Polyphonic AI Regulation

Category: Print Issues

Daryl Lim

Volume 32

, Issue 4

32 GEO. MASON L. REV. 745

Introduction

Hernan Diaz’s Pulitzer Prize-winning novel Trust offers an innovative narrative of wealth, deceit, and power set against the backdrop of 1920s New York, a time when capitalism reigned supreme.¹ The novel masterfully unfolds the same story from four distinct perspectives, each offering a different version of events and challenging the reader to discern the elusive truth.² Through these multiple voices, Diaz explores the interplay between perception and reality, illustrating how wealth and influence can shape, distort, and rewrite history depending on who controls the narrative.³

Just as Diaz’s novel compels readers to question who gets to author the truth in a world of concentrated wealth and influence, today’s debates over generative AI and market power raise similar questions about who shapes technological progress and on whose terms. As a handful of powerful tech companies dominate the development and deployment of foundation models, antitrust regulators guard against market concentration that threatens to stifle innovation and limit consumer choice.⁴ These companies train their systems on massive datasets, often comprising copyrighted works, enabling them to replicate styles, patterns, and structures embedded therein.⁵ These practices expose deep-seated doctrinal tensions between innovation and control, openness and exclusivity, public benefit and private rights.

This Article contributes to the literature by explaining how these tensions arise and how we might resolve them through polyphonic regulation. Drawing on the musical concept of polyphony, where distinct melodic lines retain their individuality while contributing to a harmonious whole, this approach envisions a regulatory framework in which antitrust, copyright, the right of publicity, and information privacy operate in coordinated, non-hierarchical interplay. Each body of law brings normative commitments and institutional logic, offering unique but complementary insights into generative AI’s societal impact. In contrast to frameworks that seek doctrinal unification or convergence through top-down harmonization, polyphonic regulation embraces structured pluralism. Just as Trust compels the reader to reconcile competing narratives to arrive at a deeper understanding, policymakers must harmonize diverse legal perspectives to craft a regulatory response that is both coherent and context sensitive.

Part I examines how generative AI restructures market dynamics across the vertically integrated AI “stack,” from infrastructure and data to foundational models and end-user applications. It analyzes how dominant firms may leverage control at each layer to restrict interoperability, foreclose rivals, and shape innovation pathways. It calls for layered antitrust scrutiny attentive to structural risks and countervailing forces, including open-source competition, synthetic data, and global market entrants. Part II explores how copyright law, particularly the fair use doctrine, mediates conflicts between innovation and authorial control in the context of AI training. It evaluates approaches to transformative use, market substitution, and licensing; and proposes borrowing market analysis tools from antitrust to improve the doctrinal coherence of fair use. Part III turns to commodifying identity and behavioral data in generative AI, where existing frameworks leave critical gaps. It advances tiered data rights, grounded in the right of publicity and information privacy, to protect against unauthorized use of personal attributes and to restore individual autonomy in AI-driven ecosystems.

A. In the Crosshairs of Antitrust

Understanding how generative AI reshapes market structures requires a layered antitrust analysis that accounts for the unique dynamics of data, compute, and model access. Unlike traditional tech markets, AI development is characterized by a vertically integrated “stack” of infrastructure, foundational models, and downstream applications, each with potential choke points. This Part explores how dominant firms may leverage their control at one or more layers to entrench their market positions, restrict interoperability, and foreclose competitors, even without traditional mergers or acquisitions.

1. Antitrust Risks

AI product development typically unfolds in three stages: development by large tech incumbents, innovation by smaller startups, and acquisition or integration into dominant platforms.⁶ This progression highlights the dynamic and competitive nature of AI product development, where innovation at the startup level frequently fuels growth and acquisition strategies at higher levels.⁷ Each stage relies on critical data, computational infrastructure, and technical expertise that have all become increasingly centralized.⁸ Incumbents enjoy disproportionate control over key computational layers and proprietary datasets.⁹ By contrast, smaller players face structural exclusion without shared standards or legal obligations for interoperability.¹⁰

The rise of partnerships like Microsoft-OpenAI and Google-Anthropic, alongside efforts by Amazon, Apple, and Meta, suggests a potential duopoly or oligopoly in foundation model development.¹¹ The high concentration of control over foundational inputs allows a handful of firms to leverage their position in ways that may distort competition, raise concerns about market dominance, and reduce innovation.¹² Determining when strategic partnerships and data-sharing agreements among AI developers cross into antitrust violations requires careful legal parsing.

Antitrust jurisprudence has long recognized the importance of preserving competitive dynamics not for the sake of competitors, but rather to safeguard competition itself, which ultimately benefits consumers through lower prices, higher quality goods and services, and increased innovation.¹³ Under Section 1 of the Sherman Act, courts focus on identifying whether a prohibited “contract, combination . . . or conspiracy” exists.¹⁴ The critical factor is concerted action involving a “conscious commitment to a common scheme designed to achieve an unlawful objective.”¹⁵ The law targets agreements that deprive the marketplace of independent decision-making, posing unique anticompetitive risks compared to unilateral action, regardless of the latter’s potentially aggressive market impact.¹⁶ Liability requires proof, often circumstantial, of an agreement or concerted action, rather than mere parallel behavior or independent conduct.¹⁷ This analytical precision helps antitrust regulators differentiate between legitimate collaborative innovation and unlawful coordination that undermines competition. This distinction is crucial to evaluating the complex interplay of firms within the layered GenAI stack, a layered model encompassing infrastructure (e.g., chips and cloud services), foundation models, and applications.

Understanding the nature of coordination is only one part of the inquiry. Dominant firms may engage in exclusionary conduct under Section 2 of the Sherman Act that exploits structural advantages across the AI stack.¹⁸ Historical antitrust precedents on exclusionary conduct, such as United States v. Microsoft Corp.,¹⁹ underscore the potential competitive risks of proprietary systems that inhibit interoperability.²⁰ Modern AI markets mirror these dynamics: Dominant firms may foreclose rivals by controlling access to training data and model Application Programming Interfaces (“APIs”).²¹ The lack of interoperability between AI systems exacerbates this problem, as smaller firms cannot develop alternative models or engage on equal footing.²² Without shared standards or legal frameworks to govern data access and use, dominant firms retain disproportionate influence over the rules of engagement in AI-driven markets.

