Can artificial intelligence generate scientific knowledge? Pre-theoretical knowledge without epistemic agency

Abstract

This paper examines whether artificial intelligence systems can generate genuine scientific knowledge. I argue that AI can produce scientific knowledge, but only in a pre-theoretical and non-agential sense. To make this thesis precise, I distinguish knowledge-as-a-product (truth-apt, methodologically warranted outputs) from knowledge-as-a-status (an epistemic standing attributable to responsible agents), and I introduce a tripartite distinction between the capacity for epistemic responsibility (categorical), the exercise of that capacity (graded), and the allocation of responsibility within scientific practice (graded). Current AI systems can contribute knowledge-as-a-product without being knowers or bearers of knowledge-as-a-status. Pre-theoretical knowledge, defined here equivalently as pre-explanatory-integration knowledge, requires truth-aptness, methodological warrant robust against ML-specific failure modes, and the absence of explanatory integration. Engaging with the theory-ladenness thesis (Hanson 1958; Pietsch 2015; Andrews 2025), with hermeneutic and pragmatist perspectives on scientific knowledge, and with recent positions in the philosophy of machine learning (Coeckelbergh 2025; Durán and Pozzi 2026), I clarify that "pre-theoretical" denotes the absence of explanatory integration within the output, not the absence of theoretical influence on the knowledge-production process. The framework is tested against a sustained case study of AI-supported mammography screening. The analysis clarifies who bears responsibility for AI-assisted results, credits AI contributions without implying authorship, and suggests governance criteria for promoting AI outputs to full scientific knowledge.