Valuing the Foundation: Frameworks for Fair Creator Compensation in the Generative AI Era

Introduction

The advent of generative artificial intelligence (AI) represents a technological inflection point, driven by models trained on vast quantities of digital content. This unprecedented scale of data ingestion has created a fundamental conflict with the principles of intellectual property law, which are designed to incentivize and protect human creativity.¹ The core of this tension lies in the dual nature of generative AI: it is simultaneously a powerful engine for innovation and a potential existential threat to the creators whose work forms its foundation.³ For many artists, writers, and publishers, the ability of AI to generate content that can substitute for, and thereby devalue, their original creations poses a significant economic risk.¹

This report navigates the complex landscape of creator compensation in the AI era. It examines the central question of how to structure a system that fosters technological progress without enabling it at the expense of the human creativity it relies upon. The analysis begins by dissecting the current legal quagmire surrounding copyright and the “fair use” doctrine, which has become the primary battleground for this conflict. It then explores the burgeoning market-based solutions, such as direct and collective licensing, that have emerged in response to this legal uncertainty. The report subsequently evaluates the divergent regulatory frameworks being developed in the United States, the European Union, and other key jurisdictions. It also assesses the critical role of technical underpinnings, including content provenance and attribution technologies, which are essential for any fair compensation system to function. By drawing lessons from analogous technological disruptions in the music and news industries, this report culminates in a set of integrated, forward-looking recommendations for policymakers, AI developers, and creators, outlining a path from an adversarial posture to a sustainable partnership.

Section I: The Copyright Crucible: Fair Use in the Age of AI

The ambiguity of existing copyright law, particularly the doctrine of “fair use,” is the primary catalyst for the current conflict over creator compensation. This legal uncertainty has simultaneously fueled litigation and incentivized the development of the very market-based solutions now taking shape. Understanding this legal framework is essential to grasping the dynamics of the entire compensation debate.

Deconstructing the Fair Use Doctrine in the AI Context

The fair use doctrine, codified in Section 107 of the U.S. Copyright Act, allows for the limited use of copyrighted materials without permission from the rights holder. Its application is determined through a case-by-case analysis of four statutory factors ⁵:

The purpose and character of the use, including whether it is commercial or for nonprofit educational purposes.
The nature of the copyrighted work.
The amount and substantiality of the portion used in relation to the work as a whole.
The effect of the use upon the potential market for or value of the copyrighted work.

In recent years, courts have placed significant emphasis on whether a use is “transformative” under the first factor. A use is considered transformative if it adds new expression or meaning, or puts the work to a different purpose.⁶ AI developers argue that training models is an archetypal transformative use; the objective is not to reproduce or republish the input data but to learn statistical patterns from it to generate entirely new outputs.⁶ This argument has found a receptive audience in some courts, which have unequivocally found the use of copyrighted works for AI training to be “highly transformative”.⁸ In one notable case, a judge described the use of large language model (LLM) technology as “among the most transformative we will see in our lifetimes”.⁹

This broad judicial interpretation contrasts with the more measured guidance from the U.S. Copyright Office (USCO). The USCO posits that transformativeness is a “matter of degree” that ultimately “depend[s] on the functionality of the model and how it is deployed”.⁷ According to this view, training a model for a non-substitutive task, such as academic research within a closed system, is highly transformative. Conversely, training a model to generate expressive content that directly competes with the creative intent of the original works is far less so and may be considered derivative rather than transformative.⁷

Judicial Precedents: A Fractured and Fact-Specific Landscape

The ongoing wave of copyright litigation has produced a fractured legal landscape where outcomes are highly dependent on the specific facts of each case, making it difficult to establish a universal rule.⁸

A landmark victory for copyright holders came in Thomson Reuters v. Ross Intelligence. In this case, the court rejected the defendant’s fair use defense, finding that Ross Intelligence had used copyrighted Westlaw headnotes to train a competing legal research tool. The court ruled that this use was commercial, not transformative, and was intended to create a “market substitute” for Westlaw’s product.¹¹ Critically, the court’s analysis of the fourth fair use factor—market effect—recognized the existence of a “potential derivative market of ‘data to train legal AI models'”.¹² This finding provides a crucial legal underpinning for the licensing markets that are now emerging, as it validates the idea that rights holders possess a licensable interest in the use of their works for AI training.

In contrast, cases like Bartz v. Anthropic and Kadrey v. Meta have produced more favorable outcomes for AI developers, at least in part. In both instances, courts found the process of training an LLM to be “highly transformative”.⁵ However, these decisions also revealed deep divisions on other critical aspects of the fair use analysis. A significant point of divergence is the provenance of the training data. In

Bartz, Judge William Alsup established a clear distinction, ruling that using legally purchased books for training constituted fair use, while using pirated books did not. He gave “dispositive weight” to the illegal acquisition, calling piracy “inherently, irredeemably infringing”.⁹ This ruling suggests the emergence of a “provenance principle,” where the lawfulness of data acquisition may be a threshold question for any fair use defense.

The courts also split on the assessment of market harm. In Bartz, the court dismissed the plaintiffs’ concerns about market dilution from AI-generated competitors, comparing it to “complain[ing] that training schoolchildren to write well”.⁹ Conversely, in

Kadrey, Judge Vince Chhabria was more sympathetic, endorsing the concept of “indirect market substitution.” He argued that LLMs possess a unique ability to rapidly “flood the market” with content that serves as a substitute for original works, a factor that should be considered in the market harm analysis.⁸

Despite their differences, the judges in Bartz and Kadrey agreed on one point: they rejected the argument that the market for licensing content to LLM developers should be considered when evaluating market harm. They reasoned that this argument is circular, as it presupposes that the activity in question—AI training—is not fair use and requires a license in the first place.⁸ This creates a legal paradox: while the

Thomson Reuters case recognized a potential licensing market, other courts are hesitant to factor that market into their fair use analysis.

