Abstract
As courts grapple with whether training AI on copyrighted material constitutes fair use, companies have adopted divergent compliance strategies. This piece analyzes recent litigation, industry practices, and legislative proposals aimed at balancing innovation with creator rights.
Background
The development of generative artificial intelligence (AI) systems hinges on access to extensive datasets that are essential for enhancing model performance and reducing algorithmic biases. However, the reliance on such data has created a complex interplay with copyright law, as many AI models are trained using copyrighted content without authorization from the rights holders. This legal ambiguity has led companies to choose a variety of approaches to navigate the rights landscape—from utilizing rights-cleared datasets to permissibly scraping content—and these choices carry significant implications for compliance, innovation, and liability.
Output-Phase vs. Training-Phase Fair Use
When an AI model generates content, that output is typically evaluated through traditional copyright principles. If a generated work exhibits “substantial similarity” to an original piece, the four-factor fair use test kicks in. This test considers the purpose and character of use, the nature of the original work, the amount and substantiality of the portion used, and its market effect. A noteworthy example arises from The New York Times Co. v. OpenAI, where plaintiffs claim that ChatGPT sometimes reproduces content from the Times verbatim. This situation raises vital questions: At what point does AI-generated content cross over into copyright infringement?
Courts may draw on precedents like Authors Guild v. Google, Inc., where the Second Circuit deemed digital indexing for search purposes transformative. However, the applicability of this standard to AI outputs—especially those mimicking a creator’s style or structure—remains uncertain. More contentious is the question of whether using copyrighted works for training qualifies as fair use. As outlined in a May 2024 report from the U.S. Copyright Office, transforming expressive works into nonhuman-readable data could be deemed “transformative,” especially in the context of research. The report parallels model training with the Google Books case, which permitted large-scale digitization for analytical purposes. Yet, critics assert that extensive training on copyrighted materials—even in a transformed state—may overstep fair use, particularly when generated outputs reflect recognizable elements of original works.
In Kadrey v. Meta Platforms, Inc., the court opted not to classify a language model as a derivative work without explicit outputs that mirrored the plaintiffs’ books. As this case is still at the motion-to-dismiss stage, future rulings hold potential to redefine these boundaries.
Divergent Industry Practices and Policy Proposals
With legal uncertainties looming, industry responses have diverged widely. A notable image-generation platform shifts the burden of infringement liability onto its users, including disclaimers in its terms of service and implementing watermarking for attributions. Conversely, some models utilize the LAION dataset, a web-scraped repository tagged with open licenses. This approach faces scrutiny over its provenance accuracy and the limited enforceability of creator opt-outs. These varying practices underscore a broader concern: that AI systems may reproduce a creator’s distinctive style without securing consent or compensation.
In light of these issues, some stakeholders have rallied behind legislative initiatives such as the Preventing Abuse of Digital Replicas Act (PADRA). This proposed law would allow artists to pursue legal action when AI uses their unique style for commercial purposes. PADRA is carefully tailored, requiring both “demonstrable intent” and a “commercial purpose.” The aim is to close the legal gaps left by traditional copyright frameworks concerning stylistic appropriation. In addition to PADRA, other widespread proposals include exemptions for text-and-data mining specifically for research, the establishment of opt-out registries for creators, and collective licensing models inspired by practices in the music industry.
The Evolving Landscape
As the landscape of generative AI and copyright law continues to evolve, rights-cleared development models—characterized by licensed datasets, rigorous moderation, and robust provenance tools—reflect one compliance-oriented approach but do not offer a panacea. Courts face the crucial task of determining how fair use doctrine applies to both the training and output of AI-generated content, while lawmakers explore targeted reforms like PADRA. The future legal framework surrounding generative AI is unlikely to crystallize through a single court ruling or legislative act. Instead, it will emerge through a patchwork of judicial decisions, negotiated licensing arrangements, industry standards, and innovative legislative efforts. The challenge ahead involves not just balancing innovation with creator rights, but also constructing a predictable, equitable system that operates effectively at scale.