Coupled with access bottlenecks is the risk that incumbents will control the direction of innovation. Dominant AI platforms may benefit from what Hideyuki Matsumi and Daniel Solove describe as a “self-fulfilling prophecy problem.”²³ Platforms reflect and shape demand by designing systems that predict and amplify certain user behaviors, such as what they will click, consume, or create.²⁴ These feedback loops can entrench incumbents, marginalize alternative creators, and bias algorithmic training toward commercially dominant content, thereby reinforcing the incumbents’ market position under the guise of user choice or personalization.²⁵

Regulators examine AI markets through the “GenAI stack” lens to counteract anticompetitive threats.²⁶ For instance, the DOJ and FTC have launched parallel antitrust investigations into AI market leaders Microsoft, OpenAI, and Nvidia, reflecting a coordinated federal effort to assess potential anticompetitive conduct.²⁷ The “GenAI stack” framing allows agencies to identify potential bottlenecks and anticompetitive conduct at each layer, including exclusive access to compute and interoperability restrictions in end-user tools. Regulators also focus on conditional dealing, including bundling, exclusivity arrangements, and loyalty discounts.²⁸ These practices may serve procompetitive functions, such as improving user experience or reducing transaction costs, but they can also foreclose competition if not properly limited or disclosed.²⁹

As the U.S. Supreme Court famously articulated, “antitrust laws were passed for ‘the protection of competition, not competitors,’” emphasizing that their purpose is “not to protect businesses from the working of the market; it is to protect the public from the failure of the market.”³⁰ Moreover, antitrust laws target conduct “which unfairly tends to destroy competition itself,” driven by a broader concern for public interest rather than private entities.³¹ This foundational antitrust principle reinforces the need for careful regulatory oversight in AI markets to prevent exclusionary or anticompetitive practices without stifling legitimate competitive vigor. Hence, a fuller understanding of generative AI markets also requires engagement with countervailing narratives that caution against regulatory overreach and highlight the dynamic potential of open-source innovation and market entry.

B. Countervailing Narratives

Not all observers agree that AI markets require aggressive antitrust intervention. For instance, Satya Marar cautions against overreliance on killer acquisition rhetoric without case-specific analysis of innovation effects and market context.³² Marar reported that a 2020 survey of 175 tech acquisitions by major platforms found only one potential killer acquisition and argued that acquisitions integrate talent and technology to accelerate innovation by eliminating double marginalization, enabling economies of scale, or subsidizing model deployment across platforms.³³

Policymakers can also overestimate the entry barriers in AI markets by focusing primarily on the dominance of large incumbents and the control they exert over foundational technologies. For instance, few would have predicted twenty years ago that Google and Apple would become the most powerful companies in the world. Google controls the internet by solving the problem of information overload.³⁴ Similarly, Apple solved the mobility problem by introducing the iPhone.³⁵ Yet despite Google’s dominance in search, Android phones have not eclipsed Apple’s dominance in mobile hardware.³⁶

In AI markets, open-source AI models are powerful forces that can mitigate barriers to entry. Their meteoric rise offers a compelling counternarrative to incumbent entrenchment. For instance, DeepSeek, launched in early 2025, was trained with second-tier chips but outperformed Meta’s best models on some benchmarks at a fraction of the cost.³⁷ DeepSeek R-1 achieved strong performance despite constrained hardware access, suggesting architectural efficiency can compensate for infrastructure gaps.³⁸ Open-source platforms allow a global community of researchers, developers, and innovators to pool collective expertise, refine models, and address technical challenges collaboratively. As Kai-Fu Lee observed, DeepSeek’s disruptive pricing and technical efficiency have led companies like 01.AI to abandon proprietary model development entirely in favor of customizing DeepSeek’s offerings.³⁹ According to Lee, the “biggest nightmare for Sam Altman is that his competitor is free,” a shift already prompting cancellations of paid ChatGPT subscriptions.⁴⁰ This suggests that innovation may come from system-level efficiency and creative workarounds, not just brute-force computing and data dominance.

Regulators are convinced that open standards will help ensure viable competition in AI markets.⁴¹ At the same time, the FTC has flagged its concern over “open first, closed later” strategies, where firms release open-source models to attract users and accrue data advantages, only to close access later and lock users into proprietary ecosystems.⁴² This practice may undermine long-term competition by creating data-driven “moats” while superficially appearing procompetitive.⁴³ Many companies train open models on corpora curated or hosted by dominant firms.⁴⁴ Moreover, as firms impose new licensing restrictions on model weights and outputs, the promise of openness may increasingly resemble a strategic deployment of selective transparency. In this light, open-source AI serves less as a structural corrective and more as a contingent, contestable layer within an ecosystem still shaped by concentration in data, compute, and distribution.

A second source of competitive pressure comes from off-the-shelf AI models, available through open-source platforms and commercial APIs that enable smaller companies to customize and personalize AI applications in niche markets without massive data resources. For example, Hugging Face and its Model Hub provide pre-trained models that smaller firms can fine-tune for specific applications, reducing the need for costly training runs.⁴⁵ This accessibility reduces the competitive gap between large incumbents and smaller players, allowing startups and mid-sized companies to build innovative applications that meet specific market demands. In this environment, the advantage increasingly lies in leveraging and refining foundation models for specialized use cases, not necessarily in controlling vast datasets. As a result, the source of competitive advantage may shift away from access to vast proprietary datasets to foundation models that can be fine-tuned for specific applications.

The third source of competitive pressure comes from global competition. The DeepSeek model and the wave of Chinese AI models that followed illustrate how global competition is reshaping the economics of generative AI. DeepSeek’s success triggered a proliferation of high-performing, open-source models across China’s tech sector, including from firms like Baidu, Alibaba, and Tencent.⁴⁶ Competitors rapidly introduce these models abroad, undercutting premium offerings from firms like OpenAI and Google.⁴⁷ China’s rapid catch-up in generative AI has reframed assumptions about innovation and access. The resulting business model shockwaves raise regulatory questions about access and the sustainability of closed-model economics in a fast-moving AI ecosystem. It also suggests that U.S. enforcement priorities must consider competitive dynamics beyond national borders.

Fourth, access to data is becoming less relevant as an entry barrier to AI markets. Synthetic data allows firms to bypass the need for large-scale data collection.⁴⁸ This trend reflects a growing recognition that we may have reached “peak data,” as experts like Demis Hassabis and Ilya Sutskever have observed.⁴⁹ Firms can use synthetic data generated by AI models to simulate real-world datasets to train and fine-tune models without requiring access to vast amounts of sensitive or proprietary data.⁵⁰ Synthetic data can also empower smaller companies to develop competitive AI models that rival data-rich incumbents by generating high-quality, diverse, and scalable training data.⁵¹

Further, the finite nature of internet-scale datasets limits the scalability of current training paradigms, pushing developers toward efficiency, architectural innovation, and synthetic augmentation. Facing the limitations of the “more data, more compute” paradigm, OpenAI has shifted toward developing “reasoning” models like o3 and o4-mini which are capable of outperforming state-of-the-art models.⁵² The model breaks from prior large language models (“LLMs”) by focusing on step-wise reasoning through problems, offering more deliberate and accurate outputs in areas previously susceptible to hallucination or guesswork.⁵³ OpenAI’s Noam Brown noted that in some cases, adding twenty seconds of reasoning time improved performance as much as scaling the model 100,000-fold.⁵⁴ These constraints elevate the strategic importance of curated small datasets and may shift the locus of competitive advantage away from raw data accumulation.

