Report on the Ascendancy of Generative AI and the Asymmetric Conflict for Critical Rights

by Gemini 2.5 Pro, Deep Research. Warning, LLMs may hallucinate!

Chapter I: The 2025 Copyright Doctrine: A Fractured Foundation

The legal and regulatory landscape governing artificial intelligence (AI) in 2025 is defined by profound uncertainty, fragmentation, and a series of contradictory judicial rulings. This state of ambiguity is not a neutral condition; it is a strategic terrain that overwhelmingly favors the operational velocity and vast capital resources of Big Tech developers. While rights holders across the creative industries seek clear, defensible lines, the AI industry thrives in the legal “doldrums of discovery,” ¹ leveraging the slow pace of the courts to achieve market saturation before the law is settled. With no fewer than 51 copyright lawsuits pending against AI companies as of October 2025 ¹ and no final summary judgment decisions on the core issue of fair use expected until mid-2026 at the earliest ¹, the definitive legal battle remains perpetually on the horizon.

This protracted uncertainty, which forces a highly fact-specific analysis in every individual case ², allows AI developers to sustain a multi-front legal war that smaller rights holders cannot afford to wage. The current judicial sentiment is best understood as a “Fair Use Triangle,” ¹ a trio of conflicting district court opinions that have left the foundational doctrines of copyright law unstable and ill-equipped for the generative era.

Section 1.1: The “Fair Use” Triangle: Conflicting Judicial Rulings of 2025

The legal architecture for AI training is currently being built on the shifting sands of three core, and conflicting, district court opinions.

First, in Bartz v. Anthropic, Judge William Alsup of the Northern District of California established the cornerstone of Big Tech’s legal strategy. In a June 2025 summary judgment ruling, he found that the act of training an AI model on lawfully acquiredcopyrighted works was “highly,” “spectacularly,” and “quintessentially” transformative.²This ruling, while not binding on other courts, provided the entire AI industry with its primary legal beachhead: the argument that training is a non-expressive, mechanical process of learning that constitutes a protected fair use.

Second, in Kadrey v. Meta, Judge Vince Chhabria, sitting in the same district, provided a stark contrast just two days later.³ While he also granted summary judgment to Meta, it was a procedural victory, not a substantive one. Judge Chhabria openly criticized Judge Alsup’s reasoning, calling his analogy of AI training to “training schoolchildren to write well” “inapt”.³ He found the training use “highly transformative” ³ but emphasized that the plaintiffs—a small group of authors—had simply failed to provide any evidence of market harm.³ He issued a stern warning that his decision “does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful,” but only that “these plaintiffs made the wrong arguments and failed to develop a record in support of the right one”.¹⁰

Third, in Thomson Reuters v. Ross, an earlier ruling from the District of Delaware provides the primary counter-narrative for rights holders.² In this case, the court rejected the AI company’s fair use defense, finding that its use of copyrighted legal headnotes was not transformative and directly harmed the market.¹¹ This case, now on interlocutory appeal ¹, represents the core legal argument that rights holders are advancing: that AI training is nothing more than mass, mechanical reproduction for a commercial purpose that directly usurps the original work’s market.

This judicial fracturing is summarized in Table 1 below.

Table 1: 2025 AI Copyright & Fair Use Rulings (U.S. Federal Courts)

This legal chaos has been further complicated by the rights holders’ own legal maneuvers. In a recent partial dismissal in the Authors Guild v. OpenAI case, Judge Paul Engelmayer’s ruling included language suggesting that for a work to have copyright, it must have “a specific human author whose protectable expression can be identified”.¹⁷This language, intended to bolster the Authors Guild’s claim that only humans can create copyrightable work, has created a bizarre and dangerous legal fissure.

Legal scholar Matthew Sag, analyzing the decision, warned of a potential “copyright winter for Wikipedia”.¹⁷ The logic follows that if a work must have a single, identifiablehuman author, then collectively-authored works like Wikipedia, created by thousands of anonymous volunteers, may not qualify for copyright protection at all.¹⁷ The very lawsuit intended to protect authors from AI scraping may have inadvertently created a legal precedent that strips copyright from one of the world’s largest and most critical training datasets, effectively classifying it as public domain content. This demonstrates a critical strategic failure: rights holders, in their haste to litigate, risk creating unintended legal precedents that benefit Big Tech far more than any courtroom victory could.

