News

What happens when one of the world’s largest technology companies needs millions of hours of video content to train its next-generation artificial intelligence model? According to a newly filed class action lawsuit, Amazon allegedly took a shortcut that could reshape how AI companies acquire training data—and expose them to significant legal liability.

On April 3, 2026, a group of YouTube content creators filed suit against Amazon in the United States District Court for the Western District of Washington, alleging that the tech giant systematically circumvented YouTube’s security measures to scrape millions of copyrighted videos for use in training its Nova Reel AI video generation platform. The complaint raises fundamental questions about the boundaries of lawful data acquisition in the age of generative AI, and it may establish important precedent for content creators seeking to protect their digital intellectual property.

The Parties and What’s at Stake

The plaintiffs include Ted Entertainment, Inc. (the company behind the popular YouTube channels “h3h3 Productions” and “H3 Podcast Highlights”), individual content creator Matt Fisher (known as “MrShortGame Golf”), and Golfholics, Inc. These creators represent a proposed nationwide class of YouTube content producers whose videos were allegedly accessed and used by Amazon without authorization. Collectively, the named plaintiffs have amassed billions of views and millions of subscribers on the platform.

The defendant, Amazon.com, Inc., launched Nova Reel in December 2024 as part of its suite of commercial AI foundation models. Nova Reel is a text-to-video AI model commercially available through Amazon Bedrock, where enterprise customers pay per second of video generated. The commercial nature of this product sits at the heart of the plaintiffs’ claims.

The Core Allegations

The complaint centers on a single federal statute: the anti-circumvention provisions of the Digital Millennium Copyright Act (DMCA), specifically 17 U.S.C. § 1201(a). This provision prohibits bypassing technological protection measures (TPMs) that control access to copyrighted works—essentially making it illegal to pick digital locks, even if no traditional copyright infringement follows.

According to the plaintiffs, YouTube does not simply host videos for public viewing. Instead, YouTube employs a sophisticated suite of technological protections designed to prevent users from downloading the underlying video files. These include:

• A “rolling cipher” encryption system that scrambles the location of video files
• IP-based monitoring and rate limiting to block automated high-volume access
• Session-bound, short-lived URLs that expire after brief windows
• CAPTCHA challenges triggered by suspicious activity patterns
• Proof-of-origin tokens that verify requests from authorized client environments

The plaintiffs allege that Amazon circumvented each of these protections to download videos at scale. On information and belief, Amazon used descrambling tools such as yt-dlp (an open-source program designed to bypass YouTube’s encryption), deployed virtual machines to rotate IP addresses and evade detection, and programmatically renewed session credentials to maintain continuous access to content.

The complaint further alleges that Amazon relied on several industry-standard machine learning datasets—including HD-VG-130M, HD-VILA-100M, Panda-70M, and HowTo100M—to identify which YouTube videos to scrape. These datasets contain references and location identifiers pointing to hundreds of millions of YouTube video clips but do not contain the actual video files themselves. To use them, a company must independently download each referenced video from YouTube—a process the plaintiffs contend necessarily requires circumventing YouTube’s protective measures.

Notably, several of these datasets carry explicit license restrictions limiting their use to academic or non-commercial research purposes. The complaint alleges that Amazon knowingly disregarded these restrictions in developing a commercial AI product.

The Significance of This Case Beyond the Parties to the Proceeding

This lawsuit is significant for several reasons that extend well beyond the immediate parties.

First, the plaintiffs have strategically grounded their claims exclusively in the DMCA’s anti-circumvention provisions rather than traditional copyright infringement. This sidesteps a major obstacle: most YouTube videos are not registered with the U.S. Copyright Office, typically a prerequisite for copyright infringement claims and statutory damages. The DMCA’s focus on whether access controls were bypassed—rather than on registration status—makes it a potentially powerful tool for content creators whose works are protected by platform security but not formally registered.

The distinction between “not registered” and “not copyrighted” is critical here. Under U.S. law, copyright protection arises automatically the moment an original work is fixed in a tangible medium—no registration, notice, or other formality required. Registration is a separate, voluntary administrative process that confers procedural and remedial benefits but is not a prerequisite to protection. The DMCA anti-circumvention provision prohibits circumventing measures that control access to “a work protected under this title”—meaning protected under Title 17, the Copyright Act—not “a work registered under this title.” Because virtually every original YouTube video meets the minimal originality threshold and is fixed upon upload, it qualifies as a protected work from the moment of creation, regardless of registration status.

This matters because traditional copyright infringement claims face steeper barriers. Under 17 U.S.C. § 411(a), registration is generally required before filing an infringement suit for U.S. works. More significantly, 17 U.S.C. § 412 limits statutory damages and attorney’s fees to works registered before infringement began or within three months of first publication. Most YouTube creators never register their videos, which would leave them proving difficult-to-quantify actual damages in an infringement suit. Section 1201’s independent statutory damages regime under Section 1203 avoids this problem entirely, allowing creators to seek $200 to $2,500 per act of circumvention without regard to registration timing. For class treatment, this is transformative: registration status would inject individualized issues into an infringement class, potentially defeating predominance, whereas a Section 1201 class can treat all original videos as protected works without reference to Copyright Office records.

