This would only work if using "content" for training is not fair use, or if updating model weights through exposure to content is governed by copyright. For something published on the internet, there's no indication either of those is true.
The "ModelShare License" is designed to grant users the non-exclusive right to utilize specific content for training, developing, or testing artificial intelligence models. In exchange for this right, users are obligated to provide the content owner with complete access to the resultant AI model, including its architecture and weights. The license ensures that any third-party recipients of the model are also bound by these obligations. The primary goal is to foster transparency and sharing when the content is used for machine learning purposes.