Generative AI: When Fair Use Becomes Unfair Competition

Category: Print Issues

Generative AI: When Fair Use Becomes Unfair Competition

Kristian Stout, Ben Sperry, & Subiksha Ramakrishnan

Volume 32

, Issue 4

32 GEO. MASON L. REV. 779

Introduction

The rise of artificial intelligence (“AI”) systems trained on large datasets, which often include copyrighted works, has led to thorny questions of how to apply intellectual property law in this developing space. Around the world, policymakers have been considering different approaches to striking a balance between protecting copyright holders and promoting innovative new uses of copyrighted material, such as generative AI.¹ Courts have also considered how copyright law, which in the United States includes doctrines like fair use, applies to the use of copyrighted material as inputs to train datasets.²

Assuming that Congress is not about to step in with some form of a “text and data mining” exemption in copyright law,³ there are, broadly speaking, two ways to facilitate training. First, the content can be licensed. However, the sheer scale of data needed, coupled with a very low marginal value for individual works, potentially makes this approach impractical. Second, developers hope to use “fair use” as an affirmative defense to any infringing acts. The issue here is that, technically, this constitutes an infringement in the first place—otherwise, an affirmative defense would not be necessary.

Moreover, it is not facially obvious that an excused infringement would not violate other bodies of law outside of copyright—for instance, competition law. Thus, even if copyright law permits certain uses through fair use or other exceptions, consumer protection or competition authorities could step in as well and police attempts to use copyrighted material as inputs for training as “unfair.” For instance, the Federal Trade Commission (“FTC”) offered a comment to the U.S. Copyright Office in October 2023 arguing that “under certain circumstances, the use of pirated or misuse of copyrighted materials could be an unfair practice or unfair method of competition under Section 5 of the FTC Act.”⁴ The FTC went so far as to even say “conduct that may be consistent with copyright laws nevertheless may violate Section 5.”⁵

Below, we consider the intersection between copyright and competition policy in the context of using copyrighted works as inputs for generative AI. Part I provides context by introducing an economic approach to the questions of copyright and fair use for AI inputs, including looking at cases which have been considered so far in the United States. Part II then considers the intersection of FTC Section 5 with copyright law, outlining the FTC’s proposed approach in its comment to the Copyright Office. Part III then argues the FTC’s approach would lead to contradictions between copyright and competition and consumer protection law reducing the incentives for innovation and competition that underlie both.

I. Law and Economics of AI and Copyright

To properly understand the intersection between copyright law, AI, and competition policy, a background in the economics of copyright is necessary. While this is an evolving area without established answers, the Copyright Office has conducted considerable work to frame the economic issues at play when examining the use of copyrighted works in artificial intelligence training.⁶ Courts have also begun to issue decisions in cases involving copyright and AI training.

A. The Economic Foundations of Copyright and AI Innovation

Copyright law functions as a crucial mechanism for fostering innovation by establishing exclusive property rights that incentivize creative development.⁷ These rights enable creators to prevent unauthorized exploitation of their work, providing essential economic incentives that ultimately enhance collective social welfare. Creative works, being inherently non-excludable and non-rivalrous, constitute public goods that would face significant underproduction without copyright protection, as unfettered copying would substantially diminish expected returns on creative investment.⁸

This protective framework necessarily grants rightsholders a degree of market power that can elevate prices and potentially restrict access to resources.⁹ To maintain an appropriate equilibrium, copyright systems incorporate temporal limitations and exceptions—particularly fair use—that serve as critical counterbalances promoting broader societal interests.¹⁰ This delicate balance between creator compensation and public access represents the fundamental challenge underlying copyright policy.

The system exhibits what might be characterized as a “hydraulic relationship” between various elements of copyright protection.¹¹ Similar to fluid dynamics, where pressure applied in one chamber creates corresponding movement elsewhere, strengthening protections in one domain often necessitates increased flexibility in another to preserve systemic balance.¹² This hydraulic principle becomes particularly evident when transformative technologies like AI challenge established frameworks. Imposing significant restrictions at the input stage may require corresponding flexibility regarding outputs and vice versa.

AI systems present unprecedented challenges for policymakers and courts attempting to balance creative production incentives against beneficial utilization of copyrighted materials for technological advancement. While AI training on copyrighted materials could potentially alter creative incentives, excessive licensing requirements or competition-oriented obligations could significantly impair innovation and utility in AI development.

Several fundamental principles warrant consideration in this context. First, copyright’s inherent limitations prove highly relevant to AI training scenarios. Copyright law protects specific expressions rather than underlying ideas—a distinction that is potentially significant when AI systems transform inputs into novel expressions.¹³ The fair use doctrine further permits transformative applications that generate new expressions from protected works.¹⁴ While fair use remains contentious in AI contexts, the substantial good-faith debate demonstrates the uniqueness of AI training and the difficulty of applying traditional copyright frameworks to this innovation.¹⁵

Second, requiring individualized licensing for all copyrighted works used in AI training would impose prohibitive practical challenges and transaction costs.¹⁶ Contemporary AI models commonly train on billions of text fragments and images from across the internet. Negotiating separate licenses for each protected work would necessitate millions of transactions, rendering comprehensive licensing practically impossible at the scale required for effective AI development.¹⁷

While proponents have proposed collective licensing models to address these challenges through clearinghouses that negotiate blanket licenses for rightsholder groups,¹⁸ significant obstacles persist. Not all content owners would participate in such collectives, creating coverage gaps particularly for independent creators.¹⁹ Mandatory collective licensing could potentially create concerning monopolistic structures that eliminate competitive pressure.²⁰ As the Supreme Court recognized in Broadcast Music, Inc. v. CBS,²¹ even voluntary collective licensing arrangements require careful antitrust scrutiny to ensure they do not create arrangements that “threaten the proper operation of our predominantly free-market economy.”²² In the absence of clear market failure, voluntary solutions might be preferable, as evidenced by some media companies already negotiating direct arrangements.²³ The current landscape features an inconsistent patchwork of approaches, with some major content owners demanding licenses while numerous smaller creators remain outside any collective framework.²⁴

Furthermore, determining appropriate licensing fees presents fundamental valuation challenges. AI models utilize billions of inputs, making it nearly impossible to determine the value of individual contributions during training. Given the vast volume of necessary input data, even conventionally valuable content may have extremely small marginal value in training contexts.

These practical difficulties suggest that focusing regulatory attention on outputs rather than inputs represents a more economically sound approach. This would prioritize addressing copyright violations in final outputs while maintaining the creativity-innovation balance that copyright law aims to achieve.

This output-focused approach circumvents the implementation challenges associated with comprehensive input licensing. Once outputs are generated and commercially deployed, valuation becomes more feasible in defined markets. The contribution assessment of particular inputs would rely on similarity analysis—a familiar approach in copyright adjudication. Additional protections for creators’ distinctive characteristics would facilitate commercial negotiations,²⁵ while standard copyright mechanisms would remain available to address infringing uses.

B. Recent Cases

A growing body of litigation has been examining whether AI developers may legally use copyrighted materials to train their models. Although most cases remain in early procedural stages, two significant judicial decisions provide important precedent. The following analysis examines the courts’ reasoning in these pivotal cases.

1. Thomson Reuters v. Ross Intelligence

In Thomson Reuters Enterprise Center GmbH v. Ross Intelligence, Inc.,²⁶ the Delaware District Court examined Ross Intelligence’s use of “Bulk Memos” to train its AI legal-research tool.²⁷ A third party, LegalEase, created these memos based on Westlaw headnotes. Initially, Ross sought to license Westlaw headnotes directly from Thomson Reuters but was refused because Thomson Reuters didn’t want to assist a potential competitor. As an alternative strategy, Ross purchased Bulk Memos from LegalEase—documents described as “lawyers’ compilations of legal questions with good and bad answers.” ²⁸ LegalEase directed its lawyers to create these memos using Westlaw headnotes, which Thomson Reuters had developed by extracting legal principles from judicial opinions. Ross subsequently used these Bulk Memos to train an AI system designed to compete in the legal research market.

