So-called neural networks, especially Generative Adversarial Networks (Goodfellow et al. 2014), form a central method of machine learning in music production. This method is particularly interesting for music and art because it promises to be able to produce something new and surprising, in other words to show a kind of machine creativity (Cádiz et al. 2021). This creativity, be it something unexpected, inspiring or a unique sound characteristic, appears at the limits of AI through its limitations and malfunctions, the moments of failure of these technologies (Claussen 2020). In my lecture, I therefore contrast a kind of AI of clean musical notation formats on the one hand with a more phonographic or sound studies oriented AI on the other and discover the roughness, the special uniqueness in the sound of the neural networks. In doing so, references are made to positions in media studies as well as to the genre of the so-called Clicks & Cuts (Cascone 2000; Großmann 2003).