AI news ·
Yale researchers propose copyleft licensing to keep generative AI models open
Yale researchers propose a license requiring AI models trained on open-source code to remain fully open. This copyleft model enforces reciprocity to stop "open washing."
Yale researchers have put forward a new licensing model designed to close a fairness gap in artificial intelligence: companies that train AI models on free, open-source code would be required to keep those models fully open themselves. The proposed Contextual Copyleft AI License (CCAI), detailed in the International Journal of Law and Information Technology, extends a decades-old software licensing tactic to generative AI, giving open-source developers a legal tool to enforce reciprocity.
The free and open-source software (FOSS) community underpins much of modern computing-from operating systems and web servers to scientific tools and internet infrastructure. Yet many AI systems built on that code do not disclose their training data or model architecture, leaving developers with no insight into how their work is being used. The CCAI treats generative AI models as derivative works, so anyone training a model on open-source code would need to release their architecture and training data under the same open terms.
The copyleft mechanism
Copyleft licensing flips standard copyright. Instead of restricting reuse, it mandates that derivative works stay as free as the original. The Yale team, led by DEC Junior Fellow Grant Shanklin, proposes applying that principle to AI training. Under CCAI, a developer who trains a generative model on copyleft-licensed code must make the resulting model's core components public.
"Our analysis showed that extending the copyleft concept to generative artificial intelligence has the potential to give open-source software developers meaningful control over how AI developers use their code," Shanklin said. "Importantly, it would incentivize the formation of a community committed to building AI tools aligned with the values of the free and open-source movement, which could help ensure that AI models are developed openly and responsibly."
Legal standing and regulatory alignment
The researchers' legal analysis finds the license feasible under current copyright law-provided that training AI models on copyrighted code is not ruled fair use. The fair use doctrine, which allows unlicensed use of protected works in specific contexts, remains the key legal uncertainty. The study also notes that generative AI carries a higher risk profile than traditional software because it can directly produce deceptive or harmful content, such as highly effective phishing emails.
The authors point to regulations like the European Union's AI Act, which prohibit systems that use subliminal or manipulative techniques to distort behavior. Their analysis suggests that CCAI licensing would work alongside such rules to mitigate the specific dangers of openly available AI models.
Blocking "open washing"
One goal of the license is to stop what the researchers call open washing-companies marketing models as open while keeping key details hidden. Claudio Novelli, a de Vries-Sherif Associate Research Scientist at the DEC, said, "AI companies have benefited from using open-source code, but their resulting models are not really open. They might be transparent about some aspects, but other key components remain closed." The CCAI license would require full disclosure, not selective transparency.
The study outlines three main benefits: giving developers control over code reuse, expanding the pool of truly open-source generative models, and disincentivizing deceptive openness claims. By linking access to valuable training data with a commitment to openness, the license aims to create a self-reinforcing community of AI builders who share resources rather than hoard them.
Why this matters for science and research
For researchers who rely on open-source software to run experiments, manage data, or build computational models, opaque AI systems create a reproducibility and trust problem. A copyleft-style license would not only protect the code they contribute but also push more AI training datasets and architectures into the public domain. That openness can accelerate peer review, enable independent safety audits, and lower the barrier for academic labs that cannot afford proprietary alternatives. The license is not yet a market reality, but the legal pathway exists-and the debate over fair use will likely determine whether it becomes a practical tool for the open-science community.