This very legal stalemate has become a powerful economic driver. The combination of courts affirming the “transformative” nature of AI training—a strong defense for AI companies—while simultaneously acknowledging the potential for severe “market harm”—a potent weapon for rights holders—has resulted in profound legal uncertainty. Litigation is costly and outcomes are unpredictable.⁸ Faced with this risk, many AI developers have concluded that the price of a license is preferable to the cost of a lawsuit, prompting them to voluntarily enter the market and negotiate deals.¹⁵ Thus, the ambiguity of the fair use doctrine has not prevented compensation; it has become the primary incentive for the creation of a private, market-based compensation system.

The Ethical Undercurrents of the Legal Debate

The legal arguments over fair use are deeply intertwined with the ethical concerns of creators. The legal concept of “market harm” is a direct reflection of creators’ fears of financial devaluation and the potential for their livelihoods to be undermined by AI-generated substitutes.² Many creators and their advocates argue that the “wholesale reproduction of copyrighted works” for training is a form of theft, regardless of whether the final AI output is a direct copy.² From this perspective, the act of copying an entire work without permission—a common practice in AI training and a point of analysis under the third fair use factor—is an inherent violation. The term “AI data laundering” has emerged to describe the process by which the transformative nature of AI training is used to “clean” copyrighted content of its protected status, effectively bypassing traditional licensing and compensation mechanisms.¹⁷ These ethical arguments underscore the belief that the value exchange in the AI ecosystem is fundamentally imbalanced, benefiting technology developers at the direct expense of content creators.

Section II: The Rise of the Data Licensing Economy

In response to the legal pressures and ethical imperatives, a dynamic, industry-led market for data licensing has rapidly emerged. This movement toward creating formal compensation structures is occurring even in the absence of definitive legal or legislative mandates, demonstrating a pragmatic shift from courtroom battles to commercial negotiations. This new data economy is evolving along two parallel tracks: large-scale direct deals between industry giants and more scalable solutions designed for the broader creator community.

Direct Voluntary Licensing: The Billion-Dollar Handshake

The most visible manifestation of this new economy is the burgeoning market for direct licensing agreements. In these deals, AI developers pay substantial sums to publishers and other large content holders for the legal right to use their archives for training models.¹⁸ This market is already valued in the billions and is projected to grow significantly.¹⁹

The players involved are the titans of their respective industries. On one side are AI leaders like OpenAI, Google, Microsoft, and Meta. On the other are major content publishers such as News Corp, The New York Times, the Associated Press (AP), and Axel Springer.³ The financial terms are often significant, with reports of OpenAI offering publishers between $1 million and $5 million per year, and its deal with News Corp valued at over $250 million over five years.²⁰

These agreements represent more than a simple cash-for-data transaction. An analysis of their common terms reveals a more complex and strategic value exchange. “Compensation” is being defined not just by licensing fees but also by non-monetary benefits that are crucial for publishers to remain relevant in an AI-driven world. Key provisions frequently include:

Content for Training: AI companies gain licensed access to high-quality, often paywalled, content archives.²¹
Attribution and Traffic: Publishers secure promises that their content, when used to generate answers in chatbots like ChatGPT, will receive clear attribution and links back to their websites, a vital mechanism to drive traffic and maintain brand presence.²⁰
Access to Technology: Publishers often gain access to the AI company’s technology and tools, enabling them to develop their own AI-powered products and features, thus becoming participants in the new ecosystem rather than just suppliers.²⁰

The motivations are clear and complementary. AI companies mitigate legal risk and gain access to reliable, structured data, which is far superior to indiscriminately scraped web content.¹⁵ Publishers, facing declining revenue from traditional search and advertising, secure a significant new income stream and a strategic partnership that provides them with the tools to innovate.²⁰

Table 1: The Evolving Landscape of Direct AI Content Licensing Agreements

AI Company	Publisher/Content Partner	Date Announced	Known/Reported Financial Terms	Key Provisions	Source(s)
OpenAI	News Corp	May 2024	Over $250 million over five years	Content from major news brands for training and display in ChatGPT with attribution. News Corp gains access to OpenAI tech.	²⁰
OpenAI	Axel Springer	Dec 2023	Not disclosed	Content from Politico, Business Insider, etc., summarized in ChatGPT with attribution; content used for training.	¹⁵
OpenAI	Associated Press (AP)	July 2023	Not disclosed	Use of AP’s text archive for training; AP gains access to OpenAI tech.	¹⁵
OpenAI	The Guardian	Feb 2025	Not disclosed	Compensation for use of journalism in ChatGPT; Guardian gains access to OpenAI tech.	²⁰
OpenAI	Dotdash Meredith	May 2024	Reportedly at least $16 million	Content from 40+ titles for training and display; OpenAI tech to improve ad-targeting tool D/Cipher.	²⁰
Microsoft	Axel Springer	April 2024	Not disclosed	Partner to develop new AI-driven chat experiences; expands adtech collaboration.	²⁰
Microsoft	Informa	May 2024	Over $10 million in first year	Access to B2B publisher’s data until 2027 to improve AI systems.	²⁰
Google	Reddit	Feb 2024	Reportedly ~$60 million per year	Use of Reddit content for training AI tools.	²⁰
Amazon	The New York Times	May 2025	Reportedly $20-25 million per year	Real-time display of summaries in Amazon products (e.g., Alexa) and use for training Amazon models.	²⁰
Meta	Reuters	Oct 2024	Not disclosed	Real-time Reuters content used in Meta’s AI chatbot to answer news queries with summaries and links.	²⁰
Perplexity	Gannett	July 2025	Not disclosed	Content from USA Today and 200+ local brands to appear in AI search answers; ad revenue sharing and tech access for Gannett.	²⁰
Prorata.ai	500+ Publications	June 2025	50% revenue share	Content used in Gist.ai search engine; proportional compensation allocated to publishers based on use.	¹⁹

Collective Licensing: A Scalable Solution for the Long Tail

While direct deals are effective for major players, they are impractical for the millions of independent creators and smaller publishers who lack the resources and leverage to negotiate individually. To address this gap, collective licensing models are emerging as a scalable solution.¹⁹