Section 1.2: “Transformativeness” vs. “Market Harm”: Deconstructing the Core Legal Arguments

The entire legal war is being fought over the four factors of fair use, as outlined in Section 107 of the Copyright Act.¹⁸ Big Tech has focused its legal and rhetorical strategy almost exclusively on Factor 1: “The purpose and character of the use”.¹⁸ Their core argument is that using a work for AI training is a “non-expressive” use.¹⁹ The AI, they claim, is not “reading” the book for its creative expression but is merely ingesting it to learn statistical patterns of language.¹⁹ This, they argue, is inherently “transformative” and thus a protected fair use.³

Rights holders, including the Authors Guild and the Motion Picture Association (MPA), counter that this is a disingenuous legal fiction. They argue that AI companies seek out high-quality, professionally authored works precisely because of their expressive content. It is the “high-quality, professionally authored” expression that is “vital to enabling an LLM to produce outputs that mimic human language, story structure, character development, and themes”.¹⁹

The U.S. Copyright Office (USCO), in its (non-binding) May 2025 report on generative AI training, largely validated the rights holders’ position.¹⁸ The USCO guidance stated:

• “Transformativeness” Must Be Meaningful, Not Mechanical: The office rejected the idea that any non-human, mechanical use is automatically transformative.¹⁸

• Competing Use is Not Transformative: The USCO stated that training an AI model to produce expressive content that competes with the originals (e.g., training on novels to write new novels) is “at best, modestly transformative”.²⁰

• Market Harm is Central: The report emphasized that Factor 4, “The effect of the use upon the potential market for or value of the copyrighted work,” remains a “central concern”.¹⁸

This guidance, combined with the Kadrey v. Meta ruling, reveals a crucial shift in the legal battlefield. Judges appear increasingly willing to accept the input (the act of training) as transformative, or at least to find the question too complex to dismiss. This has effectively shifted the entire burden of proof to Factor 4: market harm.

Judge Chhabria’s Kadrey opinion is the blueprint for this new reality. He explicitly called Factor 4 “the single most important element”.⁷ He argued that the scale of generative AI—its ability to create “literally millions of secondary works” in a “miniscule fraction of the time”—distinguishes it from any previous technology.³ For this reason, he concluded that future plaintiffs who bring actual evidence of market harm (rather than “mere speculation”) will likely “decisive_ly win the fourth factor_—and thus win the fair use question overall”.³

The strategic implication is profound. Big Tech has successfully established that the actof training is presumptively transformative. They have shifted the burden of proof to rights holders, who no longer win simply by proving their work was copied. Rights holders must now prove specific, quantifiable market harm resulting from the output of the models. This is a much more difficult, expensive, and data-intensive legal test to meet, one that heavily favors the tech incumbents.

Section 1.3: The Getty v. Stability AI UK Ruling: A Blow to Global Copyright Enforcement

While U.S. law remains fractured, the UK High Court delivered a significant, albeit narrow, victory to AI developers in November 2025.²³ In Getty Images v. Stability AI, the London-based AI firm successfully resisted Getty’s core copyright infringement claims.

The ruling was a strategic disaster for Getty. Its primary claim of copyright infringement failed because it could not provide evidence that the training (the infringing act of copying) had taken place in the UK.²³ Stability AI, the defendant, is a UK-based company, but the physical location of its training infrastructure was not established as being within the court’s jurisdiction for that claim.

More devastatingly, the judge, Mrs Justice Joanna Smith, ruled that an AI model like Stable Diffusion, which does not store or reproduce the original images, is not an “infringing copy” under UK law.²³

This ruling effectively green-lights a massive, exploitable jurisdictional loophole that can be described as “copyright laundering.” The case establishes two critical precedents: first, that legal leverage for an input claim is tied exclusively to the physical location of the training servers ²³; and second, that the resulting model is not itself an infringing copy.²³

The logical strategy for any AI developer is now clear: conduct all “infringing” training activities in a jurisdiction with weak or non-existent IP laws (a “data haven”). Once the model is trained, it can be deployed globally, including in the UK, with impunity, as the model itself is “clean.”

