In this talk, I will discuss generative artificial intelligences such as Midjourney and DALL-E in order to examine their semiotic functioning. First, I will propose an operational definition of AIs as co-enunciating machines, devoid of intentionality and initiative, which nevertheless produce visual utterances in collaboration with a human operator and on the basis of structured and reconfigurable archives. Next, I will draw on Pierluigi Basso Fossali’s understanding of the semiotic discipline as the science that studies all the social mediations in the construction of meaning. This conception describes four fundamental spheres of meaning mediation: perception, enunciation, communication and transmission. First, I will examine the relationship between AI and transmission from the perspective of archives and databases. Next, I will relate the dimension of perception to the tools of computer vision. Finally, I will address the theory of enunciation in the context of computational image generation.I argue that generative AIs function semiotically through two new configurations: through archival perception and compositional circuits. With archival perception, I intend to describe the particular way in which Ais see, hear and learn. These operations are performed on large quantities of data and through a reconfigurable genealogy of computational operations. As far as compositional circuits are concerned, I intend to highlight the singular coupling of the logic of visual composition and verbal predication; I am referring to the process of composition through visual “denoising” carried out by diffusion models, guided by the predicative prompts provided by human operators.