Abstract

This paper examines the epistemological challenges posed by large language mod- els (LLMs) through the metaphor of spatial disorientation. I argue that LLMs create what I call "semantic weightlessness" – a condition where linguistic fluency becomes decoupled from epistemic grounding, defined precisely as the absence of causal-historical chains connecting utterances to their truth-makers. Drawing on established theories of testimony, reference, and epistemic vigilance, I demonstrate how models like GPT-4, Claude, and Gemini generate texts that possess syntactic coherence and semantic plausibility while lacking genuine intentionality or truth- connection. Against emergentist interpretations, I show that statistical correla- tion, however sophisticated, cannot bridge the gap between pattern-matching and genuine understanding. I develop "semantic vigilance" as a virtue-epistemological framework, proposing concrete practices and institutional frameworks for navigat- ing machine-generated text. The paper contributes to philosophy of AI by: (1) providing a rigorous account of why LLM outputs constitute a novel epistemic cat- egory, (2) connecting AI epistemology to broader debates about extended cognition and distributed knowledge, and (3) proposing practical frameworks for institutional response.