training data, only acknowledging use of "publicly available" and licensed stock media content. The company has implemented some filtering to prevent generating clips with trademarked characters, but the underlying training methodology remains a critical legal question. As AI technology advances, the emerging legal landscape surrounding AI training methodologies could have far-reaching implications for creative industries, potentially reshaping how intellectual property is understood and protected in the digital age.