OpenAI has recently introduced Sora, its AI for video generation, and while the exact data sources for Sora’s training are not publicly confirmed, there are hints that it may have included streaming footage from platforms like Twitch and video game walkthroughs.
Sora, launched on Monday, allows users to create 20-second videos from text prompts or images, offering various aspect ratios and resolutions. Upon its initial reveal, OpenAI mentioned training Sora with Minecraft videos, raising curiosity about other potential game sources used in its development.
The capability of Sora reveals itself through the creation of videos that emulate different gaming styles. For instance, it generates footage akin to a Super Mario Bros. knock-off (albeit with glitches), first-person shooter experiences reminiscent of Call of Duty, and fighting games similar to those featuring Teenage Mutant Ninja Turtles. It even seems to mimic the nuances of a Twitch stream, suggesting it has been exposed to such content.
Quite notably, Sora also appears to have recreated likenesses of popular Twitch streamers like Raúl Álvarez Genes, known as Auronplay, and Imane Anys, or Pokimane. This suggests a familiarity with their content, albeit rendered with artistic flair.
However, generating certain iconic video game images isn’t straightforward. Despite inventive prompting, Sora cannot create direct clips of trademarked games like Mortal Kombat, indicating some level of filtering in place to prevent trademark infringement.
The ambiguity surrounding Sora’s training data raises legal questions, primarily centered on the inclusion of copyrighted materials in its dataset. Training AI on video game footage can lead to unauthorized use of intellectual property, as pointed out by several IP experts. They caution that feeding video game playthroughs into AI models poses potential legal risks, particularly if protected textures or gameplay mechanics are used without permission.
Currently, there are legal precedents forming around the training of AI models on copyrighted materials. Yet, success in legal battles for AI developers doesn’t necessarily protect users who might distribute copyrighted outputs unknowingly. Ensuring compliance, AI companies often embed indemnity clauses, but these might not always extend to individual users.
As AI technology like Sora evolves, its outputs raise concerns beyond copyright, including trademark issues and likeness rights. The burgeoning sphere of “world models,” which facilitate real-time video game creation, complicates these matters further, potentially challenging established norms in game development and intellectual property law.
As AI continues to absorb publicly available data to enhance its functionality, questions around fair use and the rights of original artists and developers remain areas of significant concern. Legal experts believe that AI’s impact, including transformational models like Sora, may bring profound changes to the creative industry landscape, prompting a reevaluation of existing laws and protections.