In sum, AI markets are complex, and antitrust enforcement in AI markets should be flexible, calibrated, and empirically grounded. Under neo-Brandeisian leadership, the FTCchallenged the narrative that technological progress lifts all boats.⁵⁵ Instead, it saw a real risk that unchecked technological progress would exacerbate inequities beyond cost, access, and innovation concerns.⁵⁶ On the other hand, critics have argued that recent antitrust investigations into AI companies like Nvidia depart from the consumer welfare standard, the traditional benchmark in U.S. competition law, instead reflecting a “big is bad” mentality. For instance, Jennifer Huddleston warns that regulators increasingly emphasize competitor protection and hypothetical market harms over tangible consumer outcomes such as price, innovation, and choice.⁵⁷ Favoring structural presumptions and rejecting the consumer welfare standard may impose high administrative and error costs.⁵⁸ Marar contends that such approaches risk chilling procompetitive mergers, especially in innovation-driven sectors like AI, by overemphasizing concentration and underestimating market dynamism and efficiency gains from integration.⁵⁹

Generative AI developers often rely on copyrighted materials to train their models, incorporating vast datasets that include books, music, images, and other creative works.⁶⁰ This practice has prompted copyright infringement lawsuits.⁶¹ While proponents of generative AI argue that training datasets transform copyrighted works into new, independent outputs, critics contend that these practices exploit copyrighted works without acknowledgment or compensation.⁶² This conflict underscores the intersection of copyright and antitrust law, as dominant firms’ ability to harness copyrighted materials without oversight raises infringement questions and implicates antitrust concerns by concentrating market power. For the reasons above and those that follow, antitrust regulators should be slow to condemn text and data mining using copyrighted materials under antitrust law.

C. Antitrust Law in Copyright Markets

In December 2023, responding to the U.S. Copyright Office’s inquiry on AI and copyright, the FTC signaled its concern over the competitive implications of generative AI technologies.⁶³ First, the FTC warned that permissive copyright regimes might enable powerful AI developers to leverage proprietary data to monopolize downstream markets further, potentially disadvantaging smaller creators and innovators.⁶⁴ Second, the FTC highlighted that using copyrighted content to train machine learning models could create anticompetitive effects by diminishing the future value of original works.⁶⁵ When this erosion of value results from the misuse of copyrighted content or the monopolistic control of training data, it may cross the line into anticompetitive behavior. Third, the FTC flagged concerns regarding indemnification practices, where AI companies may shield themselves from liability by shifting the legal risk to downstream users.⁶⁶ The concern is that this practice could further entrench market dominance by creating asymmetrical risk distribution, disadvantaging smaller competitors who lack the resources to absorb such liabilities.

Significantly, the FTC’s letter also made clear its willingness to deploy antitrust law to curb anticompetitive behavior by AI developers, even in scenarios where those developers have not technically infringed upon copyrighted material or have relied on fair use defenses.⁶⁷ As Peter Yu and the Author observed, the interface between antitrust and copyright faces technological and ideological disruption.⁶⁸ Technologically, the elements that antitrust law seeks to regulate, such as access to data, computing, and distribution channels, are essential to the success of AI developers.⁶⁹ Ideologically, the policy terrain has shifted from Chicago School orthodoxy to Neo-Brandeisian structuralism, complicating how regulators approach AI markets.⁷⁰ Targeting indemnification practices as anticompetitive could endanger AI developers from shielding users, who are often entangled in emerging copyright disputes, from legal risk. Yet banning indemnity clauses in only one direction (e.g., from users to developers but not vice versa) may do little to curb coercive dynamics unless policymakers address mutual indemnification more comprehensively. Indemnification is a standard industry practice that promotes user adoption and mitigates uncertainty in emerging technologies.⁷¹ Treating this practice as suspect risks chilling innovation and deterring responsible AI development.⁷²

Whether antitrust law should be the vehicle for addressing copyright-adjacent harm remains a matter of institutional design, particularly where copyright law already incorporates competition-sensitive doctrines such as fair use. Doctrinally, the FTC blurs the line between competition policy and copyright law by signaling its willingness to use antitrust law even where AI developers succeed in their fair use defense.⁷³ This approach raises concerns about regulatory overreach, potentially subjecting AI firms to duplicative and conflicting enforcement regimes.

II. Copyright’s Competition Policy

If antitrust law focuses on competitive structure and market access, copyright law governs generative AI’s expressive inputs and outputs. Together, they help frame the legal boundaries of innovation by addressing concentration and exclusion while also defining permissible use and market harm. This Part examines how copyright doctrine, particularly fair use, shapes the development and dissemination of AI-generated content. Current opt-out mechanisms designed to protect copyrighted works, such as robots.txt, are ineffective at scale.⁷⁴ The result is a system where non-consensual use has become the default, not the exception.⁷⁵ Therefore, whether AI developers can use content without prior authorization has far-reaching consequences for the creative and AI industries.

A. Fair Use

Copyright law promotes creativity by granting authors exclusive rights to their work and ensuring creators receive economic rewards.⁷⁶ Fair use limits those exclusive rights for uses aimed at developing new technologies that provide significant public benefits.⁷⁷ Courts apply a four-factor test: (1) the purpose and character of the use, including whether it is transformative; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used; and (4) the effect of the use on the potential market for or value of the work.⁷⁸ Courts have often emphasized the first and fourth factors as particularly important.⁷⁹

In an amicus brief filed in Kadrey v. Meta Platforms, Inc.,⁸⁰ a few law professors argued that Meta’s ingestion of entire works to train its LLaMA models was not “transformative.”⁸¹ Unlike parody or commentary, which reinterpret a particular work, Meta’s training purpose was to merely absorb the expressive content to generate substitute outputs, often in the same market.⁸² The brief highlights that plaintiffs’ works were intended to cultivate knowledge and skill, precisely the functional outcome Meta’s models aim to reproduce.⁸³ Consequently, there is no divergence in purpose sufficient to sustain a transformative use defense.