The question at issue was whether the copies of the Bulk Memos made by Ross for the training of its AI model infringed the rights held by Thomson Reuters in the Westlaw headnotes.

First, the court considered whether the record established a case of copyright infringement. Thomson Reuters had to show that “(1) it owned a valid copyright and (2) Ross copied protected elements of the copyright work.”²⁹ This was important here because Thomson Reuters claimed to hold a copyright over headnotes that were created by quoting and summarizing parts of judicial opinions that are not copyrightable.³⁰ After concluding that (most of) the headnotes were copyrightable,³¹ the court had to determine whether copying occurred, which means Thomson Reuters had to prove “(a) actual copying, and (b) substantial similarity.”³² The court found the actual copying was present because the “Bulk Memo questions for this batch closely resemble the headnotes’ text and that the headnotes differ significantly from the text of the judicial opinions.”³³ The court, under the same reasoning, found most of the Bulk Memos were substantially similar to the Westlaw headnotes. Accordingly, the court granted summary judgment “on the headnotes [for which the] language very closely tracks the language of the Bulk Memo question but not the language of the case opinion.”³⁴

Then, applying Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith,³⁵ the court emphasized that when a new use shares the same or highly similar purpose as the original work, this strongly indicates a lack of transformation.³⁶ In this case, Ross’s use of Westlaw headnotes to develop a competing legal research tool meant the underlying purpose of the original expression remained fundamentally unchanged.³⁷ Rather than adding new meaning, commentary, or purpose, Ross merely repackaged the headnotes in a format that served virtually identical functionality—helping users locate relevant legal cases. Consequently, the court determined that factor one (which considers the purpose and character of the use) weighed heavily against a finding of fair use because the use was commercial and not transformative.³⁸

Ross attempted to draw an analogy to “intermediate copying” cases, such as Sega Enterprises Ltd. v. Accolade, Inc. ³⁹ and Sony Computer Entertainment, Inc. v. Connectix Corp.,⁴⁰ where courts permitted limited copying of functional code for the purpose of reverse engineering.⁴¹ Ross’s analogy compared AI training (converting protected works into numerical weights) with reverse engineering (copying application programming interfaces to create compatible software).

The court, however, distinguished these cases on the basis that they involved functional computer code—primarily comprising unprotected ideas—and were necessary for ensuring compatibility between systems.⁴² By contrast, the Westlaw headnotes constitute creative expression, and the copying at issue was not essential to uncover any unprotected elements.⁴³ This distinction further undercut the argument that Ross’s use could be deemed transformative.

The commercial nature of Ross’s use also contributed to the court’s adverse view. By seeking to profit from a product that competes directly with Westlaw, Ross’s use maintained the same “research” or “reference” purpose inherent in Westlaw’s headnotes. This commercial aim, coupled with the absence of any significant recontextualization of the original material, reinforced the conclusion that Ross’s copying failed to achieve the requisite transformation.⁴⁴

Ultimately, because Ross’s appropriation of the headnotes did not introduce any new meaning or commentary beyond their original function, the court held that the copying was insufficiently transformative to qualify for fair use. The convergence of an unaltered purpose and the direct competition with Westlaw’s established market for headnotes served as a pivotal basis for rejecting the fair-use defense in this case.

On factor two, which focused on the nature of the original work, the court found that the Westlaw headnotes were “not that creative” compared to the original judicial opinions, thus siding with Ross.⁴⁵ Factor three, on how much of the work was used, also went to Ross.⁴⁶ The court found that “[t]here is no factual dispute: Ross’s output to an end user does not include a West headnote. . . . Because Ross did not make West headnotes available to the public, Ross benefits from factor three.”⁴⁷

Factor four, which is the most important factor of fair use, is the “likely effect on the market for the original.”⁴⁸ Here, the court again pointed out that Ross intended to create a market substitute to compete with Westlaw. The court emphasized that while the public may have an interest in judicial opinions, the “public has no right to Thomson Reuters’s parsing of the law. Copyrights encourage people to develop things that help society, like good research tools. Their builders earn the right to be paid accordingly.”⁴⁹ On balance, the court decided that the factors weighed against the fair-use defense. This case has become the first AI and copyright case to be appealed to a circuit court.⁵⁰

Despite the Ross decision, general purpose large language models (“LLMs”) might successfully establish fair use under its market-effects analysis. Unlike Ross’s specialized legal research tool, general purpose LLMs typically operate in different markets than the original works they process during training.⁵¹ The Ross court itself explicitly differentiated between Ross’s product and generative AI systems:

Ross was using Thomson Reuters’s headnotes as AI data to create a legal research tool to compete with Westlaw. It is undisputed that Ross’s AI is not generative AI (AI that writes new content itself). Rather, when a user enters a legal question, Ross spits back relevant judicial opinions that have already been written.⁵²

Unlike Ross’s product, which directly competed with Westlaw’s legal research service, general purpose LLMs serve fundamentally different purposes than the original works in their training data. These models create new content rather than reproducing existing works, potentially transforming input data into novel outputs like stories or images.⁵³ This creative function distinguishes them from tools that merely repackage copyrighted content.

This distinction strengthens the transformative-use argument for general purpose LLMs. When an AI generates original content inspired by—but not duplicating—existing works, the transformation process substantially alters both the form and function of the source material.⁵⁴ This transformative quality presents a stark contrast to the Ross case, where the copied content served essentially the same purpose in both original and derivative forms.

However, while general purpose LLMs may not directly compete with original works, they could potentially impact derivative markets by diminishing future licensing opportunities for rightsholders. Despite this concern, under the Ross framework, the distinct market positioning of general purpose LLMs might still favor AI developers in fair-use analyses. Nevertheless, given the multifaceted approach of fair use, this market distinction alone cannot guarantee protection without careful consideration of all relevant factors that safeguard creators’ economic interests.

2. Tremblay v. OpenAI

In Tremblay v. OpenAI, Inc.,⁵⁵ the Northern District of California considered whether AI systems trained on copyrighted materials could infringe copyright law through their outputs.⁵⁶ The court evaluated a motion to dismiss focusing specifically on allegations of vicarious copyright infringement.

The central question was whether ChatGPT’s outputs violated copyright law when the underlying model had been trained on plaintiffs’ copyrighted books.⁵⁷ Unlike Ross, which focused on the copying of works during training, Tremblay addressed whether the outputs themselves constituted infringement.

The court framed its analysis around vicarious infringement,⁵⁸ first assessing whether plaintiffs had adequately alleged direct infringement in ChatGPT’s outputs. The plaintiffs argued that because OpenAI had copied their books for training purposes, “every output” of ChatGPT automatically qualified as an infringing derivative work. However, the court rejected this reasoning, emphasizing that copyright infringement requires proof of both actual copying and unlawful appropriation (substantial similarity).⁵⁹ Critically, the “[p]laintiffs fail to explain what the outputs entail or allege that any particular output is substantially similar—or similar at all—to their books.”⁶⁰

This decision highlights an important principle: Copyright infringement should be evaluated at the output stage based on substantial similarity to original works. While the litigation continues, this holding contributed significant clarity to the application of copyright law to AI systems. It properly places the emphasis on analyzing specific outputs rather than automatically deeming all generations from a model trained on copyrighted materials as infringing. Copyright holders retain protection against substantially similar outputs, but training an AI on copyrighted works does not render all its outputs presumptively infringing.

