The paper "Rethinking open source generative AI: open-washing and the EU AI Act" by Andreas Liensenfeld and Mark Dingemanse explores the challenges and implications of open-source generative AI systems, particularly in the context of the upcoming EU AI Act. The authors argue that while the term "open source" is widely used, many models are only "open weight," meaning they provide access to model weights but lack transparency in training and fine-tuning data. This practice, known as "open-washing," undermines the core principles of open-source software and hinders innovation, research, and public understanding. The EU AI Act, the world's first comprehensive AI law, places significant emphasis on open-source systems, exempting them from certain regulatory requirements. However, the act's definition of "open source" is vague, allowing for lobbying and the potential dilution of its meaning. The authors propose a composite and gradient approach to assessing openness, recognizing that AI systems are complex and that openness can vary in degrees. A survey of 40 text generators and 6 text-to-image generators reveals that while some systems are highly open, many are only open in name, with limited or no transparency in training data and methods. The authors highlight the importance of detailed documentation, scientific papers, and peer review in assessing the openness of these systems. They also discuss the implications of open-washing, including the risk of false assurance and the potential for corporate interests to co-opt the concept of openness. The paper concludes by emphasizing the need for a nuanced and evidence-based approach to assessing openness, which can help foster a transparent and accountable AI ecosystem. Full disclosure of training data is crucial for ensuring safety, reproducibility, and legal compliance. The authors provide a framework for assessing openness, which can be used to make informed decisions about the deployment of generative AI systems.The paper "Rethinking open source generative AI: open-washing and the EU AI Act" by Andreas Liensenfeld and Mark Dingemanse explores the challenges and implications of open-source generative AI systems, particularly in the context of the upcoming EU AI Act. The authors argue that while the term "open source" is widely used, many models are only "open weight," meaning they provide access to model weights but lack transparency in training and fine-tuning data. This practice, known as "open-washing," undermines the core principles of open-source software and hinders innovation, research, and public understanding. The EU AI Act, the world's first comprehensive AI law, places significant emphasis on open-source systems, exempting them from certain regulatory requirements. However, the act's definition of "open source" is vague, allowing for lobbying and the potential dilution of its meaning. The authors propose a composite and gradient approach to assessing openness, recognizing that AI systems are complex and that openness can vary in degrees. A survey of 40 text generators and 6 text-to-image generators reveals that while some systems are highly open, many are only open in name, with limited or no transparency in training data and methods. The authors highlight the importance of detailed documentation, scientific papers, and peer review in assessing the openness of these systems. They also discuss the implications of open-washing, including the risk of false assurance and the potential for corporate interests to co-opt the concept of openness. The paper concludes by emphasizing the need for a nuanced and evidence-based approach to assessing openness, which can help foster a transparent and accountable AI ecosystem. Full disclosure of training data is crucial for ensuring safety, reproducibility, and legal compliance. The authors provide a framework for assessing openness, which can be used to make informed decisions about the deployment of generative AI systems.