Moreover, the professors challenge Meta’s reliance on analogies to “non-expressive use” and “intermediate copying,” arguing that the training process is fundamentally expressive: LLaMA models are optimized to reproduce linguistic patterns drawn from copyrighted works.⁸⁴ Meta’s “commercial” deployment of these models further weakens its fair use claim, particularly given its revenue-sharing arrangements with model-hosting platforms.⁸⁵ The brief underscores that allowing AI firms to copy works en masse without a license erodes the copyright system’s incentives for human creativity and puts authors at a structural disadvantage.⁸⁶

These arguments echo the arguments heard elsewhere on this issue.⁸⁷ The conflict itself echoes earlier battles over library photocopying and corporate research copying. In Williams & Wilkins Co. v. United States,⁸⁸ the court found fair use in scientific research photocopying by nonprofit institutions, emphasizing minimal market harm and public benefit.⁸⁹ By contrast, in American Geophysical Union v. Texaco,⁹⁰ the court ruled against fair use where photocopying by corporate researchers undermined the emerging licensing market operated by the Copyright Clearance Center.⁹¹ These cases suggest that fair use defenses for AI training may hinge on whether the courts view current uses as transformative and whether robust licensing alternatives exist.

B. Transformative Use and Public Interest

The Supreme Court’s decision in Campbell v. Acuff-Rose Music, Inc.⁹²510 U.S. 569 (1994). introduced the concept of “transformative use” as a touchstone for fair use.⁹³ This concept was later expanded in Google v. Oracle America, Inc. ⁹⁴ to consider public benefits, particularly in the context of technological advancement.⁹⁵ In that case, Google’s use of Oracle’s Java API was deemed fair use, partly because it enabled the development of a new and innovative platform that benefited developers and consumers alike.⁹⁶

The Kadrey amici argue that training LLMs on copyrighted works is not transformative because the models do not parody, comment on, or add meaning to individual works.⁹⁷ However, this is a reductive reading of Campbell and Google, which explicitly recognizes that transformation may arise not from commentary on a specific work but from new analytical or functional use.⁹⁸ LLMs use copyrighted inputs to detect semantic patterns at scale, not to re-express those works but to build statistical relationships among words, grammar, and ideas.⁹⁹ This functional repurposing represents a paradigmatic example of transformative use.

As discussed in Section I.B, the evolution of “reasoning systems” further proves that LLM training serves a fundamentally functional and not expressive purpose. Modern chatbots “think[] through” problems by decomposing tasks, iterating on solutions, and rechecking intermediate steps before producing an answer.¹⁰⁰ This behavior, facilitated by reinforcement learning, reflects not the expressive reproduction of source texts but the development of generalizable problem-solving strategies.¹⁰¹ These processes rely on input data but do not merely reproduce that input’s expressive character. As search engines or plagiarism detectors rely on copyrighted content without copying expressive meaning, so do LLMs extract insights to enable new interaction and understanding, not substitution. It is also worth mentioning that the brief’s reliance on the “intermediate copying” doctrine misapplies cases like Sega Enterprises Ltd. v. Accolade, Inc. ¹⁰² and Sony Computer Entertainment, Inc. v. Connectix Corp.,¹⁰³ which validated fair use precisely because intermediate duplication was necessary to understand and interact with underlying systems.¹⁰⁴

From a competition policy perspective, AI-powered tools can lower barriers to entry for small businesses, entrepreneurs, and underserved communities by enabling more efficient operations, personalized learning, and improved decision-making.¹⁰⁵ For example, AI can empower small businesses to access sophisticated analytics, automate routine processes, and expand their market reach without requiring the resources of large corporations.¹⁰⁶ A 2023 survey by 99designs found that over half of freelance designers use generative tools like Midjourney and Adobe Firefly for concept exploration and layout experimentation.¹⁰⁷ These tools accelerate workflow and democratize access, particularly for small firms that lack the resources of major studios. This broadening of creative participation may strengthen arguments for fair use under the first factor by framing AI as an accessibility-enhancing tool rather than a replacement for human creativity.

Moreover, while the concerns about AI-driven job displacement and rising inequality are valid, historical trends suggest that technological progress creates new opportunities and industries that ultimately offset the losses caused by automation.¹⁰⁸ Throughout history, disruptive technologies, from the Industrial Revolution to the advent of the Internet, have initially displaced certain types of labor but eventually led to the emergence of new sectors and job categories that were previously unimaginable.¹⁰⁹ AI, rather than solely eliminating jobs, has the potential to augment human capabilities and shift the workforce toward more creative, strategic, and cognitively demanding tasks that machines cannot easily replicate.

Governments, industry leaders, and educational institutions are critical in designing retraining programs, supporting workforce transition, and fostering lifelong learning initiatives to equip workers with the skills necessary to thrive in an AI-driven economy. These are labor challenges, not copyright or antitrust ones. Countries that prioritize workforce adaptation and create pathways for reskilling can mitigate the negative effects of technological disruption and harness AI’s potential to enhance workforce productivity and economic inclusivity. In the copyright context, the goal should not be to prevent disruptive technologies but to ensure that legal doctrines and enforcement strategies adapt to protect competition and innovation in tandem.

C. Market Harm

Courts have weighed the market effect factor heavily in their fair use analysis.¹¹⁰ Challenged uses that substitute and undermine the core commercial value of the original work or harm the potential market for or value of the original work are disfavored.¹¹¹ However, not all market harm is cognizable under copyright law. Pamela Samuelson’s analysis of disruptive technology cases suggests that courts often uphold fair use when evidence of market harm is speculative or indirect, especially when the secondary use is transformative and serves public or educational purposes.¹¹²

1. Where Content Owners Are Likely to Succeed

Content owners will likely succeed when unauthorized copying supplants or forecloses a reasonably anticipated licensing market. In Thomson Reuters Enterprise Centre GMBH v. Ross Intelligence Inc.,¹¹³ the court held that Ross Intelligence’s use of Westlaw headnotes and the key number system to train a competing AI legal research tool was not protected by fair use.¹¹⁴ Ross Intelligence’s use was not transformative because it served the same purpose as Westlaw’s systems, even though the copied headnotes were not visible in the final product.¹¹⁵ Significantly, harm to potential derivative markets was sufficient to defeat fair use.¹¹⁶ Even if Thomson Reuters had not yet licensed its headnotes for AI training, the court accepted that such a market could reasonably be expected to develop.¹¹⁷ Licensing markets for AI training data alter the fair use calculus and place the burden on AI developers to avoid foreseeable substitution effects.