3. Bartz v. Anthropic PBC ⁶¹

In Bartz v. Anthropic PBC,⁶² the Northern District of California considered another putative class action against an AI firm, alleging infringement for unauthorized copies of their works to train their LLM.⁶³ Anthropic motioned for summary judgment on the grounds of fair use.⁶⁴

Rather than treating all copying as a monolithic “AI training” use, the court carefully parsed Anthropic’s conduct into separate analytical buckets:

(1) Copies used specifically to train large language models, which the court found transformative under traditional fair-use analysis;⁶⁵

(2) The format conversion of legitimately purchased print books into digital library copies, which qualified as fair use under a distinct format-shifting rationale;⁶⁶ and

(3) The creation of a general-purpose digital library using pirated materials, which failed fair-use scrutiny entirely.⁶⁷

This tripartite structure proves crucial because each category of copying serves different purposes, involves different acquisition methods, and implicates different copyright concerns. Determining infringing use cases therefore requires independent analysis under the four-factor fair-use framework.⁶⁸

Importantly, the court’s analysis was limited to input-side copying and did not address whether AI-generated outputs might themselves infringe copyright, noting that “[a]uthors do not allege that any LLM output provided to users infringed upon [a]uthors’ works,”⁶⁹ and explicitly stating that “if the outputs were ever to become infringing, [a]uthors could bring such a case.”⁷⁰

The court’s methodical separation of these use cases demonstrates how courts will likely approach future AI copyright cases: not as broad categorical determinations about AI training, but as nuanced examinations of specific copying practices and their individual justifications.

The court’s analysis of the use of copies in AI training represents perhaps the most significant judicial endorsement to date of AI development under fair-use doctrine. The court grounded its reasoning in an analogy to human learning, noting that “[l]ike any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them—but to turn a hard corner and create something different.”⁷¹

This framing proved crucial in distinguishing the case from Thomson Reuters Enter. Centre GmbH v. Ross Intel. Inc., where the court found against fair use.⁷² The key distinction lay in the nature of the AI system: Ross involved training “a competing AI tool for finding court opinions in response to a given legal topic,” which was “not transformative.”⁷³ By contrast, Anthropic’s Claude represents generative AI that creates new content, rather than merely replicating the function of existing databases.

The court’s analysis was bolstered by a critical factual finding: No infringing content ever reached users. The court emphasized that “[a]uthors do not allege that any infringing copy of their works was or would ever be provided to users by the Claude service.”⁷⁴ This absence of direct downstream infringement proved dispositive, as the court noted that filtering software prevented any exact copies or substantial reproductions from reaching the public.

The decision also rejected the authors’ argument that computers should be treated differently from humans in learning contexts. In particular, in rejecting the argument that an LLM learning the creative techniques of source material did not constitute fair use, the court observed that copyright “does not extend to ‘method[s] of operation, concept[s], [or] principle[s],’”⁷⁵ thus analogizing a work’s intangible elements to a “method of operation.”⁷⁶

In a separate but equally important holding, the court found that Anthropic’s conversion of purchased print books to digital formats constituted fair use under a format-shifting theory. Drawing on precedents from Sony Corp. of America v. Universal City Studios,⁷⁷ American Geophysical Union v. Texaco, Inc.,⁷⁸ and the Authors Guild v. Google, Inc. ⁷⁹ litigation, the court concluded that “[s]torage and searchability are not creative properties of the copyrighted work itself but physical properties of the frame around the work or informational properties about the work.”⁸⁰

This analysis proved crucial because it established that legitimate acquisition followed by format conversion differs fundamentally from unauthorized copying. The court emphasized that “every purchased print copy was copied in order to save storage space and to enable searchability as a digital copy. The print original was destroyed. One replaced the other.”⁸¹ Importantly, there was no evidence that digital copies were shared outside the company.⁸²

The decision carefully cabined this holding, noting that while the authors “might have wished to charge Anthropic more for digital than for print copies,”⁸³ the U.S. Constitution’s language “nowhere suggests that [the copyright owner’s] limited exclusive right should include a right to divide markets or a concomitant right to charge different purchasers different prices for the same book . . . .”⁸⁴ This reasoning reinforces the first-sale doctrine, while accommodating technological needs for format conversion.

The court’s treatment of pirated library copies represents the decision’s most significant limitation on AI companies. Despite finding the ultimate training-use transformative, the court firmly rejected the notion that transformative downstream use can cure upstream piracy. As the court colorfully noted, citing Anthropic’s oral arguments:

You can’t just bless yourself by saying I have a research purpose and, therefore, go and take any textbook you want. That would destroy the academic publishing market if that were the case.⁸⁵

The court identified the fundamental flaw in Anthropic’s approach: “[b]uilding a central library of works to be available for any number of further uses”⁸⁶ constituted a separate use from training LLMs. Critically, Anthropic “retained pirated copies even after deciding it would not use them or copies from them for training its LLMs ever again.”⁸⁷ This retention for general purposes, rather than specific transformative use, proved fatal to the portion of the decision judging the fair-use defense on the storage of pirated materials.

The court’s analysis drew important distinctions from cases where intermediate copying was excused. Unlike Perfect 10, Inc. v. Amazon.com, Inc. ⁸⁸ or Kelly v. Arriba Soft Corp.,⁸⁹ where copies were “immediately transformed into a significantly altered form” and deployed directly into transformative uses, Anthropic’s pirated copies were “downloaded and maintained ‘forever’” for a general purpose library.⁹⁰

The decision explicitly rejected the argument that eventual transformative use can retroactively justify initial piracy, emphasizing that “[e]ach use of a work must be analyzed objectively” under Warhol’s framework.⁹¹ This holding establishes that AI companies cannot rely on fair use as a blanket justification to acquire copyrighted materials through unauthorized means.

4. Kadrey v. Meta

On June 25, 2025, Judge Vince Chhabria granted Meta’s cross-motion for partial summary judgment on fair-use grounds in Kadrey v. Meta Platforms, Inc.⁹² Thirteen well-known authors, led by Richard Kadrey, had alleged that Meta’s Llama models were trained on copies of their books scraped from “shadow libraries.”⁹³ The court acknowledged the direct copying but held the training use lawful under 17 U.S.C. § 107, emphasizing that the plaintiffs’ two principal theories of market harm—(i) occasional “regurgitation” of snippets and (ii) the loss of a putative licensing market for training data—were “clear losers.”⁹⁴

Yet the opinion pointedly identified a third, undeveloped theory that might change the calculus: If AI outputs “flood the market with similar works,” the resulting market dilution could undercut authors’ ability to sell their original books.⁹⁵ The judge characterized that argument as “potentially winning,” but found the record devoid of evidence tying the Meta Llama model’s current or foreseeable outputs to any substitution effect for the plaintiffs’ works.⁹⁶ In consequence, the ruling is a victory for broad-purpose training on the facts presented, not a blanket endorsement of unlicensed training writ large.

The decision therefore extends the line begun in Bartz v. Anthropic by reaffirming that large scale ingestion of text can be transformative and thus fair use. What is new—and worrisome—is the court’s willingness to frame market harm around speculative output substitution rather than the copying of inputs itself. If accepted by courts, that conflation opens a path for future plaintiffs to survive early motions even without locating infringing passages: They need only posit that model-generated books might depress demand for their own. If embraced, such a dilution theory would shift the fair-use inquiry toward policing downstream creativity, a result at odds with copyright’s traditional focus on actual substitution.

For present purposes, Kadrey underscores two messages that animate the next Part of this paper. First, courts remain skeptical of licensing-market and snippet-regurgitation claims, signaling a judicial appetite for doctrinal clarity over rhetoric. Second, and more importantly, the “market dilution” dicta reveal how quickly copyright analysis can become entangled with broader competition-policy concerns when judges conflate learning inputs with generative outputs. Untangling those strands—protecting training while addressing genuinely infringing outputs—will be essential to preserving both innovation incentives and the integrity of authors’ markets.

