Anthropic Settlement Has Historic Implications—Beyond the Cost

Anthropic’s $1.5 billion settlement agreement with a class of authors—alleging the artificial-intelligence platform had engaged in mass copyright infringement by downloading millions of pirated books for model training purposes—might be historic for more than one reason.

A district court judge has rejected the deal as incomplete, throwing the agreement into limbo, but Anthropic’s readiness to settle the case suggests a shifting balance of power between content owners and model and app developers in the AI ecosystem.

Anthropic has abundant cash resources and could have chosen to continue the litigation, as it recently raised $13 billion at a $183 billion valuation. So, its decision to agree to what would be the largest copyright settlement in history (if approved) may indicate a lack of confidence it would prevail on its fair use defense and avoid even greater statutory damages.

That assessment may be right. The US Supreme Court’s 2023 decision in Andy Warhol Foundation for Visual Arts v. Goldsmith, together with recent lower-court decisions, reflect three doctrinal shifts that have curtailed the scope of the fair use exemption, departing from about 20 years during which courts have mostly expanded its scope.

The pending Anthropic settlement, set against the background of the Warhol decision and other rulings, suggests that the pendulum of copyright law may be swinging toward copyright owners and artists.

As the likelihood of a litigation loss to copyright owners increases, AI developers may abandon the courtroom for the negotiating table. It’s no accident that the shift in fair use jurisprudence has been accompanied by some licensing agreements between AI developers and major content companies.

Overly Broad Applications

First, the Supreme Court in Warhol cautioned against overly broad applications of the “transformative use” concept in fair use analysis that “would swallow the copyright owner’s exclusive right to prepare derivative works.”

Similarly, last year the US Court of Appeals for the Second Circuit rejected the fair use defense brought by the Internet Archive, a “digital lending” service for books, on the ground that the use wasn’t transformative and substituted for sales by publishers and authors.

Assessing Harms

Second, lower courts have re-emphasized that the fair use defense requires finding that the infringing use doesn’t cause significant economic harm to the copyright owner.

In Kadrey v. Meta Platforms, the district court reaffirmed this principle (though it dismissed the case due to plaintiffs’ failure to show market harm). None of this is novel doctrine. Rather, it constitutes a return to long-standing guidance by the Supreme Court in 1994 and 1985, holding that commercial harm is the most important element of a fair use analysis.

Commercial Struggles

Third, some lower courts have exhibited skepticism toward fair use defenses brought by commercial entities. In 2018, the Second Circuit rejected a fair use defense brought by a company that had offered subscribers a searchable database of television programming.

In 2021, Locast, a service that retransmitted broadcast television signals in exchange for “donations” from subscribers, failed in its effort to qualify for a statutory fair-use-like exemption for nonprofits. Earlier this year, a district court rejected a fair use defense brought by a firm that had trained an AI-enabled legal search tool using materials that included copyright-protected headnotes from the Westlaw database.

This shift toward a narrower application of fair use departs from about a two-decade period during which courts have mostly set a low threshold for showing transformative use and a high threshold for showing commercial harm. This approach is hard to reconcile with governing Supreme Court precedent and, as the court observed in Warhol, has allowed the fair use exemption to grow so large that it threatens the core of copyright itself.

Fair Use’s Future

To be clear, the Anthropic settlement may not be approved, other litigations between content owners and AI developers are ongoing, and the state of the law remains in flux. Nonetheless, the fact that some courts have rejected fair use defenses in the AI context, and the strong language used by some judges concerning AI developers’ unlicensed use of copyrighted content, contrast with courts’ stance toward infringement claims against platforms and other content aggregators at the onset of the digital economy.

In decisions such as Viacom et al. v. YouTube in 2010 and the “Google Books” decisions in 2014 and 2015, courts had adopted broad understandings of the fair use exemption and the Section 512 safe harbors under the Digital Millenium Copyright Act.

This also tracked prevailing views in academia and the tech community that expressed skepticism concerning the necessity of copyright in digital environments. As a result, platforms were largely immunized against liability for mass user infringement while content owners were forced to play a futile “whack a mole” enforcement game.

There remains room for debate about how copyright law should be adapted, and how licensing and revenue-sharing arrangements should be structured, in a manner that addresses the interests of all stakeholders in the AI ecosystem. That debate will play out in the courts and the marketplace.

However, it is increasingly clear that those arrangements will have to exhibit a meaningful effort to deliver adequate compensation to the artists and entities that ultimately fuel the digital ecosystem.

This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.

Author Information

Jonathan Barnett is an author, law professor, and director of the media and technology law program at the University of Southern California’s Gould School of Law.

Write for Us: Author Guidelines