Valeriano Grasso

The dual-use nature of artificial intelligence is often treated as a ‘downstream’ problem, one thatarises at the point of deployment or misuse. This paper argues otherwise: a significant proportionof the dual-use risk arises ‘upstream’, in the modelling, training and validation practices that con-fer epistemic authority upon systems prior to their adoption. Design choices, data selection andevaluation criteria do not operate as neutral technical steps, but rather establish what counts asadmissible evidence and reliable output, facilitating the legitimate circulation of these very arte-facts across diverse institutional contexts. In this context, adoption in high-impact settings tendsto shift responsibility away from human and organisational collectives towards the “model” asthe presumed bearer of decisions. Under these conditions, regulation focused exclusively on de-ployment intervenes on systems whose epistemic credibility has already been established. Thepaper therefore proposes considering dual-use as a problem of cross-context circulation of epis-temic artefacts and suggests three levers for quality of regulation aimed at intervening upstream:benchmark governance, auditable data traceability and minimum contestability requirements.

Extended abstract, accepted at MBR026, Rome, June 2026, forthcoming in Springer SAPERE series Recent epistemological work on large language models has focused on questions of knowledge, reasoning, and normative competence, typically evaluating model outputs in terms of reliability, bias, or misuse. While these analyses have clarified important aspects of how LLMs perform epistemic tasks, I argue that they leave underexamined a distinct problem that emerges when language models are treated as models within institutional reasoning practices, rather than merely as tools. The central claim is that the primary epistemic risk associated with the institutionalization of LLMs is not misuse, but a phenomenon I call context drift.