C. Summary

The economic analysis of copyright’s foundational principles, coupled with emerging judicial interpretations in cases involving artificial intelligence, is beginning to clarify the legal landscape for AI systems trained on copyrighted materials. This evolving jurisprudence reflects the courts’ efforts to balance the tension inherent in copyright’s hydraulic system—protecting creators’ economic incentives while facilitating technological innovation that serves distinct markets. Through decisions such as Thomson Reuters v. Ross Intelligence, Tremblay v. OpenAI, and Bartz v. Anthropic PBC, courts are contributing toward a nuanced framework that distinguishes between different types of AI systems based primarily on their economic impact on copyright holders’ markets.

The developing jurisprudential framework appears to establish two significant principles. First, some courts have rejected the proposition that all outputs from an AI system trained on copyrighted materials are per se infringing; rather, plaintiffs must demonstrate substantial similarity between specific outputs and protected works, consistent with traditional copyright infringement analysis. Second, whether the resulting AI system directly competes with the market for the original copyrighted works may substantially influence the fair-use inquiry at the training stage. This analytical approach potentially creates a viable path for general-purpose LLMs to qualify for fair-use protection where their outputs serve distinctly different market functions than the original works used in training.

If the Anthropic decision gains traction in other courts, it would seem to draw a clear roadmap for AI companies navigating copyright law. First, strategic sourcing beats piracy: Companies should acquire training materials through legitimate channels—purchase, licensing, or authorized access. The court’s finding that format shifting of legitimately acquired materials constitutes fair use further provides a viable path for companies needing digital copies for computational purposes.

Second, output liability emerges as the next frontier. Because no infringing content reached users in this case, future litigation will likely focus on whether AI-generated outputs themselves infringe copyright. The court explicitly noted that “if the outputs were ever to become infringing, [a]uthors could bring such a case.”⁹⁷

Third, the decision signals potential legislative solutions. It very well may be the case that generative AI will create hurdles for creators seeking to monetize their work. But it also could be that new markets emerge to allow artists to be remunerated on, for example, some sort of new property right in “name, image, and likeness.”⁹⁸ That would, however, likely require some sort of legislative enactment, perhaps by amending the Lanham Act⁹⁹ with an extension of the concept of trademark.

Notwithstanding these initial judicial determinations, it would be premature to conclude that all such uses of copyrighted materials will ultimately receive official sanction. As the subsequent Part explores, competition authorities may exercise their regulatory discretion to characterize even technically non-infringing or excused infringing uses of copyrighted materials as constituting unfair methods of competition. This introduces an additional layer of legal complexity that transcends traditional copyright analysis.

II. How Can Fair Use Be Unfair?

Assuming judicial recognition of a fair-use exception for the incorporation of copyrighted materials as training inputs for AI systems, a complex regulatory question emerges regarding the FTC’s potential jurisdiction over such otherwise permissible activities. This jurisdictional inquiry necessitates examination of Section 5 of the Federal Trade Commission Act (“FTC Act”), which confers broad authority to proscribe “unfair methods of competition” and “unfair or deceptive acts or practices” in commerce.¹⁰⁰ The subsequent analysis first delineates the Commission’s statutory authority under Section 5 jurisprudence and attendant agency interpretations, followed by an examination of the Commission’s recent commentary to the United States Copyright Office regarding the potential application of this authority to artificial intelligence training methodologies that utilize copyrighted materials.

A. Section 5 Authority

Section 5 of the FTC Act states that “[u]nfair methods of competition in or affecting commerce, and unfair or deceptive acts or practices in or affecting commerce, are hereby declared unlawful.”¹⁰¹ The Commission may not declare an act or practice unlawful “on the grounds that such act or practice is unfair unless the act or practice causes or is likely to cause substantial injury to consumers which is not reasonably avoidable by consumers themselves and not outweighed by countervailing benefits to consumers or to competition.”¹⁰²

Case law and FTC guidance have further developed these high-level principles. Below, unfair methods of competition (“UMC”) and unfair or deceptive acts or practices (“UDAP”) will both be introduced in order to better understand the FTC’s argument on how the use of copyrighted materials as AI inputs could violate the FTC’s Section 5 authority.

1. UMC Unfairness

The jurisprudential framework defining “unfair methods of competition” has evolved primarily through case law interpretation and administrative guidance.¹⁰³ Understanding this evolution is essential to evaluating the FTC’s current approach to AI training and copyright.

a. Historical Development and Judicial Foundation¹⁰⁴

From its inception, Congress intended Section 5 to provide the FTC with greater flexibility than that afforded by the Sherman and Clayton Acts, allowing it to address anticompetitive practices that might not fit neatly within the strict definitions of existing antitrust statutes. This design aimed to prevent the new antitrust framework from becoming static and unable to adapt to novel forms of anticompetitive conduct.

A landmark Supreme Court decision, FTC v. Sperry & Hutchinson Co.,¹⁰⁵ significantly affirmed the FTC’s expansive authority under the UMC prong.¹⁰⁶ The Court held that the FTC could proscribe practices as “unfair” even if those practices did not violate the letter or the spirit of the antitrust laws.¹⁰⁷ The Sperry & Hutchinson ruling empowered the Commission to consider broader “public values beyond simply those enshrined in the letter . . . of the antitrust laws” when determining whether a method of competition was unfair.¹⁰⁸ This decision is foundational to understanding the FTC’s capacity to act against conduct that might otherwise be permissible under traditional antitrust analysis.

More recently, the Commission’s 2022 “Policy Statement Regarding the Scope of Unfair Methods of Competition Under Section 5 of the Federal Trade Commission Act”¹⁰⁹ signaled a renewed commitment to a broad interpretation of its UMC authority. This statement explicitly asserts that Section 5’s reach extends beyond the Sherman and Clayton Acts, targeting conduct that may violate “‘the spirit’ of the antitrust laws” or is in its “incipiency.”¹¹⁰ The 2022 Policy Statement emphasizes that the FTC will focus on conduct that “goes beyond competition on the merits”¹¹¹ and “tend[s] to negatively affect competitive conditions.”¹¹² Notably, it suggests that demonstrating actual anticompetitive harm or market power, often a rigorous requirement in traditional antitrust litigation, may not be necessary in all Section 5 UMC cases.¹¹³ This approach effectively lowers the evidentiary burden for the FTC compared to standard antitrust enforcement, enabling challenges to conduct that might survive scrutiny under the Sherman or Clayton Acts.