The most prominent example is the UK’s initiative to develop a Generative AI Training Licence. Spearheaded by the Copyright Licensing Agency (CLA) in partnership with organizations representing authors (ALCS) and publishers (PLS), this framework aims to create a centralized marketplace.²² AI developers would pay a fee to the collective for a license to use a broad catalog of works, with the revenue then distributed to the participating rights holders.²⁴ This model is designed to complement, not replace, direct deals, providing an accessible pathway for those “not in a position to negotiate” them.²² A crucial element of this debate is the mechanism for consent. Creator groups have strongly criticized “opt-out” systems, where consent is assumed unless a creator takes action to withdraw their work, arguing it places an unfair burden on them.²⁵ In response, the developing UK framework is leaning toward an “opt-in” model, where creators must affirmatively choose to include their works in the licensing pool, giving them greater control.²⁴ Other collective initiatives, such as those by the Copyright Clearance Center (CCC) in the US and intermediaries like ProRata.ai, are also building platforms to streamline licensing for a wider range of creators.¹⁹

Micro-Compensation and Creator-Centric Platforms

A third tier of compensation models is being pioneered by technology companies whose platforms are built on contributions from a large base of individual creators. These companies are implementing systems to directly reward creators for the value their content provides to AI training.

Adobe has integrated an “AI training bonus” into its Adobe Stock contributor program. The bonus is paid annually and is weighted based on the number of licenses a creator’s content has generated and the total number of approved images they have submitted, linking compensation to both value and volume.²⁷
Canva has established a $200 million fund to compensate creators who allow their content to be used for AI training. Payments are based on a mix of factors, including the total amount of content contributed and how often each piece is used.²⁷
Stability AI, for its audio generation tool, has partnered with the stock audio company Audiosparx to create an opt-in, revenue-sharing model, echoing proposals from the music industry for more direct royalty splits.²⁷

These various market-driven approaches are creating a two-tiered compensation ecosystem. Large, institutional rightsholders are securing high-value, bespoke deals that provide guaranteed revenue and strategic advantages. Meanwhile, individual creators and smaller publishers are increasingly being directed toward platform-based or collective systems where compensation is often determined by algorithmic formulas or standardized rates. This structure is economically efficient for AI developers, allowing them to manage risk with major players through direct negotiation while addressing the “long tail” of creators through scalable, lower-cost solutions. However, it also creates a potential power imbalance, mirroring dynamics seen in other digital media markets where the value of individual creative works is set by platforms rather than by the creators themselves.

Section III: The Regulatory Compass: Navigating Global Policy Frameworks

While market forces are rapidly shaping compensation practices, governments and regulatory bodies worldwide are simultaneously developing policies to govern the use of copyrighted material in AI. These top-down interventions are creating a complex and divergent global landscape, forcing AI developers to navigate a patchwork of legal requirements that reflect fundamentally different philosophies on how to balance innovation with creator rights.

The United States: A Common Law and Guidance-Based Approach

The United States has thus far adopted a cautious approach, relying on existing copyright law, judicial interpretation, and administrative guidance rather than sweeping new legislation. The U.S. Copyright Office (USCO) has positioned itself as a central interpreter in this domain, launching a comprehensive AI initiative that includes public consultations and a multi-part report on the key legal issues.²⁸

A cornerstone of the USCO’s position is the “human authorship” requirement. The Office has consistently affirmed that copyright protection extends only to works created by humans and that purely AI-generated content is not eligible for registration.⁵ This stance means that while a human can copyright their creative arrangement or modification of AI-generated material, they cannot copyright the raw output of the machine itself.²⁹ The USCO has explicitly stated that, with current technology, user prompts alone do not provide sufficient creative control to be considered authorship.²⁹

Regarding the crucial issue of AI training, the USCO’s guidance suggests that the legality hinges on a case-by-case fair use analysis.¹⁰ Significantly, the Office has signaled a preference for allowing voluntary, market-based licensing solutions to evolve, expressing reservations about imposing compulsory or extended collective licensing schemes at this time.⁷

In parallel, the U.S. Congress is considering several legislative proposals aimed at creating clearer rules. The most impactful of these is the bipartisan AI Accountability and Personal Data Protection Act. This bill would establish a new federal cause of action against the use of copyrighted works for training commercial AI systems without the “express, prior consent” of the rights holder.³⁴ Such a requirement would effectively override the fair use defense in many scenarios, fundamentally altering the legal landscape for AI developers.³⁵ Other proposed bills, like the

TRAIN Act and the Generative AI Copyright Disclosure Act, focus on transparency, seeking to compel AI companies to disclose the datasets used to train their models.³⁵

The European Union: A Regulatory, Rights-Based Framework

In contrast to the U.S., the European Union has pursued a comprehensive, top-down regulatory approach with its landmark AI Act. This legislation imposes specific obligations on providers of General-Purpose AI (GPAI) models concerning copyright.³⁸ The two central requirements are a transparency mandate and respect for existing copyright law. GPAI providers must “publish a sufficiently detailed summary about the content used for training” their models.³⁸ Furthermore, they are required to implement policies to respect EU copyright law, which includes honoring the right of creators to “opt out” of having their works used for text and data mining (TDM) as established under the EU Copyright Directive.⁴⁰

The practical implementation of these rules presents significant challenges, particularly the high transaction costs associated with tracking and respecting opt-outs from millions of rights holders.⁴² To address this, the EU’s Code of Practice for the AI Act attempts to soften the transparency requirement, suggesting that a summary of only the “most relevant” data domains may suffice.⁴² While this pragmatic adjustment aims to make compliance more feasible, it introduces its own issues, such as the potential for training datasets to become biased against smaller publishers and minority language content that may not be deemed “most relevant”.⁴²

Other Jurisdictions: A Divergent Global Landscape

The approaches in the U.S. and EU are not the only models. Other nations are charting their own courses, creating a fragmented global regulatory environment.