While Ross offers an early judicial roadmap, its precedential value for generative AI may be limited. The case involved non-generative AI, and a narrow dataset (25,000 “bulk memos”) closely modeled on copyrighted headnotes.¹¹⁸ The court did not assess whether the model’s outputs infringed upon or examine how the AI learned from the inputs. While courts are receptive to arguments about actual or probable market harm, they are reluctant to infer harm solely from the failure to license when no coherent or scalable licensing system is in place.¹¹⁹ As Pamela Samuelson argued, allowing such claims would transform fair use into a dead letter and incentivize litigation as a tool to manufacture markets.¹²⁰

It is also worth noting that while licensing markets for AI training data may seem like a good middle ground between uncompensated use and outright prohibition, input licensing regimes may paradoxically entrench incumbent power structures. Large publishers and rights holders are best positioned to negotiate favorable licensing terms, enforce rights at scale, and integrate into commercial AI pipelines. By contrast, independent creators, academics, and small rights holders often lack the bargaining power, visibility, or infrastructure to participate meaningfully in these arrangements. The result is a bifurcated licensing landscape: Dominant firms obtain lawful access to premium inputs, while marginalized creators remain excluded or exploited. Moreover, the emergence of high-value licensing deals, such as those between major academic publishers and AI developers, may channel resources to already-powerful intermediaries, reinforcing their gatekeeping roles and reducing the relative visibility or discoverability of smaller creators. Thus, even well-intentioned licensing frameworks risk reifying informational hierarchies and converting copyright into a pay-to-play system that mirrors existing inequalities in the creative economy.¹²¹

2. Where AI Developers Are Likely to Succeed

In the lawsuit the New York Times (“NYT”) filed against OpenAI and Microsoft, the plaintiffs argued that the unlicensed use of NYT content to train AI models like ChatGPT and Bing Chat directly undermines their subscription-based business model.¹²² They contend that these AI tools allow readers to bypass the NYT’s paywall by enabling users to access summaries or near-verbatim reproductions of their articles without subscription, thereby “obviat[ing] the need” to purchase access.¹²³ According to the complaint, this poses a significant threat to the newspaper’s ability to fund its journalism and maintain its subscriber base.¹²⁴

A closer examination of the facts suggests that this narrative is overstated and fails to account for the realities of the media landscape. The NYT remains a thriving enterprise, far from being financially harmed by generative AI. In 2024, The New York Times Company reported total revenue of $2.59 billion, marking a 6.6% increase from the previous year.¹²⁵ It added 350,000 subscribers in just one quarter to its total subscriber count of over 11.4 million.¹²⁶ This robust performance undermines the claim that AI models trained on NYT content are causing substantial harm to its market position.

Moreover, the NYT’s insistence that OpenAI’s failure to obtain a license amounts to copyright infringement is a circular argument. Licensing is not a prerequisite for fair use; the very purpose of the doctrine is to allow certain uses without permission when the public interest outweighs the potential harm to the copyright holder.¹²⁷ After all, copyright is utilitarian.¹²⁸ Its constitutional purpose is to promote the progress of science and the useful arts, not to shield incumbents from fair competition.¹²⁹ This perspective supports a balanced approach that avoids presuming harm in emerging AI contexts. If AI-generated outputs do not serve as market substitutes and instead provide new, non-substitutive insights or functionalities, the absence of a license becomes irrelevant to the fair use analysis.

The Kadrey brief similarly overstated the relevance of commerciality. While Meta may earn revenue through LLaMA-based applications, Campbell instructs that commercial uses can still be fair when they are transformative.¹³⁰ U.S. copyright law has consistently allowed functional repurposing of works, even by for-profit actors, when the new use furthers access to information, expands knowledge, or enables research.¹³¹ The amici’s assertion that AI model outputs inherently compete with human-authored works also lacks nuance. Generative outputs may occupy overlapping expressive domains, but this is not dispositive. Market harm must be concrete, not speculative, and must account for whether the use usurps a traditional, reasonable, or likely-to-be-developed licensing market.

A further complication in assessing fair use claims in generative AI is what Hideyuki Matsumi and Daniel Solove call the “unfalsifiability problem.”¹³² Predictions about future behavior, such as whether an AI system will substitute for a human author, are unverifiable when the claim is made. This temporal indeterminacy complicates the application of fair use’s fourth factor, which relies on credible evidence of actual or potential market harm. Courts accustomed to evaluating harm based on past or current facts are poorly positioned to adjudicate speculative or emergent markets, especially where AI tools are still evolving.

While concerns about attribution, transparency, and fair compensation for creators are valid and deserving of attention, it is essential to move beyond the overly simplistic narrative of “intellectual theft” that often dominates public discourse. This narrative obscures the fact that AI training is transformative, often resulting in outputs distinct from the underlying works. As Ross demonstrated, courts can fully address genuine instances of market substitution and competitive harm without resorting to expansive theories of infringement that conflate transformative AI uses with direct competition.¹³³

In sum, when generative AI models ingest copyrighted works to generate new, distinct outputs, and when plaintiffs fail to show actual or likely harm to existing or traditional licensing markets, fair use defenses may be viable. The assertion that every use must be licensed does not hold if the use is transformative and does not replace the market for the original. Aggressive copyright enforcement against AI firms may stifle beneficial experimentation and generate false positives that deter innovation in the public interest, especially in education, healthcare, or research.

3. Borrowing from Antitrust

The FTC’s 2023 submission to the Copyright Office drew sharp criticism from scholars who argue that antitrust intervention could upset copyright’s delicate balance between access and exclusivity, especially if fair use is recharacterized as anticompetitive.¹³⁴ Rather than supplanting copyright law, the FTC’s approach—if carefully calibrated—could serve as a backstop to address conduct that copyright law either permits or cannot reach, such as exclusionary access arrangements or self-preferencing within AI ecosystems. The key is ensuring this supplementary role remains clearly defined, lest regulatory polyphony devolve into doctrinal cacophony.

Antitrust doctrine can also inform fair use analysis. Chris Sprigman proposes importing analytical tools from antitrust law, particularly the hypothetical monopolist test and market definition methodologies, to assess whether a defendant’s work substitutes for the plaintiff’s.¹³⁵ Courts applying antitrust law know how to define relevant markets, assess substitutability, and determine whether the accused work impairs the plaintiff’s ability to monetize the original.¹³⁶ For example, if a small but significant increase in the licensing price of the plaintiff’s work would cause consumers to switch to the defendant’s work, the works may be substitutable. If not, they may operate in distinct markets, bolstering the case for fair use.

