Superman in Delft

This article takes a position on several key issues that have emerged in recent semiotic and enunciational debates on Visual Generative Artificial Intelligence (VGenAI). These positions draw on an experiment conducted with approximately twenty VGenAI platforms, consisting in producing images depicting Superman in the style of Vermeer (Section 1). Throughout the article, the technical literature on VGenAI is also examined in order to identify the operating principles that render the outcomes of this experiment intelligible (Sections 3 and 4). This analysis shows that verbal, rather than visual, processing models—namely Transformers—play a crucial role in image generation. These models operate through a process of fragmentation and de-semantisation of the original discourse (training phase), followed by a purely syntactic recomposition of fragments (generative phase). Drawing on terminology from the economic–political debate on algorithmic capitalism, the article discusses the connection between processes of dissemblage and ressemblage. The article identifies three major themes in the semiotic debate. The first concerns the applicability of the concept of enunciation to VGenAI: the article argues that this is possible, provided that enunciation is understood in a purely productive and material sense (Section 2). The second concerns whether the “latent” or virtual spaces of VGenAI can be considered “archives”: this possibility is rejected, since processes of dissemblage eliminate the individuality of the items in the training corpora (Section 3). The third theme concerns the salience and reconfiguration of the concept of “style”: the article argues that, thanks to the processes of ressemblage performed by VGenAI, styles are omnipresent yet stripped of historical and cultural grounding (Section 4). The conclusion reverses the approach adopted in the preceding sections and asks how the debate on VGenAI may prompt a rethinking of enunciation theory within post-Greimasian semiotics. The article proposes defining enunciation as an event of production or transformation of forms and distinguishes three broad types of enunciative practices: productive, which intervene upon material forms; interpretative, which manipulate cognitive forms; and delegative, which organise shifts in responsibility related to the other two and thus operate within the domain of social forms. This framework enables semiotics to monitor the fundamental risk highlighted throughout the article: that the productive enunciational delegation assigned to VGenAI may be extended—automatically and surreptitiously—to the interpretative dimension.