Len Krawczyk (Bio)

Beyond the Black Box: LLM Evaluation Devices as Objects of Enquiry

Approaches to AI are often trapped in a reductive binary: they either treat it as an isolated computational entity, erasing its social conditions of production, or reduce it to a mere reflection of human biases, denying its specific technical agency. As a result, it becomes difficult to ascribe responsibility or formulate an adequate ethical framework for AI. To move past this deadlock, AI can be conceptualized as an algorithmic assemblage. Following Deleuze and Guattari, an assemblage is not a collection of static objects but a socio-technical multiplicity in which heterogeneous human and non-human elements are mutually constituted. AI systems, within this framework, do not operate in a vacuum: their production is embedded in values, institutional arrangements, and cultural imaginaries. Such an understanding of algorithms is already well established within Science and Technology Studies. While existing literature has examined AI systems in-the-making, this presentation identifies possible objects for a genealogical analysis of algorithmic assemblages: the processes of LLMs’ performance evaluation, understood as heterogeneous practices and devices through which models become knowable, comparable, and actionable. These include metrics (technical components that allow for translation of institutional demands into computational infrastructures); benchmarks (standardized evaluation tasks that establish criteria of model “goodness”); and evaluation arenas (where human judgments are aggregated into comparative rankings). Rather than functioning as neutral measures, these devices operate as inscription devices: traces through which the values, assumptions, and priorities guiding LLM development can be examined. They also are sites where opaque computational properties encounter socially desired standards of performance. This tension should be understood not merely as a constraint, but as a productive force within the field. It shows that AI development is always culturally embedded, challenging both techno-deterministic narratives and purely ideological critiques of AI.

*This abstract was language-edited for grammar and clarity with the assistance of ChatGPT-5.5 and Gemini 3.5 Flash. Only language editing was applied; the argument and content remain the author’s own.