After Cruise-Pitt Deepfake Fallout, ByteDance’s Seedance 2.5 Blocks Hollywood Rip-Offs

Artificial intelligence has prompted a significant reckoning in Hollywood over the past year, with few moments capturing this tension as vividly as the viral fight video that took the internet by storm earlier this year. This hyper-realistic, AI-generated brawl between Tom Cruise and Brad Pitt quickly spread across social media, convincing many viewers it was real footage before the truth emerged.

This clip originated from Seedance 2.0, an AI video generator developed by ByteDance, the Chinese company behind TikTok. The model produced more than just fake celebrity fights; it also created unauthorized remixes of ‘Avengers: Endgame,’ a mashup featuring Optimus Prime battling Godzilla, and even a ‘Friends’ scene reimagined with otters taking on the roles of Rachel and Joey.

The backlash was immediate and intense. The Motion Picture Association condemned the tool for what it described as unauthorized use of American copyrighted works on a large scale. Meanwhile, the Human Artistry Campaign, which includes members from SAG-AFTRA and the Directors Guild of America, labeled the launch as an attack on creators everywhere.

Now, months later, ByteDance seems to have made significant strides in addressing these concerns. According to exclusive research shared with Deadline, the company’s latest Seedance 2.5 model is effectively blocking a large majority of attempts to recreate copyrighted Hollywood material, marking a substantial improvement from the chaotic debut of its predecessor.

The testing was conducted by VN.ai, a startup focused on stress-testing generative AI tools for copyright vulnerabilities. Their findings indicate that Seedance 2.5 blocks approximately 80% of prompts involving copyrighted material, showcasing a notable enhancement in consistency compared to Seedance 2.0 and earlier ByteDance models.

Interestingly, the model’s method of blocking content does not rely on simple prompt filtering. VN.ai explained that Seedance 2.5 generates the video first, then moderates the final output to prevent it from being served to the user. This approach implies that the underlying model has been trained on copyrighted material, as it needs to identify what it created before deciding to suppress it—a detail that may still concern rights holders, despite improved blocking rates.

However, the system is not infallible. VN.ai reported that it was still able to generate a highly accurate depiction of Oscar, the fish character from DreamWorks’ ‘Shark Tale,’ without ever directly referencing the character’s name or the film’s title, indicating that determined users can still find ways around the system.

A ByteDance spokesperson addressed these findings, stating that the company launches its models with AI transparency measures and safeguards designed to prevent unauthorized use of intellectual property. They continue to work on strengthening these protections in collaboration with creative industry partners, echoing the company’s earlier commitment to enhance safeguards following the fallout from Seedance 2.0’s launch in February.

Seedance 2.5 was launched in late July to considerable enthusiasm from AI creators, who were already impressed by the raw capabilities of its predecessor. This meant ByteDance had to balance improving copyright protections while maintaining the tool’s appeal to its user base.

Whether an 80% blocking rate will suffice to satisfy Hollywood studios and talent unions, who remain cautious about AI’s encroachment into their industry, is still an open question, especially given the workarounds that researchers have been able to discover.

Have something to add? Let us know in the comments!