Bartz v. Anthropic: Copyright Dispute Over GPL-licensed Code in AI Training
Escalating AI Copyright Disputes: Bartz v. Anthropic Ignites Global Debate on the Legality of Open-Source and Commercial AI Training Data
In spring 2024, the case of Bartz v. Anthropic, filed in the U.S. District Court for the Northern District of California, quietly emerged—only to rapidly trigger a seismic wave across technology ethics, open-source ecosystems, and AI governance. Initiated jointly by Bradley M. Kuhn, former Executive Director of the Free Software Foundation (FSF) and a prominent open-source advocate, along with several developers, the lawsuit accuses Anthropic of systematically violating the terms of the GNU General Public License version 3 (GPLv3) by incorporating multiple GPLv3-licensed open-source projects—including critical infrastructure toolchains maintained by the plaintiffs—into the training corpus for its Claude series of large language models, all without permission, attribution, or compensation. This is no isolated incident. Rather, it marks the third major judicial challenge globally—following GitHub Copilot (2022) and Getty Images v. Stability AI (2023)—to directly confront the core legal question: Does training an AI model constitute “reproduction” or creation of a “derivative work” under copyright law? Crucially, in its formal statement issued in April 2024, the FSF explicitly asserted: “If the training process substantively extracts and embeds the protected expression’s structure, sequence, and organization (SSO), it may trigger compliance obligations under strong copyleft licenses such as the GPL.” This legal framing is shifting the legitimacy boundary of AI training data—from the hazy terrain of “fair use” debates—into the clear, enforceable domain of contractual license obligations.
The “Silent Clause” of Open-Source Licenses: Systemic Neglect of Training-Data Compliance
For years, mainstream AI vendors have routinely invoked Section 107 of the U.S. Copyright Act—the “fair use” doctrine—to justify model training, emphasizing its non-expressive, transformative purpose and purported “zero substitution effect” on the original works’ markets. Yet Bartz is the first case to precisely center its argument on the contractual obligations embedded in open-source licenses themselves. GPLv3 Section 0 explicitly defines “propagation” to include “making copies… in any manner,” while Section 2 mandates that “any propagation of this Program or a Modified Version must be done in full compliance with the terms of this License.” In its statement, the FSF sharply observed: When model weights internalize specific algorithmic logic, API design patterns, or even comment styles from GPL code during training—and subsequently reproduce functional structures during inference—the resulting model ceases to be a neutral “tool” and becomes a “functional derivative” of the GPL code. This view aligns with the European Union’s draft Artificial Intelligence Act (2023), which imposes a mandatory requirement that “high-risk AI systems ensure the lawful origin of training data,” and echoes findings from the Max Planck Institute for Innovation and Competition: “If neural network weights can be reverse-mapped to the original creative expression in specific training samples, such mapping constitutes de facto reproduction.”
Notably, several recent open-source AI projects surfacing on Hacker News—such as OpenCode, a lightweight programming agent trained exclusively on MIT- and Apache-licensed code—demonstrate the feasibility of compliant alternatives. Its developers explicitly declare: “All training data underwent manual review to exclude any GPLv3-or-later licensed code; core model weights are released with a complete data provenance report.” This “license-aware training” paradigm is rapidly evolving from a fringe practice into an emerging industry consensus.
The “Data Debt” of Commercial AI: Structural Tensions Between Training Black Boxes and Compliance Costs
The deeper dilemma confronting leading AI firms like Anthropic lies in the fundamental tension between their commercial models and open-source ethics. To maximize model performance, the Claude 3 series was trained on datasets comprising trillions of tokens—spanning public GitHub repositories, Stack Overflow Q&A, technical documentation, and academic papers. Performing itemized license-compliance reviews across such massive, heterogeneous, and dynamically updated data streams entails prohibitive engineering overhead and computational cost. Industry estimates suggest scanning a 1 TB code dataset for GPLv3 compatibility consumes roughly 200 GPU-hours—a burden nearly untenable at the petabyte-scale training volumes typical in industrial AI. Hence, “data debt” has arisen: Vendors capitalize on long-term compliance risks to achieve short-term performance gains, banking on judicial leniency toward broad interpretations of “fair use.”
But Bartz is rewriting these rules of engagement. Key evidence submitted by plaintiffs shows statistically significant homology (p < 0.001) between code snippets generated by Claude 3 on specific programming tasks—and the function signatures, error-handling logic, and memory-management patterns found in libgplutils, a GPLv3 project maintained by the FSF. This directly undermines the technical defense that “models learn only abstract concepts,” revealing instead the phenomenon of “ghost residue”: the persistent, legally cognizable imprint of protected expression within training data. When AI evolves from a passive mirror into an active encoder, its outputs inherently carry the legal DNA of their inputs.
Divergent Global Governance: EU’s Strict Regulation, U.S. Judicial Experimentation, and China’s Search for Balance
The ripple effects of Bartz extend far beyond national borders. The EU is leveraging its Artificial Intelligence Act to build a “full-lifecycle compliance framework,” mandating that high-risk AI systems disclose training-data inventories and copyright authorization documentation. The U.S., by contrast, is gradually clarifying boundaries through case law—most notably Andy Warhol Foundation v. Goldsmith (2023), whose ruling emphasized that “transformative use” must yield “new meaning, new information, or new aesthetics,” thereby raising the bar for AI training claims. In China, Article 12 of the Interim Measures for the Administration of Generative AI Services requires stakeholders to “respect intellectual property rights,” yet remains silent on training-data specifics. Notably, leading domestic open-source communities—including OpenI Qizhi—are proactively launching initiatives for “AI-friendly licenses,” proposing amendments to Apache 2.0 that add appendices mandating “training-data transparency”—a pragmatic, governance-oriented response.
A Paradigm Shift in the Open-Source Community: From “Code Donors” to “Data Sovereignty Defenders”
The FSF’s statement is no nostalgic protest—it is a quiet declaration of paradigm revolution. It signals that open-source contributors are transforming from passive data suppliers into active defenders of data sovereignty. In the future, platforms like GitHub may integrate license-compliance scanning plugins to automatically flag GPL code’s AI-training risks; Model Cards—the standardized documentation for AI models—will likewise need to expand their “data provenance” fields, mandating disclosure of license-type distributions within training sets and the audit methodologies employed. As one widely discussed post on Hacker News put it: “If an email application inspired by the Arc browser (Show HN: Email App) can reconstruct human–computer interaction through minimalist design, why not reconstruct AI training ethics with the same rigorous engineering mindset?”
The Bartz case will eventually conclude—but its true legacy lies in compelling the entire industry to acknowledge a foundational truth: The height of AI intelligence can never transcend the moral integrity of its data foundations. In an era of surging algorithms, open-source licenses are not obsolete paper shackles—they are the genetic checksums of digital civilization.