China has emerged as a notable counterpoint to the U.S. on the issue of authorship. Chinese courts have granted copyright protection to AI-generated images in cases where a human user’s creative input—through the crafting of prompts and subsequent modifications—was deemed sufficient to meet the standard of originality.⁴³ This establishes a legal precedent that recognizes the creative act in guiding an AI tool, a position the U.S. Copyright Office has so far rejected.
The United Kingdom is pursuing a hybrid strategy. It is actively encouraging industry-led solutions, most notably the development of a collective licensing framework, while simultaneously leaving the door open for future legislation and grappling with the legal legacy of its former EU membership.⁴⁵

This global divergence is creating significant strategic challenges for AI companies operating internationally.⁴⁷ The lack of a harmonized approach means that the same act of training an AI model may be considered fair use in one country, require compliance with transparency and opt-out rules in another, and be subject to different standards of authorship in a third. This fragmentation incentivizes a form of “regulatory arbitrage,” where companies may choose to conduct data-intensive training operations in jurisdictions with the most permissive copyright laws to gain a competitive advantage, potentially undermining the policy goals of stricter regimes.

Furthermore, a chasm is opening between the direction of the market and the focus of some proposed legislation. While the industry is rapidly building a transactional ecosystem based on compensation through licensing, some legislative proposals, particularly in the U.S., are centered on a more rigid standard of consent. A consent-based regime, which grants rights holders an absolute veto over the use of their work, is fundamentally different from a compensation regime, which ensures payment for use. A shift in law from a framework that allows for compensation to one that requires explicit prior consent could upend the burgeoning licensing market, making it practically impossible to clear the rights for the massive datasets required for AI training.

Section IV: Technical Underpinnings for a Fair Ecosystem

Legal frameworks and market-based agreements for creator compensation are ultimately theoretical without a robust technical foundation to make them enforceable and scalable. The ability to track a piece of content from its origin through the AI training pipeline to its potential influence on a new output is a prerequisite for any fair system. Several key technologies are emerging to address this challenge, each with its own potential and limitations.

Content Provenance and Authenticity: The C2PA Standard

A critical enabling technology is the development of open standards for content provenance. The Coalition for Content Provenance and Authenticity (C2PA), an alliance of major technology and media companies including Adobe, Microsoft, Intel, and the BBC, is leading this effort.⁵⁰ The C2PA has developed a technical standard for “Content Credentials,” which function as a tamper-evident “nutrition label” for digital assets.⁵²

This standard allows a creator or publisher to embed secure metadata directly into a file at the moment of its creation or editing. This metadata, which is cryptographically signed using standard X.509 certificates to ensure its integrity, can contain information about the content’s creator, the tools used (including generative AI), and a log of any modifications.⁵⁰ The result is a verifiable, machine-readable history that travels with the asset, allowing anyone to inspect its provenance.⁵⁶ While these credentials can be stripped from a file, the act of removal itself can be a red flag, and the manifest can be designed to show that information has been deleted.⁵⁸ It is important to note that C2PA verifies the history of a file but does not make a judgment about its truthfulness.⁵⁶

Digital Watermarking: Embedding an Invisible Signature

Digital watermarking offers another method for identifying AI-generated content. This technique involves embedding an imperceptible but algorithmically detectable signal directly into the output of an AI model.⁵⁹ Leading technology companies are actively developing and deploying these systems. Google’s SynthID, for example, can embed watermarks in images, audio, text, and video that are invisible to humans but can be identified by a corresponding detector.⁶¹

However, the effectiveness of watermarking is a subject of intense debate. A successful watermark must balance three competing properties: robustness (the ability to survive edits and transformations), fidelity (imperceptibility to humans), and capacity (the amount of information it can carry).⁶² Critics, including the Electronic Frontier Foundation (EFF), argue that many watermarking techniques are fragile and can be easily defeated, either unintentionally through common actions like cropping or re-compressing an image, or intentionally through adversarial attacks designed to remove the signal.⁶³ The debate over watermarking is not merely technical; it reflects a deeper philosophical division about the future of the AI ecosystem. Proponents envision a more controlled, auditable environment where content is traceable and accountable.⁶² Skeptics favor a more open, permissionless ecosystem, arguing that robust watermarking is technically infeasible at scale and could stifle innovation, particularly in the open-source community.⁶³

Attribution and Tracking Systems: Connecting Output to Input

The most significant technical hurdle is attribution: deterministically linking a specific AI-generated output back to the individual pieces of training data that most influenced its creation. Given that models are trained on billions of data points, this is an exceptionally difficult problem.⁶⁵

Despite the complexity, innovative solutions are emerging. The company Bria.ai has developed an attribution technology that assigns a unique identifier to each piece of training content. When a new work is generated, an “Attribution Agent” creates an irreversible vector that links the output to the original works in the dataset based on a “relevance score.” This allows Bria to implement a revenue-sharing model where compensation is distributed to the creators of the most influential source works.⁶⁷

This approach has parallels in the field of digital marketing, where AI-driven attribution models are used to analyze complex, multi-touchpoint customer journeys and assign credit for a conversion to various marketing efforts.⁶⁸ These systems demonstrate that AI itself can be a powerful tool for solving complex attribution problems, suggesting a potential path forward for copyright attribution.

These technologies are not mutually exclusive; in fact, they are symbiotic. A robust system for fair compensation requires a layered technical approach. Content provenance standards like C2PA provide the foundational layer, establishing verifiable ownership of content before it enters a training dataset. Without this, any subsequent attribution is built on uncertain ground. Once provenance is established, internal attribution engines can then track the influence of this verifiably-owned content on new AI outputs, enabling compensation models to function with integrity. One technology establishes ownership, while the other tracks influence, creating an end-to-end chain of custody for creative work.

Overarching Technical Challenges in Data Tracking

Implementing any of these solutions at a global scale faces immense practical challenges. The sheer volume of data involved requires massive computational power and storage capacity.⁷¹ Ensuring the quality, consistency, and security of these vast datasets is a monumental task, with risks including data poisoning, where malicious actors intentionally corrupt training data, and privacy violations from the inadvertent inclusion of sensitive information.⁷³ Furthermore, the “black box” nature of many advanced AI models, where even their developers cannot fully explain the internal logic behind a specific output, presents a fundamental obstacle to true mechanistic interpretability and attribution.⁷⁵

Section V: Lessons from the Past, Models for the Future

The challenges posed by generative AI to creator compensation, while technologically novel, are not without historical precedent. The digital disruptions in the music and news industries over the past two decades offer powerful case studies and cautionary tales. Analyzing these earlier transformations provides a valuable lens through which to evaluate potential compensation frameworks for the AI era.

