<html><head><meta http-equiv="content-type" content="text/html; charset=utf-8"></head><body style="overflow-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;">Hi Josh,<div><br id="lineBreakAtBeginningOfMessage"><div><br><blockquote type="cite"><div>On 5 Mar 2025, at 2:24 AM, Josh Berkus <josh@berkus.org> wrote:</div><div><div><br>1. The current language does NOT require attribution (or licensing for that matter); it just requires a notice that the output was produced using some kind of AI.<br><br></div></div></blockquote><div><br></div><div>Correct. As Pamela pointed out, the current clause somewhat oversteps, as not all users will use the output to train a new model. However, it can also serve as an indicator, informing downstream users who intend to use the output as training data to check whether the source model’s license imposes any conditions on output reuse. Another option, as I suggested, is to require only users who distribute the output as a dataset to notice the original model and include the license name. This narrows the scope but adds an additional requirement.</div><div><br></div><br><blockquote type="cite"><div><div>2. Also, by my reading, the notice requirement isn't inheritable. So if I were to take the output from a ModelGo model, and use it to train a new model called "JoshAI", then there is no requirement that JoshAI have any particular notices.<br></div></div></blockquote><div><br></div><div>Partly so. If the original model is licensed under MG-BY-OS, derivatives must also be licensed under MG-BY-OS because it is copyleft. However, for other variants like MG0, no such notice is required. The reason I intend to add a notice in MG-BY-OS is to ensure open-source inheritance. Consider a scenario where I use an MG-BY-OS model and distribute its output as a dataset on Hugging Face without indicating which model I used. If another user downloads the dataset, trains a new model, and changes its license, they may unintentionally violate the MG-BY-OS license. This behavior is very common in current model development, and you can find many such extracted datasets on Hugging Face.</div><div><br></div><br><blockquote type="cite"><div><div>3. It's not clear that model output has any copyrightability, so even if you wrote a provision which contained attribution and was inheritable, it's not clear that it could be legally enforced. Lawyers?<br></div></div></blockquote><br></div><div>I think this is a complex legal question. As stated in the MG license, our clause is a condition for distributing model output, and 2.4b read:</div><br><blockquote style="margin: 0 0 0 40px; border: none; padding: 0px;"><div><div>For the avoidance of doubt, the </div></div><div><div>foregoing shall not constitute or be construed as any statement, representation or </div></div><div><div>warranty by the Licensor that there are or will be rights (including Intellectual Property </div></div><div><div>Rights) capable of subsisting in and to any Output.</div></div><div><br></div><div><br></div></blockquote>Best,<div>Moming</div></div></body></html>