The question of whether algorithmic systems generate genuinely new knowledge remains one of the most contested issues in AI research and philosophy of technology. This paper argues that the debate is unresolved because it conflates two fundamentally different concepts of novelty. The distinction between weak, information-structural novelty — new patterns, combinations, and predictions within an existing conceptual framework — and strong, epistemic novelty — transformation of the conceptual framework itself — reveals a structural property of current and foreseeable algorithmic systems: epistemic conservatism. Drawing on Kuhn's theory of paradigm shifts, Peirce's logic of abduction, and Boden's taxonomy of creativity, the paper argues that algorithmic systems are structurally intra-paradigm tools capable of impressive weak novelty but not yet shown to perform the conceptual rupture that constitutes strong epistemic novelty. Three structural mechanisms — historical data dependency, surrogate optimization, and self-reinforcing feedback loops — explain why epistemic conservatism is systemic rather than incidental. Three prominent counterexamples (AlphaFold, AI-driven scientific discovery, emergent capabilities in large language models) are examined and shown to illustrate rather than refute the distinction.
Burhan Dinler (Sat,) studied this question.