Case Study: The Music Streaming Revolution

The music industry’s shift from a model based on ownership (selling CDs and digital downloads) to one based on access (monthly streaming subscriptions) fundamentally rewired its economic structure and compensation mechanisms.⁷⁶ This transition has been dominated by a debate between two primary royalty distribution models, a framework that is directly applicable to the challenge of compensating for AI training data.

The prevailing model is the “pro-rata” system. Under this model, all revenue from subscriptions is collected into a single pool, and royalties are distributed to artists based on their percentage of the total streams on the platform.⁷⁶ This system has been widely criticized for disproportionately benefiting a small number of global superstars who command the majority of streams, while leaving many niche and independent artists with fractions of a cent per stream and making it nearly impossible to earn a living wage from streaming alone.⁷⁹

An alternative, long proposed by creator advocates, is the “user-centric” payment system (UCPS). In a user-centric model, an individual subscriber’s monthly fee is distributed only among the artists that specific user actually listened to.⁷⁶ This approach is seen as more equitable, as it directly links a listener’s payment to the artists they value, potentially increasing payouts for artists with dedicated but smaller fan bases.⁸¹

This debate provides a clear blueprint for evaluating AI compensation. A simplistic model that compensates creators based on the raw volume or frequency of their work in a massive, undifferentiated training dataset would essentially replicate the pro-rata system, likely creating a similar “superstar economy” that benefits large, mainstream publishers at the expense of smaller, specialized creators. A more equitable approach would mirror the user-centric philosophy, aiming to measure and reward the actual influence or relevance of a specific piece of training data on valuable AI outputs. This is precisely the direction that advanced attribution technologies, like the one developed by Bria.ai, are heading, moving beyond simple presence to measured contribution.

Case Study: The News Aggregation Wars

The trajectory of the relationship between news publishers and online news aggregators offers a predictive roadmap for the current AI-copyright conflict. In the early 2000s, with the rise of platforms like Google News and The Huffington Post, publishers leveled accusations of “theft” and “free-riding,” arguing that aggregators were profiting from their costly journalistic efforts without compensation.⁸² This mirrors the rhetoric used by many creator groups against AI developers today.

The ensuing legal battles were often inconclusive, mired in the same complex, fact-specific arguments over fair use and related doctrines like “hot news misappropriation” that characterize the current AI litigation.⁸² Despite the legal uncertainty, the market eventually found an equilibrium. Publishers and aggregators recognized a degree of symbiosis—aggregators drove traffic, and publishers provided content—and the initial phase of conflict evolved into a more stable, transactional relationship.⁸⁴

This history suggests that the current flurry of lawsuits against AI companies is likely not the final state of affairs, but rather a contentious but necessary period of leverage-building and price discovery. Following the precedent of the news aggregation wars, the probable endgame is not a definitive court ruling that settles the matter for all time, but rather the maturation of the voluntary licensing ecosystem that is already taking shape. Litigation will serve to define the outer boundaries of acceptable practice, while the day-to-day business of compensation will be handled through the direct and collective licensing markets detailed in Section II.

Section VI: Synthesis and Strategic Recommendations

The challenge of fairly compensating creators in the age of generative AI is multifaceted, spanning law, economics, technology, and ethics. No single solution will suffice. The path forward requires a hybrid, multi-layered framework that integrates technical standards, market mechanisms, and regulatory guardrails to create a sustainable and equitable ecosystem for both creators and innovators.

The Inevitability of a Hybrid Framework

A workable system for fair compensation must be built upon three interconnected layers:

A Technical Foundation: This layer is the bedrock. It requires the widespread adoption of open, interoperable standards for content provenance, such as the C2PA’s Content Credentials, to create a verifiable record of authorship and origin. This must be complemented by robust attribution technologies capable of tracing the influence of training data on AI outputs.
A Market-Based Mechanism: Building on this technical foundation, a dynamic, two-tiered licensing economy can thrive. This market will combine large-scale, bespoke direct licensing deals for major rights holders with accessible, efficient collective licensing systems to serve the long tail of independent creators and smaller publishers.
A Legal and Regulatory Backstop: Law and policy should not dictate the market but should set its boundaries. This includes establishing clear legal principles, such as the “Provenance Principle” that penalizes the use of illegally acquired data, and enforcing regulatory guardrails, like the transparency mandates in the EU AI Act, to ensure fairness and provide legal recourse when market mechanisms fail.

Recommendations for Policymakers and Regulators

Prioritize Transparency and Provenance: Rather than attempting to legislate a definitive answer to the fair use question, regulators should focus on mandating technical transparency. This includes supporting the adoption of open standards like C2PA and requiring AI developers to provide meaningful disclosure of the sources used to train their models.
Foster Competitive Licensing Markets: Policy should aim to create the conditions for healthy licensing markets to flourish. This means avoiding overly prescriptive interventions like compulsory licensing, which could stifle market development, while simultaneously applying antitrust oversight to prevent the concentration of power in the hands of a few large AI labs or publisher consortiums.
Harmonize Internationally: Given the global nature of AI development, nations should work through international bodies like the World Intellectual Property Organization (WIPO) to establish a baseline of shared principles on transparency and licensing, reducing the legal friction that encourages regulatory arbitrage.

Recommendations for AI Developers

Embrace Ethical Sourcing as a Competitive Advantage: AI companies should move beyond a defensive, risk-mitigation posture on data acquisition. Proactively investing in ethically sourced, licensed data and provenance-aware data pipelines can become a mark of quality and trustworthiness, differentiating “responsibly trained” models in the marketplace.
Invest in Attribution Technology: Developing robust internal systems for tracking data lineage is not merely a compliance cost but a strategic investment. It is the core infrastructure required for building scalable compensation systems, fostering trust with content partners, and gaining deeper insights into model behavior.
Support Open Standards: To avoid a fragmented ecosystem of proprietary, incompatible solutions, AI developers should actively participate in and contribute to the development of open standards for content provenance (C2PA) and interoperable watermarking and attribution technologies.

