Abstract
Large Language Models (LLMs) have facilitated the automation of language generation and elaboration. However, they have also prompted critical concerns regarding the societal impact of accelerated, machine mediated communication. Indeed, language serves as a central vehicle for the creation, maintenance, and reinforcement of social hierarchies, while the accelerating pace enabled by automation constitutes a normative force shaping contemporary power dynamics. Consequences are particularly evident in the artistic sector, where LLM-based technologies promote cultural homogenisation and advance extractive practices that undermine the economic and creative agency of artists. HCI research holds strong potential for critical insights into the role of accelerated language production. However, research remains underdeveloped in the field. Drawing on critical theories of social acceleration and ideology, this dissertation advances two complementary research approaches. Firstly, it adapts and extends Thompson’s critical theory of ideology to examine the automation of language. Secondly, it embraces adversarial design to engage with the temporality of obsolete technological devices as a material for design exploration. This thesis is empirically situated within the artistic practice, and this situatedness manifests in two aspects of the research. I involved artists in field studies to leverage their ability to critique emerging technologies through creation and disruption. Working in the dual role of artist/researcher, I defined the AI-art practice of crafting Disobedient Discursive Machines to critique rather than celebrate AI development. These machines repurpose the physical embodiment of old media devices, combining them with a LLM. Through their discursive engagement with audiences, they 'disobey' the capitalist imperatives of efficiency and overproduction embedded in generative AI (GenAI). This research makes three key contributions. First, it provides Isotta and Alessia, two embodied forms of knowledge on the power issues of LLM, and their evaluation in artistic exhibitions. Second, it offers the design tactics foregrounded by the practice of crafting Disobedient Design Machines, namely Outdated Technologies as Political Design Materials, Temporal Standards as Sites of Contestation, and Frictional Encounters with GenAI. Third, the study adapts Thompson’s methodology to HCI, articulating a theoretical framework to reflect critically on the ideological consequences of linguistic patterns in real-world LLM applications. All together, these contributions advance HCI methodologies that acknowledge language technologies as socio-artefacts situated within cultural and political contexts. Specifically, they challenge the normative efficiency-driven ideology that underpins GenAI development and support power-aware co-design practices.