Send message to Authors

Paper short abstract

We examine how LLMs become infrastructures for knowledge production through an ethnographic study of Italian AI startups. We identify two key practices enabling their diffusion: “sensorization” and “mythologization,” which shape how organizations render social reality knowable and actionable.

Paper long abstract

The rapid diffusion of Large Language Models (LLMs) is reshaping how “the social” is rendered knowable and actionable. In line with STS debates on the representation and enactment of social realities through technical devices, this paper examines how LLMs are locally implemented as infrastructures for knowledge production within firms.

Drawing on an 18-month ethnographic study of the Italian AI startup landscape, we investigate how LLM-based solutions are configured and made meaningful across diverse organizational settings. Our empirical material includes 18 in-depth interviews with startup managers, ethnographic observations at 10 business events and exhibitions, and a corpus of textual and visual materials collected throughout the research period.We identify two interrelated practices: sensorization and mythologization. Sensorization refers to the expansion of data infrastructures—through the installation of physical and digital sensors and the integration of heterogeneous data streams—that render organizational processes legible to LLMs. Mythologization captures the discursive and symbolic work – such as invoking Silicon Valley, exponential growth, or Moore’s Law – through which AI adoption is legitimized.

We argue that these practices actively participate in producing specific imaginaries of the social, privileging datafied and anticipatory forms of knowledge. In this sense, LLMs emerge as socio-technical arrangements that reconfigure what counts as “relevant” within organizations, while simultaneously aligning local actors with global narratives of AI innovation. The paper thus highlights the mutual constitution of AI imaginaries, organizational practices, and the epistemic transformations associated with the uptake of LLMs.