This policy reflects an intent to use Section 5 as a dynamic tool, potentially to counteract judicial interpretations that some argue have narrowed the scope of traditional antitrust statutes, such as the Supreme Court’s decision in Verizon Communications Inc. v. Law Offices of Curtis V. Trinko, LLP,¹¹⁴ which constrained the essential facilities doctrine.¹¹⁵

b. Analytical Framework

The test for identifying whether conduct is an unfair method of competition includes first a determination that the challenged conduct is a method of competition, and second that it is unfair.¹¹⁶

To constitute a method of competition under the Commission’s analytical framework, the challenged conduct must satisfy two threshold criteria. First, it must represent affirmative action “undertaken by an actor in the marketplace,”¹¹⁷ rather than merely reflecting structural market conditions such as entry barriers or industry concentration levels. Second, the conduct must necessarily implicate competitive dynamics. The Commission has indicated that certain behaviors outside traditional antitrust boundaries—including the misuse of regulatory processes or violations of generally applicable laws, may satisfy this requirement when they affect competition.¹¹⁸

For instance, the Commission cites Walker Process Equipment, Inc. v. Food Machinery & Chemical Corp. ¹¹⁹ for the proposition that defrauding the patent office could be a basis for an antitrust violation.¹²⁰ The Court found such abuse of regulatory process could be a basis for a Sherman Act Section 2 claim, but it was not a per se violation.¹²¹ The Commission also cites In re American Cyanamid Co.,¹²² as an example of finding misleading statements and withholding of information leading to a patent could be found to be an unfair method of competition.¹²³

The concept of “unfairness” within the Commission’s Section 5 jurisprudence encompasses conduct that “goes beyond competition on the merits.”¹²⁴ The Commission employs two principal criteria to evaluate whether conduct constitutes impermissible, non-meritorious competition: first, whether the conduct is “coercive, exploitative, collusive, abusive, deceptive, predatory, or involve[s] the use of economic power of a similar nature,” and second, whether it “tend[s] . . . to negatively affect competition conditions.”¹²⁵

These evaluative criteria operate on a sliding scale within the Commission’s analytical framework, such that compelling evidence of one criterion may diminish the quantum of evidence required for the other.¹²⁶ Significantly, the Commission has clarified that actual anticompetitive harm need not be demonstrated; rather, a tendency to produce negative competitive effects suffices to establish a violation.¹²⁷

When presented with a prima facie case of an unfair method of competition, the Commission will consider potential justifications, albeit within narrowly circumscribed parameters. The Commission categorically rejects mere pecuniary benefits accruing to the respondent as sufficient justification.¹²⁸ Any proffered justification must be legally cognizable, non-pretextual, and narrowly tailored to minimize competitive harm.¹²⁹ Furthermore, the Commission requires that the asserted benefits manifest in the same market where the competitive harm occurs.¹³⁰ Even when these stringent requirements are satisfied, the claimed benefits must outweigh the competitive harm to constitute a valid defense.¹³¹

2. UDAP Unfairness

The Commission’s authority to designate consumer practices as “unfair” or “deceptive” under Section 5 underwent significant legislative curtailment following congressional concerns regarding perceived administrative overreach in the late 1970s.¹³² In response to congressional scrutiny, the Commission promulgated its Policy Statement on Unfairness, which articulated interpretive principles aligned with the statutory parameters subsequently codified at 15 U.S.C. § 45(n).¹³³

Under this framework, unjustified consumer injury constitutes the cardinal consideration in consumer unfairness analysis. To satisfy the statutory threshold, such injury must satisfy three conjunctive elements: It must be (1) substantial, (2) not outweighed by countervailing consumer or competitive benefits, and (3) not reasonably avoidable through consumer action.¹³⁴

The substantiality requirement mandates that actionable harm transcend mere triviality or speculative injury. While economic detriment typically satisfies this criterion, substantial health and safety risks may likewise establish unfairness, whereas subjective or emotional harms generally fall outside the Commission’s enforcement purview.¹³⁵

The Commission’s balancing inquiry acknowledges that “[m]ost business practices entail a mixture of economic and other costs and benefits,”¹³⁶ necessitating careful evaluation of offsetting market advantages. Under this calculus, practices are deemed unfair only when “injurious in [their] net effects,”¹³⁷ requiring the Commission to weigh consumer injury against potential market efficiencies.

Finally, the reasonable avoidability criterion reflects the Commission’s deference to market self-correction through informed consumer choice. As articulated in the Policy Statement, “[n]ormally we expect the marketplace to be self-correcting, and we rely on consumer choice—the ability of individual consumers to make their own private purchasing decision without regulatory intervention—to govern the market.”¹³⁸ The Commission thus directs its enforcement authority not toward “second-guess[ing] the wisdom of particular consumer decisions, but rather to halt some form of seller behavior that unreasonably creates or takes advantage of an obstacle to the free exercise of consumer decision making.”¹³⁹ Paradigmatic examples include information asymmetries that prevent meaningful comparison shopping, coercive service contract tactics, and fraudulent health claims—practices that undermine the consumer’s capacity for autonomous market participation.¹⁴⁰

3. The Convergence Problem: UMC and UDAP Unfairness Standards

The distinction and potential for overlap between UMC “unfairness” and UDAP “unfairness” can create significant regulatory ambiguity with particular relevance to AI copyright issues. While UDAP unfairness is governed by the codified three-part test in § 45(n) focusing on consumer injury, UMC unfairness, particularly as articulated in the 2022 Policy Statement, appears more open-ended. It relies on “indicia of unfairness” (such as conduct being coercive, exploitative, or abusive) and a “tendency to negatively affect competitive conditions.”¹⁴¹

This divergence implies that conduct not meeting the “substantial consumer injury” threshold for UDAP unfairness might still be challenged as an “unfair method of competition” if it is deemed to adversely affect the broader competitive landscape, even if direct, quantifiable consumer harm is less apparent or does not meet the § 45(n) criteria. This offers an alternative avenue for the FTC to intervene against conduct that might have been cleared under one specific aspect of its authority or by a court applying a similar consumer injury standard.

This regulatory flexibility becomes particularly problematic in the AI training context. Consider a scenario where an AI developer’s use of copyrighted materials under fair use cannot satisfy the UDAP unfairness standard because consumers suffer no substantial injury—indeed, consumers may benefit from improved AI capabilities. Under the more expansive UMC framework, however, the same conduct could be characterized as “exploitative” of rightsholders or as having a “tendency to negatively affect competitive conditions” in content markets, even without demonstrable consumer harm.

This dual-track approach effectively allows the FTC to forum shop within its own authority, selecting the standard most likely to yield enforcement success rather than applying consistent principles. Such regulatory arbitrage undermines legal predictability and threatens to circumvent the careful limitations Congress imposed on UDAP authority following the Commission’s overreach in the 1970s.

B. FTC Comment to Copyright Office

With the foregoing analytical framework established, it becomes possible to examine the Commission’s potential application of unfairness principles to copyrighted materials used in artificial intelligence training. In 2023, the United States Copyright Office initiated an inquiry regarding artificial intelligence and copyright law,¹⁴² to which the Commission submitted a formal comment articulating its relevant enforcement interests and identifying potential competition and consumer protection concerns implicated by generative AI technologies.¹⁴³

In its submission, the Commission first catalogued its enforcement activities relating to artificial intelligence business practices before addressing competitive concerns. The Commission asserted that “rapid development and deployment of AI also poses potential risks to competition,” with particular emphasis on market concentration among “dominant” technology firms possessing critical AI inputs including technological infrastructure, computational resources, and training data access.¹⁴⁴ Based on these market dynamics, the Commission maintained that its unfair methods of competition authority constituted an essential regulatory tool.

With specific reference to copyrighted works as training inputs, the Commission stated that “under certain circumstances, the use of pirated or misuse of copyrighted materials could be an unfair practice or unfair method of competition under Section 5 of the FTC Act.”¹⁴⁵ Certain applications of this principle align logically with established unfairness jurisprudence. For instance, the Commission noted that copyright-violative conduct—such as unauthorized training on protected expression or commercializing outputs that mimic creators’ distinctive attributes—may constitute unfair methods of competition or unfair practices, particularly when such conduct “deceives consumers, exploits a creator’s reputation or diminishes the value of her existing or future works, reveals private information, or otherwise causes substantial injury to consumers.”¹⁴⁶ Such reasoning comports with traditional unfairness analysis where copyright infringement (absent fair use) serves as the predicate violation.

However, the Commission’s position extends considerably further, asserting that “conduct that may be consistent with the copyright laws nevertheless may violate Section 5.”¹⁴⁷ This expansive interpretation potentially encompasses lawful uses of copyrighted material—including those protected under the fair use doctrine—within the ambit of unfairness regulation.

