Suno, Yout, Perplexity AI and §1201: AI Training and another piece of the DMCA

“No person shall circumvent a technological measure that effectively controls access to a work protected under this title.” 17 U.S.C. § 1201(a)(1)(A)

“a technological measure “effectively controls access to a work” if the measure, in the ordinary course of its operation, requires the application of information, or a process or a treatment, with the authority of the copyright owner, to gain access to the work.” 17 U.S.C. § 1201(a)(3)(B)

Thus far, Section 1201 of the Digital Millennium Copyright Act, the part that makes it illegal to break digital locks, like DRM, has not taken center stage in the AI copyright litigation wars. 

Authors Alliance members will know that we’ve paid quite a lot of attention to this provision, because it’s one part of the law that can seriously hamper creators and researchers’ fair use rights. So, for instance, we’ve worked closely with researchers to petition for exemptions that allow for circumvention of digital locks for purposes such as academic text data mining. 

In the last few weeks, we’ve seen a spate of new developments in AI copyright lawsuits; a number of these cases raise DMCA §1201 circumvention claims, including most recently a suit brought by UMG Recordings against Suno, an AI music generation platform.

Before diving into the legal weeds of these lawsuits, it’s worth grounding ourselves in a few core principles—ideas that have consistently shaped our thinking about copyright, fair use, and the development of generative AI.

If you’ve been following our work, you’ll recognize a few themes that cut across our analysis of both copyright more broadly and AI specifically:

1. Undermining fair use—whether through terms of service or other mechanisms—is almost always a bad idea.

Fair use has been described by the courts as both a “safety valve” and “First Amendment safeguard.” Both descriptions are apt – among other benefits, fair use bolsters our ability to quote and criticize the works of others without their permission. Without it, copyright law could quickly shape our speech in deeply concerning ways. Imagine a world where you could not write an essay quoting and criticizing the work of another author without paying prohibitively high licensing fees (or not be permitted at all). Imagine a world where virtually every time you write something, you are presented with a bill at the end. Absent fair use, both of those possibilities could become reality.   

We are deeply skeptical when rightsholders attempt to shrink fair use through private contract or technical restriction rather than through law and policy. Contractual override is one of the most concerning mechanisms to control fair use that we’ve written about (see: How to Evade Fair Use in Two Easy Steps and Contractual Override: How Private Contracts Undermine the Goals of the Copyright Act for Libraries and Researchers). 

Just as we view contractual override as bad policy, undermining fair use through the DMCA is similarly deeply problematic. In both instances, fair use rights, the safety valve and First Amendment safeguard that help the whole copyright ecosystem function properly, are being eroded by areas of the law that largely ignore or imperfectly account for the careful balance struck by the Copyright Act.  

2. Training generative AI on copyrighted works is, in our view, likely a fair use.

As we’ve said before, we believe that using copyrighted works as training data for generative AI should generally be considered fair use—and courts are increasingly open to that view. Judge Alsup’s reasoning in Bartz v. Anthropic, recognizing AI training as highly transformative, is a strong example of this emerging judicial trend.

3. “Lawful access” should not be a prerequisite for fair use in the AI context.

LLMs and generative models require massive datasets to be effective. If AI developers were required to obtain permission for each work used in training or only train on material that a rightsholder agrees to provide, fair use would collapse under the weight of transaction costs. Sustainable, diverse AI development requires avoiding a “permission-first” world – that is not how fair use is meant to operate.

Why Plaintiffs Are Experimenting with New Claims

Rightsholders suing AI companies are throwing every viable legal theory at the wall, not only copyright infringement. Increasingly, we’re seeing courts express openness to the idea that training on copyrighted works can be fair use. That means the core infringement claim may not ultimately carry the day.

Plaintiffs are looking elsewhere.

Two DMCA provisions have emerged as part of the new battleground:

• §1202 (removal of copyright management information)
  
• §1201 (anti-circumvention)

We’ve discussed §1202 before (including in our amicus brief in Doe v. GitHub). Here, we focus on §1201—a provision that could significantly impact the development of AI, even if AI training is ultimately deemed fair use.

Yout: A Key Precursor

To understand how rightsholders are seeking to apply §1201 to AI, it helps to start with a look at Yout, a platform that allows users to download audio from YouTube streams.

In 2019, the RIAA sent DMCA notices that effectively got Yout delisted from Google search, making it quite difficult to find. Yout responded by suing the RIAA for defamation and seeking a declaratory judgment that its tool did not “circumvent” anything under §1201.