Recommendations for Publishers and Creators

Audit and Assert Rights: Creators and their representatives must proactively manage their intellectual property. This involves auditing existing publishing contracts to clarify who holds the rights to license works for AI training and using collective bodies like author and artist guilds to assert these rights with a unified voice.⁸⁵
Embrace New Models and Technologies: Compensation in the AI era will extend beyond simple cash payments. Creators should engage with the full spectrum of value, including access to AI technology, prominent attribution, and performance-based revenue sharing. Adopting C2PA standards for new content is a crucial step to ensure that work is “AI-ready” for future licensing and attribution systems.
Collaborate and Aggregate: Individual negotiations with large AI companies are often impractical for smaller players. Creators and small publishers should leverage the power of collective licensing organizations and other aggregation platforms to increase their bargaining power and ensure their inclusion in the new data economy.

Concluding Thoughts: From Adversaries to Partners

The current relationship between the creative industries and the AI sector is often framed as a zero-sum conflict. However, this perspective is ultimately counterproductive. Fair compensation is not an obstacle to innovation; it is a prerequisite for a sustainable AI ecosystem. The data that powers generative AI is not a raw, unowned commodity like sunlight or air; it is the product of human labor, investment, and creativity. By establishing transparent, efficient, and equitable systems for valuing this foundational data, the industry can create a virtuous cycle. When creators share in the immense value their work generates, they are incentivized to continue producing the high-quality, diverse content that will fuel the next generation of innovation, transforming a relationship of conflict into one of enduring partnership.