The Commission’s justification for this jurisdictional expansion appears questionable upon scrutiny. The Commission contends that “[m]any large technology firms possess vast financial resources that enable them to indemnify the users of their generative AI tools or obtain exclusive licenses to copyrighted (or otherwise proprietary) training data, potentially further entrenching the market power of these dominant firms.”¹⁴⁸ This argument’s persuasiveness depends entirely on the premise that fair use would not apply to smaller AI companies seeking to utilize identical data. If fair use doctrine permits such utilization—as recent jurisprudential developments suggest it might for general-purpose systems—then both established technology firms and emerging competitors would enjoy equivalent legal access to these inputs.

Moreover, the Commission’s market power reasoning with respect to licensing deals could paradoxically support imposing affirmative dealing obligations on major copyright holders (at least with respect to their relations with small AI developers), an outcome that would seem to contradict established antitrust principles. For example, Thomson Reuters, as one of two dominant legal research providers through its Westlaw service, might under such logic be compelled to license its proprietary headnotes to market entrants like Ross Intelligence to facilitate competitive AI development. This reasoning would effectively transform copyright—a legal regime designed to grant exclusive rights—into a vehicle for mandatory licensing, at least for entities with substantial market share.

The tension becomes particularly apparent when considering the essential facilities doctrine in antitrust jurisprudence. While this doctrine has been significantly narrowed by the Supreme Court in cases such as Verizon Communications Inc. v. Law Offices of Curtis V. Trinko, LLP, the Commission’s approach could be an attempt to resurrect a more expansive interpretation specifically for copyright holders.¹⁴⁹ If a dominant technology firm’s exclusive licensing of copyrighted materials constitutes an unfair method of competition due to potential market foreclosure, then symmetrically, a dominant copyright holder’s refusal to license to potential competitors would logically present similar competitive concerns. This interpretive approach would effectively subordinate copyright law’s exclusivity principles to competition policy considerations whenever a copyright holder achieves significant market position.

Furthermore, such reasoning introduces significant doctrinal inconsistency. The Copyright Office and courts have carefully delineated when fair use applies to innovative technologies, calibrating intellectual property protections to balance creator incentives with technological progress. Allowing the Commission to determine that legally permissible uses under copyright law—including fair use—nonetheless constitute unfair methods of competition would create competing legal standards that undermine this carefully crafted equilibrium. Copyright holders and AI developers would face contradictory legal obligations, where conduct explicitly permitted under one legal regime could simultaneously be prohibited under another.

In sum, the Commission has advanced the proposition that even fair use of copyrighted materials in AI training may constitute an unfair method of competition or unfair act or practice. The subsequent analysis will demonstrate that this regulatory approach would potentially undermine innovation and competition—the foundational values that animate copyright, antitrust, and consumer protection jurisprudence.

III. The FTC Should Not Discourage Innovation by Undermining Copyright and Fair Use

The optimal approach would harmonize copyright and competition law by preventing unsuccessful copyright infringement claims from serving as the foundation for unfairness claims. Courts should preserve copyright law’s carefully calibrated balance by limiting independent unfairness claims.

This Part analyzes how the FTC’s Copyright Comment makes incorrect assumptions about the market for AI training, which colors its analysis. Then, it makes a case for why the law of unfairness should be limited to situations where there is a clear underlying violation of copyright law, unexcused by fair use. Finally, it considers how courts have handled state-level unfair competition claims as an example of reading copyright law and competition law consistently.

A. FTC’s Comment to the Copyright Office Makes Incorrect Assumptions About the AI Training Market

The FTC’s proposed approach misunderstands the dynamism in AI technology markets and risks overweighting rightsholder interests in AI training contexts. The Commission’s perspective reflects several problematic assumptions about both the technical functioning of AI and market realities.

First, the approach effectively grants rightsholders protections exceeding traditional copyright boundaries. No legal consensus exists on how AI systems conceptually engage with copyrighted works. While arguments exist that AI models “memorize” content,¹⁵⁰ research demonstrates that careful data curation can significantly reduce such memorization.¹⁵¹ Moreover, AI systems only make “true” copies during training; they transform works into tokens and abstract “weight” adjustments rather than storing original text sequences.¹⁵² The resulting model contains billions of parameters that form “a vast sea of numbers, with no direct correspondence to the original text.”¹⁵³ This tokenization process enables models to learn language patterns without reproducing creative expression in their permanent files.¹⁵⁴ Thus, fair use remains a live controversy requiring careful consideration before agencies impose competition-based restrictions.¹⁵⁵

Second, the FTC’s approach would itself create prohibitive transaction costs. Much internet content used in AI training lacks centralized management, making comprehensive rightsholder identification virtually impossible. Even if possible, the transaction costs would likely lead to less-capable AI models due to input constraints, or force reliance on synthetic data with potentially negative effects on model quality.¹⁵⁶

Third, value assignment at the input stage presents insurmountable challenges. Copyright Office analysis confirms that with foundation models ingesting billions of works, individual works’ influence becomes so diluted that even minimal licensing transaction costs would exceed that work’s share of model utility.¹⁵⁷ Even highly valuable literary collections represent mere drops in the vast training corpus. Traditional valuation methods become impractical—multiplying per-work fees by millions or billions yields astronomically high costs disconnected from actual impact on the model.

Furthermore, monetization in generative AI occurs primarily at the output stage when models produce valuable content, not during data ingestion. Without methods to connect specific generated content to particular training examples, input valuation remains highly speculative and likely de minimis for most individual works.

Finally, restrictive approaches risk creating AI systems that disproportionately reflect perspectives from entities with market power or established licensing frameworks. This undermines representation of independent creators and non-commercial knowledge sources, potentially producing AI systems biased toward commercially dominant viewpoints rather than reflecting the full spectrum of human knowledge and creative expression.

B. Unfairness Without Violation? Why the FTC Should Not Treat Alleged or Excused Legal Conduct as Unfair Competition

The FTC’s invocation of Section 5 unfairness authority in contexts involving other legal regimes bears a striking resemblance to the tort doctrine of negligence per se. Under that doctrine, a violation of a statute may serve as evidence of breach in a negligence case, but courts impose strict limits on when such violations may be used to establish tort liability.¹⁵⁸ Crucially, negligence per se applies only where there has been an actual violation of a statute, and that violation is unexcused.¹⁵⁹ If the defendant raises and successfully invokes a statutory defense—such as impossibility, necessity, or compliance under emergency conditions—then the mere fact of technical noncompliance is insufficient to establish liability.¹⁶⁰ Similarly, if the statute was not designed to prevent the type of harm at issue, courts will not apply the negligence per se doctrine.¹⁶¹

By contrast, the FTC’s theory of unfairness lacks these core doctrinal constraints. The Commission’s 2023 comment to the U.S. Copyright Office asserts that even uses of copyrighted material “consistent with copyright laws” may still constitute unfair methods of competition under Section 5.¹⁶² This approach treats legal compliance not as a defense but as a separate, potentially irrelevant question—allowing the Commission to recharacterize excused conduct (such as a use found to be fair under the Copyright Act) as independently “unfair.”

The FTC’s treatment of legal violations, more notably, non-violations, under its unfairness authority exhibits a profound internal inconsistency. In some regulatory contexts, the FTC treats clear statutory violations as proxies for anticompetitive harm. For instance, in labor and employment cases, the Commission has suggested that violations of workplace safety or labor laws may be evidence of an exercise of market power.¹⁶³ Yet in others, such as copyright and fair use, the FTC implies that conduct consistent with the governing legal framework may nonetheless be deemed “unfair.”¹⁶⁴ This inconsistency reveals a troubling elasticity in the agency’s approach: where it suits the enforcement agenda, legal compliance may be disregarded or recharacterized as evidence of harm.