The district court disagreed, finding that Yout’s software did circumvent YouTube’s access controls:“Yout’s technology clearly “bypasses” YouTube’s technological measures because it affirmatively acts to “modify” the Request URL (a.k.a. signature value), causing an end user to access content that is otherwise unavailable. As explained, without modifying the signature value, there is no access to the allegedly freely available downloadable files. Accordingly, I cannot agree with Yout that there is “nothing to circumvent.” (Yout v. RIAA, Ruling on Defendant’s Motion to Dismiss at 32)

Yout is currently appealing that decision (full docket here). Its appeal has been supported by an amicus brief from EFF. Github also filed an amicus brief and published a blog post detailing the threats posed by the district court’s expansive view of Section 1201. Among those threats, the court’s view of circumvention could undermine ad blockers, screen readers, and web archivers (noting that youtube-dl itself is used preserve video files, such as “preserving evidence in the fight for human rights.”)

In a notable recent development: two music AI companies—Suno and Udio—have now filed an amicus brief in support of Yout (nearly two years after the case was argued!)  They clearly see the case as directly tied to their own.

UMG Recordings v. Suno

Suno, a popular AI music creation platform, was sued in June 2024 by major music labels. Among several claims, plaintiffs argue that Suno violated §1201 by “stream-ripping” music from YouTube to build its training dataset. 

Section 1201 has two main types of prohibitions. The first, found in 17 U.S.C. §1201(a)(1)(A) Provides a broad prohibition on circumventing technological protection measures that control access to a work: “No person shall circumvent a technological measure that effectively controls access to a work protected under this title.” The second, found in §1201(b)(1)(A) provides a narrower prohibition on trafficking in “any technology, product, service, device, component, or part thereof, that . . . is primarily designed or produced for the purpose of circumventing protection afforded by a technological measure that effectively protects a right of a copyright owner under this title in a work or a portion thereof.” The primary “right of a copyright owner” invoked here is the reproduction right, hence this provision being referred to as a “copy control.” 

Suno argues that YouTube’s restrictions are not access controls, but rather copy controls—and §1201 treats those differently.

In support of Yout, Suno’s amicus brief cites the U.S. Copyright Office’s own 1998 summary explaining the DMCA’s design, which explains the difference, and highlights how and why Congress chose to directly prohibit only access controls, not copy controls.

Because copying can be fair use, Congress did not want §1201 to become a backdoor mechanism for restricting fair use.

Suno argues: “This statutory asymmetry was intentional: Congress enacted it specifically to protect fair users. Copy controls, by definition, “protect [the] right[s] of a copyright owner,” including by preventing unauthorized reproduction of copyrighted works. But not all reproductions are prohibited by copyright law…Enacting a statute that would penalize circumvention of copy controls would, in effect, give copyright owners the de facto right to prohibit fair uses. So while Congress enacted a prohibition on the provision of devices designed to circumvent copy controls, it declined to prohibit the act of circumventing those controls, so that it would not effectively impose liability on fair users.” (Brief for Amici Suno and Uncharted Labs at 4)

This matters because YouTube’s so-called “rolling cipher,” the mechanism at the center of these lawsuits, looks far more like a copy friction than a genuine access barrier. 

The crux of Suno’s legal argument is that the barriers to copying erected by YouTube are nothing like the encryption and scrambling methods that are properly characterized as access controls. 

In line with this argument, there’s a fair amount of support for the proposition that YouTube’s barriers to copying are not meaningful access controls: ‘Deciphering’ YouTube’s Rolling Cypher in Your Browser is a Piece of Cake. The argument that these controls were not the kind contemplated by Congress in 1998 is compelling.

Will DMCA §1201 shape AI training and development? 

We hope not, but we’re starting to see litigants resort to §1201 in new AI suits. Beyond the case Suno is defending against UMG, Perplexity and several other companies are now also named in a suit brought by Reddit based on their scraping of Reddit data from Google search results.  

From the Reddit complaint: 

“These tools are aimed at bypassing two levels of security: First, evading Reddit’s own anti-scraping measures, and second, circumventing Google’s controls and scraping Reddit content directly from Google’s search engine results. In a very real sense, these Defendants are similar to would-be bank robbers, who, knowing they cannot get into the bank vault, break into the armored truck carrying the cash instead.” (Reddit  v. SerpApi, Complaint at 1)

“Recognizing that Reddit denies scrapers like them access to its site, Defendants SerpApi, Oxylabs, and AWMProxy scrape the data from Google’s search results instead. They do so by masking their identities, hiding their locations, and disguising their web scrapers as regular people (among other techniques) to circumvent or bypass the security restrictions meant to stop them.” (Reddit  v. SerpApi, Complaint at 2)

The boundary lines between access controls and copy controls will be fleshed out in cases like these. The expansion of §1201 to cover the types of activities at issue in these lawsuits would be a troubling development. It would erect a new, indirect barrier to fair use and one that Congress sought to avoid. We believe that allowing §1201 to function as a way to shut down AI training would be a profound misuse of a statute written decades before today’s AI technologies existed.

Continued AI development shouldn’t hinge on a misapplied anti-circumvention rule. If copyright owners want to challenge AI training, they should do so directly on copyright infringement grounds or through Congress. Not through a legal side door designed to regulate encryption in 1998.