Works cited

Generative AI Is a Crisis for Copyright Law – Issues in Science and Technology, accessed September 15, 2025, https://issues.org/generative-ai-copyright-law-crawford-schultz/
Is Copyrighted Material Used by AI? – American Bar Association, accessed September 15, 2025, https://www.americanbar.org/groups/senior_lawyers/resources/voice-of-experience/2024-april/is-copyrighted-material-used-by-ai/
Generative AI Lawsuits Timeline: Legal Cases vs. OpenAI, Microsoft …, accessed September 15, 2025, https://sustainabletechpartner.com/topics/ai/generative-ai-lawsuit-timeline/
Artificial Intelligence Impacts on Copyright Law – RAND, accessed September 15, 2025, https://www.rand.org/pubs/perspectives/PEA3243-1.html
Generative Artificial Intelligence and Copyright Law | Congress.gov …, accessed September 15, 2025, https://www.congress.gov/crs-product/LSB10922
What Is Fair Use? — The Impact of AI on Fair Use – Originality.AI, accessed September 15, 2025, https://originality.ai/blog/fair-use-and-ai
Copyright Office Issues Key Guidance on Fair Use in Generative AI …, accessed September 15, 2025, https://www.wiley.law/alert-Copyright-Office-Issues-Key-Guidance-on-Fair-Use-in-Generative-AI-Training
Fair Use and AI Training: Two Recent Decisions Highlight the …, accessed September 15, 2025, https://www.skadden.com/insights/publications/2025/07/fair-use-and-ai-training
A Tale of Three Cases: How Fair Use Is Playing Out in AI Copyright …, accessed September 15, 2025, https://www.ropesgray.com/en/insights/alerts/2025/07/a-tale-of-three-cases-how-fair-use-is-playing-out-in-ai-copyright-lawsuits
U.S. Copyright Office Issues Guidance on Generative AI Training | Insights | Jones Day, accessed September 15, 2025, https://www.jonesday.com/en/insights/2025/05/us-copyright-office-issues-guidance-on-generative-ai-training
Court Rules AI Training on Copyrighted Works Is Not Fair Use — What It Means for Generative AI – Davis+Gilbert LLP, accessed September 15, 2025, https://www.dglaw.com/court-rules-ai-training-on-copyrighted-works-is-not-fair-use-what-it-means-for-generative-ai/
AI Training Using Copyrighted Works Ruled Not Fair Use, accessed September 15, 2025, https://www.pbwt.com/publications/ai-training-using-copyrighted-works-ruled-not-fair-use
First of its Kind Decision Finds AI Training is Not Fair Use – Copyright Alliance, accessed September 15, 2025, https://copyrightalliance.org/ai-training-not-fair-use/
Copyright and AI: the Cases and the Consequences | Electronic …, accessed September 15, 2025, https://www.eff.org/deeplinks/2025/02/copyright-and-ai-cases-and-consequences
What happens when your publisher licenses your work for AI …, accessed September 15, 2025, https://www.authorsalliance.org/2024/07/30/what-happens-when-your-publisher-licenses-your-work-for-ai-training/
Generative AI, Copyrighted Works, & the Quest for Ethical Training Practices, accessed September 15, 2025, https://copyrightalliance.org/generative-ai-ethical-training-practices/
ETHICAL USE OF AI: NAVIGATING COPYRIGHT CHALLENGES – GLOBSEC, accessed September 15, 2025, https://www.globsec.org/sites/default/files/2024-09/Ethical%20Use%20of%20AI%20-%20Navigating%20Copyright%20Challenges.pdf
Publishers seek compensation and attribution from AI training data – AMEC survey — Wadds Inc. | Professional advisor to agencies & comms teams, accessed September 15, 2025, https://wadds.co.uk/blog/2024/5/1/publishers-seek-compensation-and-attribution-from-ai-training-data
AI Copyright Licensing: Market Solutions to GAI Development | CCC, accessed September 15, 2025, https://www.copyright.com/blog/ai-copyright-licensing-market-solutions-gai-development/
News generative AI deals revealed: Who is suing, who is signing?, accessed September 15, 2025, https://pressgazette.co.uk/platforms/news-publisher-ai-deals-lawsuits-openai-google/
2024 in review: A timeline of the major deals between publishers and AI companies, accessed September 15, 2025, https://digiday.com/media/2024-in-review-a-timeline-of-the-major-deals-between-publishers-and-ai-companies/
CLA announces development of Generative AI Training Licence, accessed September 15, 2025, https://cla.co.uk/development-of-cla-generative-ai-licence/
Generative AI and Copyright | Licensing, Training & More | CLA, accessed September 15, 2025, https://cla.co.uk/ai-and-copyright/
UK’s Collective Licensing Initiative Aims to Harmonize AI and …, accessed September 15, 2025, https://www.ailawandpolicy.com/2025/05/uks-collective-licensing-initiative-aims-to-harmonize-ai-and-copyright-law/
Collective license for AI training – OlarteMoure | Intellectual Property, accessed September 15, 2025, https://olartemoure.com/en/collective-license-for-ai-training/
CCC Launches Collective AI License – Copyright Clearance Center, accessed September 15, 2025, https://www.copyright.com/blog/ccc-launches-collective-ai-license/
How should creators be compensated for their work training AI …, accessed September 15, 2025, https://qz.com/how-should-creators-be-compensated-for-their-work-train-1850932454
Copyright and Artificial Intelligence | U.S. Copyright Office, accessed September 15, 2025, https://www.copyright.gov/ai/
Generative AI in Focus: Copyright Office’s Latest Report: Wiley, accessed September 15, 2025, https://www.wiley.law/alert-Generative-AI-in-Focus-Copyright-Offices-Latest-Report
U.S. Copyright Office issues report on copyrightability of AI assisted and generated works, accessed September 15, 2025, https://www.hoganlovells.com/en/publications/us-copyright-office-issues-report-on-copyrightability-of-ai-assisted-and-generated-works
Copyright Office Publishes Report on Copyrightability of AI-Generated Materials | Insights | Skadden, Arps, Slate, Meagher & Flom LLP, accessed September 15, 2025, https://www.skadden.com/insights/publications/2025/02/copyright-office-publishes-report
US Copyright Office Publishes Second Part of Report on AI Copyrightability — AI: The Washington Report | Mintz, accessed September 15, 2025, https://www.mintz.com/insights-center/viewpoints/54731/2025-02-07-us-copyright-office-publishes-second-part-report-ai
Copyright Office Weighs In on AI Training and Fair Use | Skadden, Arps, Slate, Meagher & Flom LLP, accessed September 15, 2025, https://www.skadden.com/insights/publications/2025/05/copyright-office-report
New Senate bill aims to provide more certainty on rights and …, accessed September 15, 2025, https://www.mcdonaldhopkins.com/insights/news/new-senate-bill-aims-to-provide-more-certainty-on-rights-and-obligations-of-ai-developers
Senators Introduce Legislation to Curb Use of Personal Data and …, accessed September 15, 2025, https://natlawreview.com/article/senators-introduce-legislation-curb-use-personal-data-and-copyrighted-works-gen-0
Legislative Developments – U.S. Copyright Office, accessed September 15, 2025, https://www.copyright.gov/legislation/
AI, Copyright, and the Law: The Ongoing Battle Over Intellectual Property Rights – USC, accessed September 15, 2025, https://sites.usc.edu/iptls/2025/02/04/ai-copyright-and-the-law-the-ongoing-battle-over-intellectual-property-rights/
EU AI Act: first regulation on artificial intelligence | Topics | European …, accessed September 15, 2025, https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence
High-level summary of the AI Act | EU Artificial Intelligence Act, accessed September 15, 2025, https://artificialintelligenceact.eu/high-level-summary/
EU AI Act: How Far Will EU Copyright Principles Extend? – Publications – Morgan Lewis, accessed September 15, 2025, https://www.morganlewis.com/pubs/2024/02/eu-ai-act-how-far-will-eu-copyright-principles-extend
Generative AI & Copyright in the EU: Myths Versus Facts – CCIA, accessed September 15, 2025, https://ccianet.org/articles/generative-ai-copyright-in-the-eu-myths-versus-facts/
