Courts split AI training, push licensing deals

The early legal battles over AI training data spotlighted the contentious use of pirated and copyrighted materials without authorization, setting a precedent for the industry. In September 2025, Anthropic faced a landmark class-action lawsuit for training its AI agent Claude on pirated books, ultimately settling for $1.5 billion paid to roughly 500,000 authors. This case underscored the financial and reputational risks AI companies face when leveraging unlicensed content, a theme that would soon escalate with further high-profile litigation.

By early 2026, the legal landscape intensified as major publishers and prominent creatives challenged AI firms over unauthorized data usage and mimicry. In March 2026, journalist Julia Angwin sued Grammarly—recently rebranded as Superhuman—claiming its AI tool 'Expert Review' unlawfully mimicked her work without consent, highlighting the nuanced issues of creative imitation. Shortly thereafter, five major publishers alongside author Scott Turow filed suit against Meta, accusing the tech giant of training its Llama AI models on millions of copyrighted books and articles sourced from pirate sites, with allegations that Meta deliberately stripped copyright information to conceal the origins. These lawsuits not only spotlighted the ethical and legal quagmires of AI training practices but also raised pivotal questions about whether such use constitutes fair use or requires explicit licensing.

The ongoing litigation against AI companies like Meta has thrust courts into uncharted territory regarding copyright law’s application to AI training. With no definitive rulings yet, the cases are poised to shape the future of AI development by clarifying if training on publicly available copyrighted works without permission is lawful. The outcomes will likely influence industry costs and operational norms, as stakeholders await judicial guidance on balancing innovation with creators’ rights in this rapidly evolving digital ecosystem.

Anthropic’s landmark $1.5 billion settlement, approved by U.S. District Judge Araceli Martinez-Olguin, represents the largest copyright payout in U.S. history, compensating authors and publishers with $3,000 per work across roughly 500,000 titles. This settlement notably resolves disputes not over AI training itself but over Anthropic’s illicit acquisition of data from piracy platforms, setting a precedent that creators’ compensation hinges more on the legality of data sourcing than on the act of AI training with copyrighted material.

A pivotal federal ruling by Judge William Alsup established that training AI models on copyrighted text is transformative and protected under fair use, yet Anthropic’s method of obtaining data through piracy was deemed independently illegal. This nuanced distinction means that while AI training per se may not violate copyright, unauthorized data acquisition remains actionable, leaving broader questions about AI-generated outputs and future licensing unresolved and without binding precedent due to the pre-appeal settlement.

Despite the substantial financial settlement, many authors regard the outcome as a partial victory at best, since the fair use ruling potentially narrows future copyright claims to cases involving illegal data sourcing rather than unauthorized AI training. This shift implies that plaintiffs must now prove piracy in data acquisition rather than mere unauthorized use, fundamentally altering the legal landscape for creators seeking redress in the evolving AI era.

Anthropic’s business trajectory post-settlement underscores the settlement’s characterization as a manageable legal cost rather than a constraint on innovation; the company has continued deploying Claude models in high-profile contexts such as the European Parliament’s internal AI platform and has released new model versions throughout the litigation. This continuity highlights how, despite the historic payout, AI commercial operations remain robust and largely unaffected by ongoing copyright debates.

By mid-2026, federal courts have begun to clarify the application of fair use to AI training data, issuing several pivotal rulings that underscore the complexity of this evolving legal terrain. Notably, the Northern District of California's June 2023 decision in Barts v. Anthropic distinguished between pirated and legally obtained content, ruling that while the use of pirated works for AI training is unequivocally not fair use, the use of purchased and scanned books can qualify as fair use due to the transformative nature of generative AI. Judge Alup emphasized the importance of the first and fourth fair use factors and rejected claims that generative AI outputs dilute the market for original works, noting that such dilution theories lack historical acceptance in copyright jurisprudence.

Recent federal court decisions, including Reuters v. Ross, Bartz v. Anthropic, and Kadrey v. Meta, reveal a nuanced judicial approach that hinges on the concept of 'transformative use' to assess fair use in AI training contexts. Courts are increasingly weighing how transformative use impacts market harm and dilution claims, thereby shaping the contours of copyright protection in the AI era. This evolving jurisprudence draws a clear legal line between pirated content—which courts consistently exclude from fair use protections—and legally obtained materials, signaling a cautious but significant shift in how AI training datasets are evaluated under copyright law.

By mid-2026, the legal landscape surrounding AI training data has intensified with multiple high-profile lawsuits challenging the use of copyrighted works without authorization. Authors have appealed to California courts to clarify whether training AI models on copyrighted content constitutes infringement, while Anthropic faces a $75 million lawsuit alleging it used pirated books from shadow libraries to train its Claude AI model, following an earlier $1.5 billion class action settlement from which some authors opted out seeking higher statutory damages. These cases underscore escalating legal stakes and the urgent need for clearer copyright boundaries in AI training practices.

Central to these disputes is the evolving judicial interpretation of 'fair use,' particularly the role of 'transformative use' and its effect on market harm, as courts weigh whether AI training constitutes a permissible transformation or an infringement that dilutes original works. Key cases like Reuters v. Ross, Bartz v. Anthropic, and Kadrey v. Meta are pivotal in shaping this doctrine, with courts distinguishing between pirated and legally acquired content to determine liability. This nuanced legal reasoning reflects the complexity of applying traditional copyright principles to generative AI technologies.

With dozens of lawsuits progressing globally, there is mounting anticipation that unresolved cases will ascend to the Supreme Court, which could deliver a definitive ruling on copyright law’s application to AI training data. While some disputes, such as the recent settlement involving BSA | The Software Alliance over pirated works, have momentarily eased tensions, the broader implications remain uncertain. Stakeholders warn that these outcomes could impact billions of dollars in value and fundamentally reshape how AI companies source and manage training data, highlighting the high stakes and legal complexity at play.

By mid-2026, AI companies began pioneering a transformative business model that directly compensates creators for generating content tailored specifically for AI training purposes. One firm reported disbursing over $10 million to writers and musicians producing data expressly for AI, signaling a deliberate move away from relying on unlicensed, scraped content. This shift not only acknowledges the value of original creative labor but also fosters a new ecosystem where creators are active participants and beneficiaries in AI development.

Amid persistent legal ambiguity surrounding copyright and fair use in AI training data, licensing agreements have surfaced as a pragmatic industry response. By mid-2026, AI developers increasingly negotiated formal licenses with content creators, seeking to preempt litigation and establish clearer rights frameworks. This approach reflects a strategic balance between innovation and respect for intellectual property, offering a viable pathway forward while courts continue to deliberate on the complex intersection of AI and copyright law.

AI developers are forging new alliances with writers and musicians, paying for purpose-built training data and establishing licensing deals as the legal fog around fair use persists.