The theory that the GNU General Public License (GPL) propagates to AI models trained on GPL-licensed code remains unresolved as of 2025, despite a significant decline in public discourse compared to the early days of tools like GitHub Copilot. While the idea that AI models inherit the copyleft obligations of their training data was widely debated in 2021, it is no longer a dominant narrative in mainstream AI or open source discussions. However, the theory has not been legally dismissed, and ongoing litigation in the U.S. and Germany suggests it could resurface with legal force. Two key lawsuits highlight the legal uncertainty. In Doe v. GitHub (the Copilot class action), plaintiffs argue that training on public GitHub repositories—some under GPL and other open source licenses—constitutes a breach of license terms, particularly regarding attribution and the requirement to release derivative works under the same license. Although many claims were dismissed, the court allowed the core argument about license violations to proceed, especially concerning the failure to comply with open source license conditions during model output. Notably, the court did not rule on whether the model itself is a derivative work under copyright law, nor did it mandate GPL licensing of the model. Instead, it left open the possibility of injunctive relief to prevent future violations. The German case, GEMA v. OpenAI, is even more significant for its legal reasoning. The Munich I Regional Court ruled that when an AI model memorizes and reproduces copyrighted lyrics verbatim in response to simple prompts, this internal memory constitutes a "reproduction" under German copyright law, even if the data is encoded in model parameters. The court emphasized that the act of storing a work in a form capable of being reproduced—regardless of technical format—falls within the scope of copyright protection. This judgment establishes that AI models can legally contain reproductions of copyrighted material, which opens the door to arguments that models trained on GPL code may themselves be considered reproductions or derivatives of that code. While this does not directly confirm GPL propagation, it provides a foundational legal principle: if a model contains a copy of GPL-licensed code, then the distribution of that model could be seen as distributing a derivative work. If such a model is distributed without complying with GPL terms—such as providing source code in a modifiable form—the license could be violated. However, no court has yet ruled that the entire model must be released under GPL, nor has any jurisdiction established that copyleft automatically applies to AI systems. Legal and technical arguments against propagation remain strong. From a copyright perspective, AI models are statistical abstractions, not direct copies of source code. The model’s weights do not contain human-readable code, and outputs are probabilistic, not reproductions. Courts in the UK and Japan have similarly rejected the idea that models are infringing copies. The GPL itself was not designed for AI systems; its language focuses on software modifications and linking, not machine learning. Requiring disclosure of model weights or training data as "source code" under GPL is impractical, as these are not easily modifiable by humans. The Free Software Foundation (FSF) and Open Source Initiative (OSI) acknowledge this, with the OSI explicitly avoiding mandatory training data disclosure in its Open Source AI Definition, while the FSF calls for new standards to ensure freedom in AI, not a forced application of existing licenses. Practically, enforcing GPL propagation across AI models trained on diverse datasets—many containing conflicting licenses—would be unworkable. It could lead to widespread avoidance of GPL code in training data, undermining the very purpose of copyleft by limiting its utility in AI development. This could harm open source ecosystems rather than protect them. Major organizations remain cautious. The OSI promotes transparency and reproducibility without requiring GPL-style propagation. The FSF advocates for broader freedom, including training data, but treats this as an ethical ideal rather than a legal mandate under current GPL versions. The Software Freedom Conservancy warns against judicial interpretations that could harm the open source community. In conclusion, while the GPL propagation theory is not legally settled, it is far from dead. It remains a theoretical possibility that could gain traction depending on future rulings, particularly if courts recognize AI models as containing protected works. However, the prevailing view among legal experts, technologists, and open source leaders is that the theory is impractical, technically unsound, and potentially harmful to the open source movement. The focus has shifted toward developing realistic, community-driven solutions—such as improved data provenance, automated licensing, and transparent model sharing—rather than enforcing outdated licensing doctrines on new technologies. The path forward lies not in legal overreach, but in innovation that aligns with the spirit of software freedom in the age of AI.