Sprigman’s focus on substitution acts as a clarifying principle, deploying antitrust-informed tools to render the fair use inquiry more predictable, less subjective, and ultimately more faithful to copyright’s constitutional purpose. By treating fair use as a question of market dynamics, courts can move beyond artistic interpretation and evaluate whether the secondary use impairs the economic incentives copyright seeks to preserve.¹³⁷ This analytical lens aligns with the polyphonic model advocated in this Article: It invites copyright doctrine to learn from antitrust, not by collapsing normative frameworks, but by cross-pollinating methods where appropriate. Substitutability in copyright law need not adopt antitrust’s consumer welfare test wholesale. Rather, policymakers can retool it to reflect copyright’s distinct objectives, promoting creativity and balancing incentives, while benefiting from antitrust’s empirical discipline.

While antitrust and copyright law primarily address economic dimensions of generative AI—competition, innovation, and market harm—they fall short in capturing the personal and dignitary injuries that arise when AI systems co-opt individual identity or behavioral data. These non-economic harms implicate deeper concerns about autonomy, selfhood, and consent, especially as AI technologies increasingly replicate human likeness, voice, and behavioral patterns. To address these gaps, the next Part turns to the legal frameworks of the right of publicity and information privacy, which offer a foundation for recognizing and remedying the commodification of identity in AI training and deployment.

III. Towards a Data Right

With generative AI becoming increasingly capable of generating lifelike images, voices, and personas, concerns about the misuse of an individual’s identity for commercial gain have intensified.¹³⁸ In April 2023, “ghostwriter977” released a viral track titled Heart on My Sleeve using AI-generated voices of Drake and The Weeknd.¹³⁹ Copyright law does not protect the human voice itself, which prompted artists and industry groups to seek alternative remedies in the right of publicity.¹⁴⁰ A year later, in July, 2024, the SAG-AFTRA actors’ union called a strike against major video game companies, demanding stronger safeguards against AI-generated impersonations of performers’ voices and likenesses.¹⁴¹

The concentration of control over data and information in the hands of a few powerful corporations enables the shaping of narratives.¹⁴² Without effective safeguards, these forces could contribute to a society where public opinion is influenced not by facts and reason but by manipulated content crafted by AI systems. This Part argues that a data right, grounded in the right of publicity and information privacy, can help restore individual agency in data-driven economies.

A. Commercialization of Identity

As early as 1890, Samuel Warren and Louis Brandeis argued for a privacy “right to be let alone” laying the conceptual groundwork for publicity rights.¹⁴³ Traditionally used to protect individuals from the unauthorized commercial exploitation of their name, image, and likeness, the right of publicity can be extended to encompass control over the digital footprint that individuals leave behind.¹⁴⁴ Remedies may include statutory damages, disgorgement of profits, injunctive relief, or punitive damages in cases of willful infringement.¹⁴⁵ Importantly, plaintiffs are generally not required to prove actual harm, making the doctrine a powerful tool for challenging unauthorized AI-generated likenesses.¹⁴⁶

Orin Kerr’s “equilibrium-adjustment” theory posits that legal doctrines evolve to preserve individual protections as technology shifts power toward the state or private actors.¹⁴⁷ Applying this framework to AI and identity suggests that the right of publicity may require doctrinal recalibration, expanding in scope or remedy to preserve individuals’ control over their digital personhood. One salient example of this regulatory turn is Tennessee’s passage of the Ensuring Likeness, Voice, and Image Security (“ELVIS”) Act of 2024, which expanded the state’s existing right of publicity framework to explicitly prohibit the unauthorized commercial use of a person’s AI-generated voice or image.¹⁴⁸ The Act defines “voice” to include actual and simulated sounds that are “readily identifiable” as belonging to a particular individual, even if no recording of the person’s real voice was used.¹⁴⁹ In doing so, the law targets precisely the kind of voice cloning and synthetic likenesses that generative AI now enables at scale. Importantly, liability under the ELVIS Act may extend beyond creators to those who develop or distribute generative tools and platforms.¹⁵⁰

Two major federal proposals, NO AI FRAUD and NO FAKES, seek to create new intellectual property-style protections for digital replicas of likenesses and voices.¹⁵¹ Both bills define the protected right expansively, covering nearly any audio or visual material generated, altered, or simulated with digital technology.¹⁵² Notably, by labeling these rights as federal intellectual property, the legislation sidesteps Section 230 immunity, putting hosting platforms squarely in the litigation path. The pending bills also face criticism for potential First Amendment overreach and platform liability concerns.¹⁵³

Beyond publicity rights, the asymmetry of power between data controllers and ordinary citizens creates a situation where corporate algorithms, rather than individuals, choose what information is collected, how it is used, and how it influences behavior. This power imbalance undermines meaningful consent and autonomy, giving consumers little control over their digital identities, an issue related to the right of publicity. It also creates privacy concerns. The emergence of surveillance capitalism has recast consumers as suppliers of personal data, not merely users of services.¹⁵⁴

A data right could create a more transparent and accountable AI ecosystem by allowing individuals and entities to license, withhold, or monetize their data for AI training. Additionally, the right could allow for the development of fair compensation models where creators whose works contribute to the advancement of AI receive a share of the economic value generated by AI output. A data right could be a statutory entitlement akin to copyright or publicity rights, enforceable via collective licensing schemes, opt-out registries, or fiduciary data intermediaries. It would enable creators and users alike to exert granular control over how companies use their data and ensure more equitable participation in the value chains that generative AI creates.

B. Control of Personal Data

Information privacy law regulates how personal data is collected, processed, and stored.¹⁵⁵ Platforms compete to addict, track, and manipulate users to maximize behavioral advertising revenues.¹⁵⁶ Data is a competitive asset and a personal resource with intrinsic value in AI-driven markets.¹⁵⁷

The intersection between antitrust and information privacy occurs when dominant AI firms leverage their control over massive datasets to entrench market power and stifle competition.¹⁵⁸ The FTC has signaled its intention to scrutinize how data practices affect market power and whether privacy measures are being weaponized to exclude competitors.¹⁵⁹ Recent discussions around updating antitrust frameworks emphasize the need to consider data governance practices when assessing anticompetitive behavior and formulating remedies.¹⁶⁰ Maurice Stucke characterizes the evolving regulatory approach to privacy and competition in three stages: Privacy/Competition 1.0 (complete separation), 2.0 (privacy as a non-price parameter of competition), and 3.0 (emergent tensions and tradeoffs).¹⁶¹

Four issues are key in thinking about a data right based on privacy. First, while earlier stages saw privacy and competition as complementary levers to check the power of “data-opolies,” Stucke’s third stage warns that increased competition may degrade privacy, especially when firms exploit user data rather than serve user interests. Dominant platforms may use privacy compliance as a justification to limit data portability or interoperability, thereby preventing smaller firms from gaining access to the data necessary to develop competitive AI products. This dynamic can create a data “moat,” where incumbents solidify their dominance while claiming adherence to privacy standards.¹⁶² In this context, privacy laws that restrict data collection or require explicit consent can inadvertently create barriers to entry for smaller competitors who lack access to similarly rich datasets.