Consider the FTC’s increasing interest in labor market practices. In recent policy discussions, the agency has emphasized the relationship between employer conduct—such as wage suppression, no-poach agreements, or safety violations—and potential labor market concentration.¹⁶⁵ A violation of Occupational Safety and Health Administration regulations, for example, is often cited not only as a harm in itself but as a symptom of monopsony power.¹⁶⁶

Yet in stark contrast, when addressing uses of copyrighted materials in AI training, the FTC takes precisely the opposite view. The Commission’s 2023 comment to the U.S. Copyright Office asserted that “conduct that may be consistent with copyright laws nevertheless may violate Section 5.”¹⁶⁷ In other words, even where courts have determined that no copyright violation has occurred the FTC reserves the right to label that same conduct “unfair.” This reveals a reverse asymmetry: Here, non-violation still equals market abuse.

This doctrinal inconsistency erodes the distinction between unlawful conduct and conduct that the law affirmatively permits. Courts, by contrast, have consistently treated legal compliance and affirmative defenses as relevant, and often dispositive, in determining liability. In copyright law, the fair use doctrine is not a mere procedural hurdle; it is a substantive determination that the conduct in question does not give rise to liability. As the Supreme Court has held, fair use is not a lesser form of infringement but a doctrine that “permits courts to avoid rigid application of the copyright statute when, on occasion, it would stifle the very creativity which that law is designed to foster.”¹⁶⁸

Copyright law (including its fair use doctrine and related limitations) is a fully articulated legal regime, carefully constructed by Congress and interpreted by courts to balance exclusive rights with innovation and expressive freedom. By asserting that conduct “consistent with copyright law” may still constitute an unfair method of competition under Section 5, the FTC risks imposing liability for actions that the Copyright Act explicitly permits. This threatens to unravel the statutory equilibrium, subjecting developers and creators to duplicative or conflicting standards.

Allowing the FTC to treat alleged or excused violations of other laws as per se “unfair” under Section 5 creates a regulatory framework unmoored from statute, judicial precedent, or objective standards. Such an approach threatens to convert Section 5 into a floating liability regime, where enforcement decisions hinge not on defined legal boundaries but on agency discretion. In this framework, legal compliance—including compliance confirmed by judicial findings—offers no safe harbor, while allegations alone may suffice to trigger regulatory sanction.

This discretionary expansion of unfairness authority poses serious risks for legal predictability and innovation. By failing to cabin its enforcement actions to conduct that is clearly unlawful and unexcused, the FTC fosters an environment of retrospective enforcement—where conduct deemed permissible at the time may be recharacterized as unfair after the fact. This uncertainty chills risk-taking and innovation, particularly in areas like AI development where legal doctrines such as fair use remain actively contested.

To safeguard the integrity of both legal compliance and competition enforcement, the FTC should constrain its unfairness authority to situations involving clear statutory violations that are not excused by affirmative defenses or legal exceptions. This would ensure that Section 5 complements, rather than contradicts, the substantive laws it touches—and restores much-needed stability to the Commission’s enforcement approach.

C. The FTC’s Pattern of Treating Lawful Conduct as Unfair: Precedents for Copyright Overreach

The FTC’s assertion that conduct “consistent with copyright laws nevertheless may violate Section 5”¹⁶⁹ is not an isolated position but reflects a broader pattern of regulatory expansion. Examining the Commission’s enforcement history reveals multiple instances where it has challenged conduct that was arguably permissible under other legal frameworks or had been cleared by specialized authorities. These precedents illuminate the practical risks of allowing the FTC to second-guess copyright law’s careful balance through unfairness claims.

1. Circumventing Sherman Act Requirements: Invitations to Collude

The FTC’s treatment of inchoate anticompetitive conduct demonstrates its willingness to expand beyond traditional legal boundaries. In Drug Testing Compliance Group, LLC, the Commission challenged a mere invitation to collude as an unfair method of competition, even though no agreement was formed.¹⁷⁰ Under established Sherman Act jurisprudence, such conduct would not constitute a violation because it lacks the requisite “contract, combination . . . or conspiracy” element.¹⁷¹ Unlike under the 2015 UMC Policy Statement,¹⁷² the current version allows for much more expansive enforcement actions against “incipient” threats to competition.¹⁷³

This approach effectively allows the FTC to prosecute conduct that the Department of Justice could not pursue under traditional antitrust law. While the Commission characterizes this as addressing “incipient” violations, it fundamentally alters the legal landscape by criminalizing preparatory conduct that Congress chose not to prohibit directly. The parallel to copyright is clear: just as the FTC bypasses Sherman Act requirements by invoking Section 5, it seeks to bypass fair-use determinations by recharacterizing potentially lawful copying as unfair competition.

2. Meta-Regulation of Specialized Agencies: Orange Book Patent Listings

The Commission’s challenges to pharmaceutical patent listings in the FDA’s Orange Book reveal another dimension of regulatory overreach. These patents are granted by the USPTO—which presumes their validity—and listed through FDA procedures. Yet the FTC routinely challenges the listing of certain medical devices under its unfair methods of competition authority.¹⁷⁴

This creates a precedent where the FTC positions itself as a meta-regulator, second-guessing determinations made by agencies with specialized expertise. The USPTO evaluates patent validity, and the FDA administers the listing process according to established regulations. By recharacterizing the use of these regulatory entitlements as anticompetitive, the FTC undermines the coherence of the federal regulatory framework.

The copyright parallel is direct: Just as the FTC challenges patent holders’ use of USPTO-granted rights and FDA procedures, it seeks to challenge AI developers’ use of copyright materials that courts might deem protected by fair use. In both contexts, the Commission treats compliance with an established legal regime as irrelevant to its broader unfairness authority.

3. Expanding Beyond Specific Statutory Frameworks: Data Privacy Practices

Perhaps most relevant to the copyright context is the FTC’s approach to data privacy. The Commission has explicitly asserted that data practices may be “unfair” under Section 5 even when they comply with specific privacy statutes like COPPA, HIPAA, or state laws like the CCPA.¹⁷⁵ This principle directly parallels the Commission’s position on copyright: that conduct “consistent with copyright laws nevertheless may violate Section 5.”

This approach transforms the FTC into a roving commission empowered to find “unfairness” wherever specific statutory schemes permit conduct the Commission disfavors. Rather than respecting the policy choices embedded in specialized legislation, the Commission treats its general unfairness authority as superior to Congress’s specific enactments.

4. The Outer Boundary of Section 5 Authority

The Commission’s most ambitious extensions of stand-alone Section 5 authority have arisen in the so-called “patent ambush” context. Dell,¹⁷⁶ Rambus,¹⁷⁷ N-Data,¹⁷⁸ and Unocal ¹⁷⁹ show the Commission reaching past established patent and contract doctrines when it believed deception inside a collective decision-making body created ex-post market power. They represent the high-water mark of Section 5’s expansion—bold enough to secure consent orders in Dell, N-Data, and Unocal, yet fragile enough to collapse on appeal in Rambus.

All four episodes shared three elements. First, the challenged firm took part in a collective-choice process whose legitimacy depends on candor: an SSO in Dell,¹⁸⁰ Rambus,¹⁸¹ and N-Data and a public rule-making in Unocal.¹⁸² Second, the Commission alleged concrete deception capable of inducing reliance. Dell and Unocal flatly denied owning relevant patents;¹⁸³ Rambus and N-Data withheld or repudiated licensing commitments while watching rivals devote investments into the chosen technology.¹⁸⁴ Third, the deception was said to confer ex post market power by locking an entire industry into the patented technology once switching became impracticable.¹⁸⁵ Only when those three predicates lined up did the Commission feel confident invoking its stand-alone unfair-competition authority.