This development makes the creative industry’s calls for mandatory transparency the single most important and viable counter-strategy.²³ Without a legally mandated paper trail showing what data was used and where it was trained, it will become jurisdictionally impossible for rights holders to enforce their copyrights against a globally distributed AI industry.

Chapter II: The Anthropic Precedent: How a $1.5 Billion “Loss” Solidified Big Tech’s Victory

The proposed $1.5 billion settlement in Bartz v. Anthropic, announced in 2025, has been widely hailed by author groups as a historic victory.²⁴ This analysis, however, reframes this event. The settlement was not a defeat for Big Tech but a calculated “cost of business”—a one-time “piracy cleanup” fee that successfully neutralized the industry’s single greatest legal liability (past infringement via “shadow libraries”) while simultaneously cementing its most important legal victory (the right to train on legally acquired data).

Section 2.1: Deconstructing the Alsup Ruling: The “Piracy vs. Training” Bifurcation

The pivotal moment of the Bartz case was not the settlement, but Judge Alsup’s June 2025 summary judgment ruling.⁵ In a masterful stroke of judicial reasoning, Judge Alsup bifurcated the case, creating two distinct legal concepts that would define the future of the industry:

1. The Act of Training: He ruled that Anthropic’s use of lawfully acquired books for the purpose of training its AI model was “quintessentially transformative” and therefore a protected fair use.³ This was a monumental, precedent-setting victory for the entire AI industry, providing the legal certainty it had been desperate for.

2. The Act of Acquisition: He simultaneously ruled that Anthropic’s acquisition and retention of millions of books from known piracy sites (”shadow libraries” like LibGen and PiLiMi) was not fair use and constituted infringement.³

This ruling provided the “Anthropic Roadmap” for all AI development.³² Judge Alsup handed the industry its legal playbook: piracy is illegal, but training is not. The strategic imperative for every AI company became, overnight, to “clean up” its data pipeline and begin striking licensing deals.

Before this ruling, the legal status of AI training at all was a massive, existential question mark. Judge Alsup’s decision isolated the “toxic” part of Anthropic’s conduct (the piracy) from the “essential” part (the training). He effectively told the industry, “You can do this, you just have to pay for your inputs, just like any other business.” This ruling, celebrated by the Authors Guild as a win against theft ²⁴, was, in fact, the biggest win Big Tech could have hoped for. It transformed their potential existential legal crisis into a simple business negotiation over licensing costs.

Section 2.2: The Price of Piracy: Calibrating the New Licensing Market

Judge Alsup’s ruling set the stage for the settlement. After his decision, he certified a class action only for the piracy claim.⁵ This was the critical move. Because statutory copyright damages can reach $150,000 per infringed work, Anthropic was suddenly facing a credible, existential threat of hundreds of billions of dollars in potential damages.⁵

Faced with this trial, Anthropic settled for $1.5 billion.¹² This figure represents the largest copyright recovery in history ²⁹, covering an estimated 500,000 copyrighted works ¹² at a rate of approximately $3,000 per work.¹²

This settlement is not a deterrent; it is a market-calibrating event. The $1.5 billion is not a “punishment” for AI; it is a retroactive “piracy cleanup” fee.²⁸ It establishes, for the first time, a market price (approx. $3,000/work) for past, large-scale, willful infringement.²⁹

This legitimizes AI innovation.³⁶ It clears the legal and financial decks, allowing Anthropic and its competitors to move forward. Far from slowing down AI, this settlement accelerates the development of new licensing frameworks.²⁹ Big Tech, now armed with legal clarity and a price benchmark, will move to dominate this new market for training data. OpenAI, for example, already holds a 53% market share in these new licensing deals.³⁶ The settlement has simply, and expensively, reset the negotiating table.

