The Motion Picture Association (MPA) and ByteDance announced a memorandum of understanding (MOU) intended to introduce global guardrails that limit how the Chinese company’s generative AI systems can use film and television properties. The agreement covers ByteDance’s Seedance video model and Seedream image models as well as outputs on TikTok, CapCut and Dreamina, the companies said on August 17, 2026.
- MPA and ByteDance signed an MOU to apply AI guardrails across Seedance, Seedream, TikTok, CapCut and Dreamina.
- The deal follows cease-and-desist letters and viral AI-generated deepfakes that triggered industry backlash in early 2026.
- ByteDance previously added watermarks, C2PA credentials, face-blocking and filters to updated Seedance/Seedream releases.
- Red-team testing and reporting indicate these controls are helpful but not foolproof; "likeness-adjacent" outputs remain a governance challenge.
- Public reporting did not disclose detailed enforcement procedures, U.S. rollout timing, or exact technical thresholds.
What the agreement says (and what it does not)
Both organisations framed the MOU as a step toward protecting creative rights while permitting continued AI development. MPA chairman and CEO Charles Rivkin said the deal “illustrates our belief that copyright is a cornerstone of the film and television industry” and that the MOU reflects shared work to “fortify those guardrails,” according to the MPA’s statement reported by multiple outlets. ByteDance general counsel John Rogovin said the company “respects the intellectual property rights” of creative industries and called the MOU a framework for continued collaboration as the technology evolves.
News reports indicate the MOU will apply to Seedance and Seedream as well as TikTok, the TikTok USDS joint venture, CapCut and Dreamina. Neither the MPA nor ByteDance published a detailed public protocol listing technical controls, thresholds or enforcement procedures in the announcements cited by reporting, and outlets note specific technical details were not readily shared.
Why the deal came now
The MOU follows months of escalation between Hollywood and ByteDance. In February 2026 the MPA and several major studios issued cease-and-desist letters alleging widespread unauthorized use of copyrighted works and likenesses by Seedance and related tools. The immediate trigger for the dispute was a fully AI-generated viral video showing Tom Cruise and Brad Pitt fighting; that clip and several similar outputs prompted warnings from the MPA and complaints from actors’ representatives, according to contemporaneous reporting.
Industry coverage and earlier reporting of ByteDance’s product changes show the company responded to the backlash by adding watermarking, content credentials and face-blocking to Seedance and Seedream in subsequent releases. For example, ByteDance relabeled advanced versions of Seedream and Seedance—Seedream 5.0 Pro and Seedance 2.5—and added features such as visible C2PA watermarks, embedded content credentials, an “advanced invisible watermarking” mechanism and filters to block generation from real faces and copyrighted characters ahead of broader rollouts, according to technology reporting from March 2026.
What the safeguards aim to do and known limits
Reported technical safeguards implemented by ByteDance before the MOU include:
- Blocking generation from inputs that contain real faces;
- Blocking unauthorised generation of clearly copyrighted characters;
- Adding visible watermarks and embedded C2PA Content Credentials to model outputs; and
- Introducing an invisible watermarking layer meant to track content after sharing or alteration.
Reporting also highlights limits that remain unresolved. Independent red-teaming and industry observers have documented ways that creative prompting can sometimes produce “likeness-adjacent” characters—outputs that evoke a real person or copyrighted figure without technically reproducing them. Those gaps reflect a broader governance challenge: the technical ability of a model can outpace rulebooks and detection methods, making enforcement and edge-case definitions a continuing concern.
Where the tools are deployed and market context
ByteDance has rolled Seedance and related models into its content and editing ecosystem rather than offering the model in isolation. Seedance’s integration with CapCut (ByteDance’s editing app) and TikTok gives the company control over model training, generation, distribution and platform-level enforcement, a dynamic observers flagged as giving ByteDance unique levers for enforcement — but also unique responsibility to make those levers effective.
Earlier product rollouts showed ByteDance initially limiting availability of enhanced Seedance versions to specific markets via CapCut (paid users in Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand and Vietnam were listed in March reporting), with the United States and India absent from early releases. Reporting did not provide a firm timeline for U.S. availability under the updated safeguards.
Implications for creators, platforms and regulators
For film and TV rightsholders the MOU represents formalised engagement with a major AI developer rather than immediate legal victory. The MPA framed the agreement as reinforcing copyright protections; ByteDance positioned it as a commitment to “responsible innovation”. Both sides said they would continue collaboration as the technology evolves.
For platform and AI teams, the deal underlines that integrated watermarking, proactive monitoring and generation filters are now expected components of responsible deployment for generative video and image systems. However, the technical community and rights holders will be watching whether detection, prompting-resistance and enforcement hold up in real-world misuse attempts.
Regulators also have an interest: the EU AI Act and similar policies require transparency about synthetic content and may influence what constitutes adequate watermarking and disclosure. Reporting suggests ByteDance’s adoption of C2PA and embedded credentials anticipates such regulatory moves, but whether the measures satisfy regulators in the U.S., Europe or elsewhere was not reported.
Outstanding questions and next steps
- Exact technical definitions and enforcement mechanisms in the MOU were not publicly released in the reporting; observers and rights holders will await further specifics from the MPA and ByteDance.
- How the parties will handle disputed cases where output is “likeness-adjacent” remains unresolved in the public reporting.
- Whether the safeguards will be applied uniformly across markets, and when advanced Seedream/Seedance versions will reach the U.S., were not specified in available coverage.
- Independent testing and transparency about false negatives/positives for watermarking and filters will be key for external confidence; the reporting notes red-team tests found some bypasses.
Timeline (based on reporting)
- February 2026: Viral fully AI-generated Tom Cruise–Brad Pitt video and related outputs trigger cease-and-desist letters from the MPA and several studios.
- March 2026: ByteDance revises Seedance/Seedream with watermarking, face-block filters, C2PA credentials and invisible watermarking; limited CapCut rollout in several markets reported.
- August 17, 2026: MPA and ByteDance announce a memorandum of understanding establishing an AI IP-protection framework covering Seedance, Seedream, TikTok, CapCut and Dreamina.
Bottom line
The MPA–ByteDance MOU formalises a collaborative approach to guarding film and television intellectual property across ByteDance’s generative AI stack. It follows a period of industry alarm and product changes at ByteDance, and it embeds known technical protections such as watermarking and face-blocking into a broader framework. Public reporting shows the MOU is a significant diplomatic step, but key technical and enforcement details have not been disclosed and independent verification of the guardrails’ effectiveness remains pending.