Even then, the outcomes were mixed. Dell capitulated quickly,¹⁸⁶ and N-Data settled after a divided vote that drew two dissents questioning the theory’s novelty.¹⁸⁷ Rambus, by contrast, is the cautionary limit case: the D.C. Circuit vacated the Commission’s order because it could not see proof that the relevant standards body would have adopted a different standard “but for” the nondisclosure.¹⁸⁸ Unocal settled only after a merger placed negotiating pressure on the company.¹⁸⁹ Taken together, the quartet shows that Section 5 can fill gaps left by patent, contract, or fraud doctrines—but only where intentional deception distorts a decision-making forum and plausibly produces durable market power.

That historical pattern is instructive precisely because it is absent from the Commission’s current rhetoric on generative-AI training. In the patent-ambush cases, the Commission did not claim that exercising a valid patent was inherently “unfair.” The gravamen was the deceptive acquisition of that patent-enhanced position—an “extra element” beyond the lawful assertion of a statutory right. By contrast, the FTC’s 2023 comment to the Copyright Office asserts that an AI developer’s use of copyrighted materials may violate Section 5 even when a court holds the copying to be fair use. In other words, the Commission now proposes to condemn conduct because it complies with the governing intellectual-property regime, not because it secretly subverts it. That move would stretch Section 5 well past the boundary charted by Dell, Rambus, N-Data, and Unocal.

The distinction matters for both doctrine and policy. Section 5 interventions premised on deception do not threaten the predictability of patent (or copyright) law because they target conduct the specialist regime never intended to bless. Declaring “unfair” a use that copyright affirmatively permits, by contrast, would create a direct conflict between two federal statutes, depriving innovators of any stable safe harbor. Far from policing a gap, the Commission would manufacture one.

5. Implications for Copyright and AI Innovation

These precedents reveal a consistent pattern: The FTC regularly treats compliance with other legal frameworks as potentially irrelevant to its unfairness analysis. Whether the conduct involves Sherman Act requirements, USPTO/FDA procedures, or specific privacy statutes, the Commission maintains that its Section 5 authority operates independently.

Applied to copyright, this pattern suggests the FTC would likely pursue unfairness claims against AI training practices regardless of fair-use determinations. Just as it challenges disfavored patent listings despite USPTO approval, or data practices despite HIPAA compliance, the Commission would likely characterize fair use of copyrighted materials as “unfair” whenever it serves broader enforcement objectives.

This approach fundamentally undermines legal predictability and the careful policy balances Congress has struck. Copyright law, like patent and privacy law, reflects specific Congressional choices about how to balance competing interests. Allowing the FTC to override these determinations through open-ended unfairness claims effectively nullifies the specialized frameworks Congress has created.

D. State Law Unfair Competition Claims and Copyright Fair Use: Recent Jurisprudence

While the FTC operates under federal statutory authority and thus is not constrained by copyright preemption principles that limit state unfair competition claims,¹⁹⁰ the judicial reasoning in the preceding state cases offers valuable guidance. The same logic that applies to preemption of copyright claims recast as duplicative state competition claims applies in equal force to the use of open-ended federal unfairness authority to end-run the Copyright Act. The preemption of claims that recast copyright causes of action as non-copyright causes of action would prevent regulatory authorities from obtaining, in effect, duplicative opportunities to penalize identical conduct under different legal theories. Moreover, this analytical framework provides essential legal certainty for artificial intelligence developers by ensuring that conduct deemed permissible under copyright doctrine cannot be subsequently recharacterized as unfair competition without additional elements of wrongdoing. Maintaining this doctrinal consistency across legal regimes serves the fundamental interests of both regulatory predictability and technological innovation.

Recent litigation has examined whether state law unfair competition claims can succeed in cases where copyright fair use might apply to AI training. Courts have generally been skeptical of such claims, finding them preempted by federal copyright law absent additional elements beyond mere copying, and for good reason. The principle of copyright preemption stems from the Copyright Act’s express provision that state law claims are preempted when they concern works “within the subject matter of copyright” and assert “legal or equitable rights that are equivalent to any of the exclusive rights within the general scope of copyright.”¹⁹¹ Courts have consistently interpreted this provision to bar state unfair competition claims that merely restate copyright infringement allegations without an “extra element” that qualitatively differentiates the claim from copyright protection.¹⁹² This doctrine preserves the federal scheme’s integrity by ensuring that copyright’s carefully balanced rights and limitations cannot be circumvented through alternative state law theories addressing identical conduct.

As noted above, in Thomson Reuters v. Ross,¹⁹³ that court rejected Ross’s fair-use defense for copying Westlaw’s copyrighted headnotes to train an AI legal research tool. In the District of Delaware litigation, Thomson Reuters had characterized Ross’s actions as both copyright infringement and unfair competition, the court held the unfair competition claim preempted due to the “gravamen of [the] claim [being] the same as . . . [the] copyright claim.”¹⁹⁴ Notably, the FTC has previously cited this case as an example where an AI developer’s conduct might constitute an “unfair method of competition” regardless of copyright liability determinations.¹⁹⁵

In Tremblay v. OpenAI, Inc., plaintiffs alleged both copyright infringement and unfair competition related to the training of GPT models on their books without permission.¹⁹⁶ The court dismissed the unfair competition claim in July 2024 while allowing the copyright claim to proceed.¹⁹⁷ The judge determined that the state law unfair competition allegations lacked the necessary “extra element” beyond unauthorized copying (such as deception or passing off) required to escape federal copyright preemption.¹⁹⁸

Similarly, in the GitHub Copilot litigation concerning AI training on open-source code, the court dismissed unfair competition claims brought under California law to the extent they were based on other failed claims.¹⁹⁹ Without an independent wrong beyond reproduction of code snippets, the unfair competition theory could not stand.²⁰⁰

These cases illustrate a consistent principle: State law unfair competition claims involving AI training data cannot succeed when they merely restate copyright infringement allegations. Courts have required an “extra element” such as misrepresentation, passing off, or consumer deception to overcome copyright preemption. When fair use applies—or even potentially applies—to the underlying copying, courts have rightfully rejected attempts to reframe the same conduct as unfair competition without additional wrongful elements beyond the copying itself.

Conclusion

The FTC’s proposed approach risks undermining innovation without adequately addressing the core challenges at the intersection of copyright and AI. By potentially treating even fair use of copyrighted materials as “unfair competition,” the Commission would create regulatory uncertainty that could significantly hamper AI development.

A more balanced approach would recognize what might be called the “hydraulics of copyright”—when pressure is applied in one area of copyright law, it creates a counterbalancing effect elsewhere. Rather than restricting AI development at the input stage through potentially overlapping and contradictory regulatory regimes, policymakers should focus on solutions at the output stage. This would recognize both the transformative nature of AI training and the legitimate interests of creators in protecting their works from direct competition.

When AI models generate outputs that substantially resemble protected works, copyright law already provides mechanisms to challenge such outputs. Strengthening these protections while permitting fair use at the training stage would maintain the essential balance between protecting creators and encouraging innovation.

The FTC should limit its unfairness authority to cases where courts have found copyright infringement without fair-use protection. This approach would align copyright law with competition and consumer protection principles while maintaining the balance necessary for innovation. Ideally, Congress should also clarify that the preemption principles embodied in 17 U.S.C. § 301(a) for state competition claims should also guide federal agencies’ application of unfairness authority to AI training.

By focusing regulatory attention on outputs rather than inputs, AI innovation can flourish while creators receive appropriate protection and compensation. This forward-looking approach acknowledges the dynamic nature of AI markets and the shared goal of all relevant legal frameworks: promoting innovation and enhancing consumer welfare through a vibrant creative economy.

Category: Print Issues

Generative AI: When Fair Use Becomes Unfair Competition

Share this article