While antitrust can advance privacy indirectly by enhancing consumer choice and reducing lock-in, it is not well-suited to define or enforce privacy rights. The American Antitrust Institute warns that shoehorning privacy into antitrust may dilute the doctrinal coherence of competition law and distract from its central aim of preserving contestable markets.¹⁶³ A better approach is coordinated, but not conflated, enforcement that recognizes when privacy restrictions serve legitimate ends and when they distort competitive conditions.

Second, regulatory compliance often requires substantial financial and human resources, including investments in legal expertise, compliance officers, and ongoing monitoring systems. Large technology companies are better equipped to absorb these costs with their established infrastructure and deep financial resources.¹⁶⁴ By contrast, smaller firms and startups with limited budgets and lean teams may struggle to meet regulatory requirements, diverting valuable resources from innovation and product development.

Empirical data confirms this disparity. One study found that privacy compliance led to an average 20% increase in cloud data storage costs, with the smallest firms facing a 25% increase compared to just 13% for the largest firms.¹⁶⁵ These findings underscore how compliance costs scale regressively, reinforcing incumbents’ advantage by raising fixed costs that deter entry and suppress innovation among smaller firms. Privacy enforcement also increased market concentration in digital advertising and web technology markets. For example, the implementation of the General Data Protection Regulation (“GDPR”) led to a 17% short-run increase in concentration as websites reduced vendor usage, gravitating toward dominant players like Google and Meta.¹⁶⁶ Finally, privacy compliance has also been shown to dampen new firm formation. An analysis of venture investment data found a 30.7% drop in EU tech deals relative to the United States after the GDPR rollout, particularly for firms reliant on data-intensive services.¹⁶⁷ This suggests that privacy laws may deter entry and stifle the emergence of innovative competitors.

Third, antitrust and privacy law share a problematic assumption: Consumers will make rational, preference-consistent choices in data-driven markets. Antitrust assumes privacy can be priced and that privacy competition will emerge naturally; through notice and consent, privacy law presumes individuals can evaluate tradeoffs effectively.¹⁶⁸ In cases such as hiQ Labs, Inc. v. LinkedIn Corp. ¹⁶⁹ and Epic Games, Inc. v. Apple, Inc.,¹⁷⁰ courts have framed privacy as either a commercial justification or collateral concern rather than a co-equal legal interest.¹⁷¹ However, mounting evidence shows that asymmetric information, dark design patterns, and cognitive overload often frustrate these assumptions.¹⁷²

Fourth, leveraging technology also helps. As seen in Part I, synthetic data lowers the expenses associated with data acquisition and labeling, democratizes access to high-quality training data, and makes AI development more accessible for smaller companies and research institutions.¹⁷³ It is also a boon to contestable markets under the antitrust policy lens and can reduce concerns about data privacy by eliminating the need to use sensitive real-world data.

The right of publicity guards against unauthorized commercial use of persona, while privacy law protects against surveillance, manipulation, and loss of agency. Yet the boundaries blur in practice. Voice cloning implicates dignity and autonomy-based harm; extracting personality from behavioral data (e.g., affective computing) may not fall under either doctrine. Harmonizing these approaches requires a shared framework grounded in informational self-determination, not merely market-based valuation. A dual-track structure would align legal recognition with the underlying normative distinctions between expressive identity-based claims and informational control claims, echoing China’s differential treatment of biometric editing versus behavioral profiling under its Personal Information Protection Law.¹⁷⁴

C. Reform and the Burdens of Regulation

The regulatory governance of generative AI is not neatly partitioned. The four legal regimes considered in this Article—antitrust, copyright, the right of publicity, and information privacy—frequently intersect in ways that amplify or undermine one another, producing redundancies and blind spots. For instance, copyright and antitrust converge where firms use copyrighted materials to entrench market power. While copyright grants exclusive control over expression, that exclusivity may be leveraged to dominate training data markets or restrict downstream applications. The FTC has flagged such risks, particularly when licensing bottlenecks or indemnification practices foreclose entry.¹⁷⁵ However, antitrust and copyright also risk doctrinal collision: Courts may double-count harm or impose conflicting duties unless they take care to preserve their distinct logical structures—one economic, the other expressive.¹⁷⁶

Similarly, as policymakers consider establishing a data right informed by publicity and privacy rights, it is essential to distinguish between two conceptual and doctrinally distinct categories of harm: personality-based claims and behavioral or informational control claims. While both arise from commodifying identity in generative AI systems, they implicate different normative concerns and legal traditions.

Personality-based claims focus on the unauthorized use of recognizable attributes, such as a person’s name, voice, likeness, or performance. These are governed primarily by the right of publicity and are often triggered when generative AI tools replicate a celebrity’s voice in a song, simulate an actor’s face in a video game, or generate deepfakes that misappropriate a public persona. These harms are dignitary, concerned with false endorsement, reputational distortion, or losing control over one’s identity as a commercial asset. Remedies typically emphasize attribution, consent, and compensation.

By contrast, behavioral or informational control claims arise from collecting, aggregating, or repurposing personal data, including clickstreams, location histories, browsing behavior, and social media interactions, often scraped or inferred without the individual’s awareness. These practices implicate information privacy law and raise concerns about autonomy, manipulation, and surveillance. Unlike personality claims, the individual need not be famous or publicly recognizable; the harm lies in losing control over one’s digital self, not public misrepresentation. The regulatory response must prioritize transparency, purpose limitation, and control over downstream use.

Framing the data right without parsing these distinctions risks doctrinal confusion and overbreadth. A personality-based claim may warrant strong exclusivity and anti-impersonation protection, while behavioral data governance may call for softer access, consent, and erasure rights. Therefore, a well-calibrated data right must navigate this duality: protecting publicly recognizable identities from misappropriation while empowering ordinary individuals to control how their behavioral data is collected, shared, and used in AI training. This distinction is not merely semantic; it determines which legal safety measures are appropriate and what harm the law seeks to prevent.

To address the diverse harms stemming from generative AI’s use of personal data, the proposed data right should therefore adopt a tiered or dual-track structure that reflects the distinct normative bases underlying personality-based and behavioral claims. A bifurcated statutory design could preserve the expressive dimensions of identity while empowering individuals to control the economic and informational uses of their behavioral data. Such a framework would respect the divergent logics of harm and remedy, while providing courts and regulators with more administrable categories for adjudication and enforcement.