Table 2: The Bartz v. Anthropic Settlement Terms & Strategic Implications

Section 2.3: The Settlement’s Unanswered Questions: Judge Alsup’s Skepticism and the “Output” Loophole

The Bartz settlement is not yet final. Judge Alsup has repeatedly expressed deep skepticism about the agreement. In court, he has “skewered” the deal ³⁹ and, in a significant move, refused to grant preliminary approval.¹² He has demanded more information, including a “drop-dead list” of all pirated books and a clearer explanation of the claims process.³⁹ He also raised concerns about the “behind the scenes” role of the Authors Guild and Association of American Publishers, worrying they might be pressuring authors to accept a deal that is not in their best interests.³⁹ This judicial skepticism highlights the fragility of attempting to resolve such a novel and complex class action.

However, the most critical detail of the proposed settlement is what it excludes. The agreement explicitly does not release Anthropic from any claims for future conduct. Most importantly, it does not cover claims based on allegedly infringing AI outputs.²⁹

This “output loophole” is the real story. The $1.5 billion settlement ²⁵ only covers the past act of pirate data ingestion. It does nothing to prevent an author from suing Anthropic tomorrow if its Claude model generates a detailed summary of their book that constitutes an infringing derivative work.¹⁴

This aligns perfectly with the legal consensus forming in the Kadrey case ³ and the USCO guidance ²⁰: the new, defensible battleground for creators is market harm from the output. The Bartz settlement, far from ending the war, was a brilliant strategic concession by Big Tech. They paid a massive, one-time fee to take the messy, indefensible “input” (piracy) issue off the table, allowing them to focus all future legal resources on defending the “output” (market substitution), which is the battle they are far more confident they can win.

Chapter III: The “Speed and Collaboration” Doctrine: Big Tech’s Asymmetric Advantages

The hypothesis—that Big Tech wins through superior speed and collaboration—is correct. However, these “collaborative” methodologies are not benign; they are highly effective, asymmetric competitive strategies designed to outpace, overwhelm, and ultimately obsolete both rivals and regulators. This doctrine is waged on three fronts: weaponized open-source development, unified legislative lobbying, and the capture of technical standards bodies.

Section 3.1: Weaponized Open Source: Building an Unregulatable Ecosystem

Big Tech’s “collaboration” through open-source AI is its primary competitive weapon.⁴¹This strategy is led by Meta, with its Llama family of models ⁴², and French startup Mistral, whose models are also open.⁴³ By releasing powerful, high-performance models for free, they achieve several strategic goals simultaneously. First, they accelerate innovation across the entire ecosystem.⁴¹ Second, they commoditize the AI model layer, preventing a single proprietary company (like OpenAI) from achieving a durable market monopoly. This competitive pressure is so intense that in mid-2025, even the historically “closed” OpenAI released its first open-weight models to keep pace.⁴⁶

This “innovation-first” approach is explicitly endorsed by the U.S. government. The Trump Administration’s “America’s AI Action Plan” actively promotes open-source AI models as a key pillar of its strategy to accelerate innovation and ensure American global dominance in the field.⁴⁷

The true strategic brilliance of this “collaboration,” however, is that it creates an anti-regulatory moat. By open-sourcing a powerful model, Big Tech (especially Meta) creates an ecosystem that is too fast, too diffuse, and too distributed to be effectively regulated.

This dynamic can be seen in its conflict with the EU AI Act. Regulators in Brussels want to impose rules and liability on “GPAI model providers”.⁴⁹ Meta’s strategy ⁴³ is not just to compete with OpenAI ⁴⁶; it is to create millions of providers by “collaborating” with a global community of developers. How can the European Commission ⁵⁰ or the Authors Guild ²⁴ enforce rules, serve takedown notices, or file lawsuits against a million individual developers, small businesses, and researchers around the world who download and fine-tune an open-source model? They cannot.

“Open source” is therefore a strategy to outrun and obsolete the very concept of top-down regulation. It ensures that by the time a law is fully implemented, the technology is already too widespread and democratized to control.

Continue reading here (due to post length constraints): https://p4sc4l.substack.com/p/the-hypothesisthat-big-tech-wins