The European Union is still caught in an AI copyright bind – Bruegel, accessed September 15, 2025, https://www.bruegel.org/analysis/european-union-still-caught-ai-copyright-bind
Chinese Court Again Rules AI-Generated Images Are Eligible for …, accessed September 15, 2025, https://www.chinaiplawupdate.com/2025/03/chinese-court-again-rules-there-is-copyright-in-ai-generated-images/
Copyright Protection for AI generated works – Recent Developments – Bird & Bird, accessed September 15, 2025, https://www.twobirds.com/en/insights/2024/china/copyright-protection-for-ai-generated-works-recent-developments
No current plans for statutory private copying levy – Arts Professional, accessed September 15, 2025, https://www.artsprofessional.co.uk/news/no-current-plans-for-statutory-private-copying-levy
Copyright and Generative AI | Insights | Mayer Brown, accessed September 15, 2025, https://www.mayerbrown.com/en/insights/publications/2025/06/copyright-and-generative-ai
Generative AI: Navigating Intellectual Property – WIPO, accessed September 15, 2025, https://www.wipo.int/documents/d/frontier-technologies/docs-en-pdf-generative-ai-factsheet.pdf
Will copyright law enable or inhibit generative AI? – The World Economic Forum, accessed September 15, 2025, https://www.weforum.org/stories/2024/01/cracking-the-code-generative-ai-and-intellectual-property/
Global AI Law and Policy Tracker – IAPP, accessed September 15, 2025, https://iapp.org/resources/article/global-ai-legislation-tracker/
C2PA (Coalition for Content Provenance and Authenticity) in Digital Asset Management, accessed September 15, 2025, https://www.orangelogic.com/c2pa-coalition-for-content-provenance-and-authenticity-in-digital-asset-management
About – C2PA, accessed September 15, 2025, https://c2pa.org/about/
Content Provenance and Disclosure Requirements for AI Generated …, accessed September 15, 2025, https://www.globalvoices.org.au/post/content-provenance-and-disclosure-requirements-for-ai-generated-content-on-digital-and-traditional-m
C2PA | Verifying Media Content Sources, accessed September 15, 2025, https://c2pa.org/
“Did ChatGPT really say that?”: Provenance in the age of Generative AI. | Library Innovation Lab – Harvard University, accessed September 15, 2025, https://lil.law.harvard.edu/blog/2023/05/22/provenance-in-the-age-of-generative-ai/
Insights into Coalition for Content Provenance and Authenticity (C2PA) – Infosys, accessed September 15, 2025, https://www.infosys.com/iki/techcompass/content-provenance-authenticity.html
C2PA Explainer :: C2PA Specifications, accessed September 15, 2025, https://spec.c2pa.org/specifications/specifications/1.2/explainer/Explainer.html
How it works – Content Authenticity Initiative, accessed September 15, 2025, https://contentauthenticity.org/how-it-works
What is C2PA? C2PA and Digital Authenticity – YouTube, accessed September 15, 2025, https://www.youtube.com/watch?v=wMnVHeXPb6c
Digital Watermarking Technology for AI-Generated Images: A Survey, accessed September 15, 2025, https://www.mdpi.com/2227-7390/13/4/651
SoK: Watermarking for AI-Generated Content – arXiv, accessed September 15, 2025, https://arxiv.org/html/2411.18479v1
SynthID – Google DeepMind, accessed September 15, 2025, https://deepmind.google/science/synthid/
The Case for and Against AI Watermarking | RAND, accessed September 15, 2025, https://www.rand.org/pubs/commentary/2024/01/the-case-for-and-against-ai-watermarking.html
AI Watermarking Won’t Curb Disinformation | Electronic Frontier …, accessed September 15, 2025, https://www.eff.org/deeplinks/2024/01/ai-watermarking-wont-curb-disinformation
Identifying AI generated content in the digital age: The role of watermarking | EY, accessed September 15, 2025, https://www.ey.com/content/dam/ey-unified-site/ey-com/en-in/insights/ai/documents/ey-identifying-ai-generated-content-in-the-digital-age-the-role-of-watermarking.pdf
Who Owns the Output? Bridging Law and Technology in LLMs Attribution – arXiv, accessed September 15, 2025, https://arxiv.org/html/2504.01032v1
Which Contributions Deserve Credit? Perceptions of Attribution in Human-AI Co-Creation, accessed September 15, 2025, https://arxiv.org/html/2502.18357v1
Advancing Equitable AI Development | Bria’s Attribution Technology, accessed September 15, 2025, https://bria.ai/advancing-equitable-ai-development
AI in Marketing Attribution: Everything You Need to Know – MNTN, accessed September 15, 2025, https://mountain.com/blog/ai-attribution/
AI-driven marketing attribution: What it is, how it works, and why it matters – Usermaven, accessed September 15, 2025, https://usermaven.com/blog/ai-driven-marketing-attribution
Top 10 best AI-powered marketing attribution tools – CMO Alliance, accessed September 15, 2025, https://www.cmoalliance.com/best-ai-powered-marketing-attribution-tools/
6 Common AI Model Training Challenges – Oracle, accessed September 15, 2025, https://www.oracle.com/artificial-intelligence/ai-model-training-challenges/
Top 7 Data Challenges in Generative AI and Solutions for 2025 – RTS Labs, accessed September 15, 2025, https://rtslabs.com/generative-ai-data-challenges
The Hidden Dangers in AI Data Collection and How to Avoid Them …, accessed September 15, 2025, https://medium.com/@tahirbalarabe2/%EF%B8%8Fthe-hidden-dangers-in-ai-data-collection-and-how-to-avoid-them-87ec80e11e53
3 training data challenges hurting AI – Sigma AI, accessed September 15, 2025, https://sigma.ai/ai-training-data-challenges-2/
[2501.18887] Towards Unified Attribution in Explainable AI, Data-Centric AI, and Mechanistic Interpretability – arXiv, accessed September 15, 2025, https://arxiv.org/abs/2501.18887
The Evolution of Music Royalties in the Streaming Era » Flourish …, accessed September 15, 2025, https://flourishprosper.net/music-resources/the-evolution-of-music-royalties-in-the-streaming-era/
The Evolution of Music Royalties: From Physical to Digital, accessed September 15, 2025, https://royaltyexchange.com/blog/the-evolution-of-music-royalties-from-physical-to-digital
Revenue Sharing at Music Streaming Platforms | Management …, accessed September 15, 2025, https://pubsonline.informs.org/doi/10.1287/mnsc.2023.03830
The Effects of Digital Music Streaming on the Revenue Models of …, accessed September 15, 2025, https://ir.library.oregonstate.edu/downloads/9w032897m
(PDF) Consequences of platforms’ remuneration models for digital content: initial evidence and a research agenda for streaming services – ResearchGate, accessed September 15, 2025, https://www.researchgate.net/publication/360638853_Consequences_of_platforms’_remuneration_models_for_digital_content_initial_evidence_and_a_research_agenda_for_streaming_services
Economics of Streaming & the Rise of the Music Artists’ Rights and …, accessed September 15, 2025, https://ipbusinessacademy.org/economics-of-streaming-the-rise-of-the-music-artists-rights-and-compensation
(PDF) The Rise of the News Aggregator: Legal Implications and Best …, accessed September 15, 2025, https://www.researchgate.net/publication/228242270_The_Rise_of_the_News_Aggregator_Legal_Implications_and_Best_Practices
The Rise of News Aggregation and the Race for Information in the …, accessed September 15, 2025, https://www.law.columbia.edu/news/archive/rise-news-aggregation-and-race-information-digital-age
The Rise of Online News Aggregators: Consumption and Competition | Request PDF, accessed September 15, 2025, https://www.researchgate.net/publication/276250104_The_Rise_of_Online_News_Aggregators_Consumption_and_Competition
AI Licensing for Authors: Who Owns the Rights and What’s a Fair …, accessed September 15, 2025, https://authorsguild.org/news/ai-licensing-for-authors-who-owns-the-rights-and-whats-a-fair-split/

How should content creators be fairly compensated in the AI world?