The narrow category of truly “uncopyrighted” works—public domain content, U.S. government works, or material lacking any originality—would fall outside Section 1201’s reach because there is no copyright owner whose authority defines the access that a technological measure controls. But given the volume and variety of creator content on YouTube, such carve-outs are unlikely to be substantial enough to undermine the theory.

Second, the lawsuit could establish important precedent regarding the legal boundaries of AI training data acquisition. Amazon’s Nova Reel System Card reportedly acknowledges that the model was trained on “open source datasets” and “publicly available data.” The plaintiffs argue that “publicly available” does not mean “lawfully obtained,” and that the mere fact a video can be viewed through a web browser does not authorize its bulk extraction for commercial AI training. YouTube’s own CEO has publicly stated that unauthorized scraping violates creators’ rights and the platform’s Terms of Service.

Third, the class mechanism proposed here could encompass thousands of affected content creators, with damages potentially calculated on a per-violation basis. The DMCA provides for statutory damages ranging from $200 to $2,500 per act of circumvention under 17 U.S.C. § 1203(c)(3)(A), and the complaint alleges that each video clip downloaded represents a separate circumvention event. Given the scale alleged—hundreds of millions of clips across four major datasets—the potential exposure is staggering.

Key Litigation Hurdles for Plaintiffs

Class certification will be contested. Plaintiffs must show that common questions predominate over individualized issues under Rule 23(b)(3). Amazon will likely argue that determining whether each class member’s videos were accessed, whether the same protections applied at relevant times, and whether any licensing arrangements existed requires individualized proof. Establishing a reliable, classwide methodology to identify affected videos—without devolving into clip-by-clip inquiries—will be central to predominance and manageability. Plaintiffs may rely on dataset identifiers, internal logs, or forensic analysis, but the court will scrutinize whether those methods are consistent and admissible. Ascertainability poses a related challenge: if the records needed to identify class members are fragmented across third-party datasets, Amazon’s systems, and YouTube, the court may find the class unmanageable. Amazon may also challenge typicality and adequacy, arguing that variations in content type, timeframe, or licensing arrangements distinguish the named plaintiffs from the broader class.

On the merits, plaintiffs must prove that YouTube’s measures “effectively controlled access” to the copyrighted works within the meaning of 17 U.S.C. § 1201(a)(3)(B)—meaning that the measures, in their ordinary course of operation, required authorized information or processes to gain access—and that Amazon circumvented them. Amazon will likely counter that it did not “circumvent” because the content was publicly viewable, that access occurred through standard protocols, or that third parties—not Amazon—performed any circumvention. Plaintiffs will need credible technical evidence linking the alleged tools and methods directly to Amazon.

Damages present additional complexity. The complaint seeks statutory damages on a per-clip basis, which allows recovery without proof of actual economic harm—a significant advantage for class treatment. Amazon will contest that theory, arguing for a narrower unit of violation that would reduce the aggregate exposure. Amazon may also assert good-faith reliance on industry-standard dataset practices and raise statute of limitations defenses to narrow the class period. Ultimately, plaintiffs’ success will turn on the strength of their technical proof and the court’s willingness to accept scalable, classwide methodologies for identifying affected content and measuring damages.

Practical Takeaways for Businesses

This case offers several important lessons for companies operating in the AI space or managing digital content:

For AI developers: The provenance of training data matters. Reliance on “publicly available” or “open source” datasets does not insulate a company from liability if those datasets require circumventing technological protections to access the underlying content, creating potential legal exposure. Companies should conduct rigorous due diligence on training data sources, including reviewing any license restrictions and assessing whether data acquisition methods may trigger DMCA or similar legal exposure.

For content creators: This lawsuit illustrates that platform security measures can provide meaningful legal protection even absent formal copyright registration. Creators should understand the protective mechanisms employed by the platforms they use and consider whether their works might have value to AI developers seeking training data.

For platform operators: The case underscores the importance of robust technological protection measures and clear terms of service. YouTube’s layered security approach forms the factual foundation for the plaintiffs’ DMCA claims. Platforms that rely on similar protections may find themselves as fact witnesses or even interested parties in analogous disputes.

Looking Ahead

This litigation is still in its earliest stages. Amazon has not yet filed a responsive pleading, and no class has been certified. The complaint acknowledges that the full extent of Amazon’s data acquisition practices is “not yet known and will be the subject of discovery.” The next phases will likely center on motions to dismiss, class certification briefing, and discovery focused on technical systems, datasets, and acquisition methods. Given the certification and merits challenges outlined above, early expert work on both sides will be crucial. Courts will be looking for rigorous, scalable methodologies, not just high-level narratives.

If the plaintiffs succeed, this case could fundamentally alter the economics of AI development by establishing that technology companies cannot simply harvest protected content at scale and label it “publicly available.” It may also open the door to similar claims against other AI developers who have relied on YouTube and other protected platforms as sources of training data without obtaining proper licenses. If, however, the court finds that individualized proof predominates or that the measures at issue do not qualify as effective access controls, the case could set limits on the use of the DMCA in disputes over AI training data.

For now, the case stands as a stark reminder that the explosive growth of generative AI does not suspend the ordinary rules governing access to copyrighted content. As AI companies race to build ever more sophisticated models, the question of who owns the raw material—and what legal protections apply—will only grow more urgent.