Generative AI as a Knowledge Distribution System

Generative AI is increasingly integrated into scientific knowledge production, reshaping how information is generated, circulated, and legitimised across research ecosystems. This paper conceptualises generative AI as a Knowledge Distribution System (KDS): a socio-technical infrastructure that mediates access to epistemic resources at scale and thereby influences what counts as credible knowledge. While Large Language Models (LLMs) enable unprecedented forms of automation and synthesis, their deployment raises concerns about opacity, accountability, epistemic bias, and the concentration of informational power in the hands of a few dominant actors. Against this backdrop, the paper argues that relying primarily on centralised LLMs is neither ethically neutral nor epistemically sustainable. It proposes an alternative governance-oriented design approach based on Small Language Models (SLMs) as a distributed ecosystem of specialised models, trained on well-curated datasets and accountable to distinct institutional and community-based norms. Such an architecture could support pluralism, reduce systemic vulnerabilities, and enable more transparent and auditable forms of AI-assisted knowledge mediation. By reframing generative AI as a KDS rather than merely a productivity tool, the paper highlights the ethical stakes of emerging AI infrastructures and outlines a feasible pathway towards more sustainable and democratically aligned knowledge distribution.