Policymakers must also be cognizant that regulation involves balancing tradeoffs and knowing when not to intervene. Overregulating AI markets in their formative stages may weaken U.S. global leadership. Increased compliance burdens and legal uncertainty disproportionately affect small firms and researchers, driving talent and investment abroad. For instance, Europe’s digital economy is now grappling with regulatory saturation. Between 2019 and 2024, the number of EU digital regulations more than doubled, from forty-five to ninety-eight, leading to overlapping rules and complex compliance burdens, especially for small businesses.¹⁷⁷

In this environment, the EU AI Act has sparked concerns that its stringent regulatory framework may exacerbate Europe’s ongoing economic malaise and stifle AI innovation across the region.¹⁷⁸ Rather than promoting certainty, it produced fragmentation, fatigue, and hesitancy among firms to launch or expand digital products in the European Union. For instance, Meta decided to hold back the launch of its newest AI technologies due to regulatory uncertainty.¹⁷⁹ European executives called for accelerated deregulation at the 2025 World Economic Forum in Davos to boost competitiveness in AI and digital markets.¹⁸⁰

In contrast to the European Union, the United States has prioritized growth and competitiveness by driving advancements in computer-related technologies, including AI, cloud computing, and data-driven applications. The U.S. approach emphasizes market-driven innovation, where regulatory frameworks aim to create an environment that encourages risk-taking and rapid technological advancement.¹⁸¹ At the Paris AI summit in February 2025, U.S. Vice President JD Vance warned that excessive regulation would entrench Big Tech incumbents and suppress startup innovation.¹⁸² Vance argued that heavy-handed mandates like the Biden AI executive order risk creating a “Big Tech–government duopoly” while enabling authoritarian regimes like China’s to export censorship-aligned AI models.¹⁸³ He emphasized that a permissive U.S. regulatory environment has encouraged “unparalleled R and D investments” and must be preserved to keep America the “partner of choice” in global AI markets.¹⁸⁴ Vance dismissed fears that AI would inevitably automate away the American workforce, stating instead that most AI applications “involve supplementing, not replacing” human labor.¹⁸⁵ He characterized AI as a driver of productivity and potential backbone for a “new industrial revolution” that could benefit workers provided innovation is not stifled by preemptive regulation.¹⁸⁶

At that same Paris Summit, OpenAI declared that “the race for AI is effectively over” if training on copyrighted material is not deemed fair use.¹⁸⁷ In policy filings to the Trump administration, the company argued that restricting such use would allow Chinese firms, including DeepSeek, to pull ahead by accessing the same types of content that U.S. firms would be barred from using.¹⁸⁸ Framing the issue as one of “freedom to learn,” OpenAI claimed that AI model training merely extracts “patterns, linguistic structures, and contextual insights” from creative works to generate novel outputs, activities that align with the core goals of copyright law.¹⁸⁹

Regulatory coordination can help prevent interdoctrinal misfires between antitrust, copyright, publicity, and privacy laws. Singapore’s legitimate interest and business improvement bases for processing personal data illustrate how privacy law can flexibly accommodate AI development goals, provided procedural safeguards (e.g., anonymization, balancing tests) are in place. For example, Singapore’s Trusted Data Sharing Framework and Privacy Enhancing Technology (“PET”) Sandboxes create structured environments for experimentation and interagency interpretation of legal duties, particularly where competition goals (e.g., expanding access to data) intersect with privacy constraints.¹⁹⁰ These initiatives function as regulatory tuning forks, facilitating innovation while preventing data governance from becoming an antitrust or privacy impasse. Built-in calibration tools could minimize friction with competition objectives that may demand access to personal data for model training or product improvement.

This example of regulatory coordination highlights that polyphony need not entail discord. Congress and the Executive Branch should consider establishing a cross-agency task force on generative AI governance, drawing representation from the FTC, U.S. Copyright Office, National Telecommunications and Information Administration, and other relevant entities such as the Department of Commerce and the National Institute of Standards and Technology. This task force would facilitate shared jurisdictional understanding, promote interoperable regulatory standards, and prevent doctrinal redundancy or contradiction. It would also provide a platform for developing shared principles and experimental governance mechanisms, such as AI sandboxes, data trusts, and pre-certification regimes, that reflect the joint values of innovation, dignity, and competition.¹⁹¹

Without such coordination, the polyphony of legal regimes risks devolving into regulatory cacophony. Doctrinal misfires, such as antitrust interventions that duplicate copyright protections or privacy rules that inadvertently entrench incumbents, may lead to legal uncertainty, overcompliance, and chilling innovation. Similarly, raw data that is neither copyrighted nor personal, such as synthetic data, factual compilations, or aggregate behavioral patterns, may escape legal protection altogether.

The polyphonic model offers a means of doctrinal orchestration, where each regime retains its voice but engages in structured dialogue, enabling more holistic responses to AI’s multifaceted impact. Rather than doctrinal hierarchy or convergence, structured pluralism offers a more resilient foundation for generative AI governance. Through procedural harmonization, regulators can preserve the analytical integrity of each domain while advancing shared objectives like accountability, innovation, and dignity. Whether through regulatory sandboxes, flexible licensing regimes, or dual-path legal provisions, these approaches embody a concerted governance model responsive to generative AI’s multidimensional challenges.

Conclusion

The regulation of generative AI must meet the moment with a framework that reflects both the multidimensionality of the technology and the pluralism of our legal values. Each examined regime—antitrust, copyright, publicity, and privacy—carries a normative legacy, from protecting markets and creativity to safeguarding dignity and autonomy. No doctrine suffices in the face of developer hegemony, synthetic content, predictive modeling, and personalized manipulation. Like Diaz’s Trust, the regulatory project involves reconciling multiple, sometimes contradictory narratives.

A polyphonic approach to AI governance embraces this complexity by flattening differences and composing coherence through structured regulatory pluralism. Antitrust can ensure competitive access to inputs and platforms; copyright can balance innovation and expression; publicity rights can defend against identity appropriation; and privacy law can restore agency over behavioral data. If carefully orchestrated, their intersections offer the contours of a just regulatory response.

But coherence will not emerge organically. Without proactive coordination through cross-agency task forces, regulatory sandboxes, and interdoctrinal dialogue, we risk a world of overlapping mandates, doctrinal gaps, and chilling uncertainty. Generative AI, if governed well, can enhance access to information, empower creators, and broaden opportunity. If governed poorly, it may entrench monopolies, erode trust, and displace human agency. The stakes are not only technological, but they are also democratic. In the end, polyphonic regulation aims to reconcile law with innovation and to align innovation with a regulatory system that is truly AI-ready.

Category: Print Issues

Polyphonic AI Regulation

Share this article