Facticity Without Redeemable Validity - Knowledge, Turned Through the Kaleidoscope — A Habermasian Diagnosis of Frontier AI and the Fracture of Governance

This paper advances a claim about the current condition of knowledge under frontier AI systems, using Habermas's distinction between facticity and validity (Faktizität und Geltung) as its primary theoretical anchor, the present author's four-conditions account of knowledge integration as a second, independent diagnostic layer, and an empirically logged instance of cross-model context contamination as a third, concrete illustration of why the resulting governance requirement is agency rather than verification. In Habermas's account, modern law and legitimate normative order depend on facticity and validity remaining co-original: a norm's factual, coercive force is only legitimate insofar as its validity claim remains redeemable through rational discourse. The claim advanced here is threefold. First, frontier AI systems generate facticity — output with real, coercive, behaviour-structuring force — while the discourse process that is supposed to redeem the corresponding validity claim is structurally broken or absent, because no stable content-core stands behind the resolved output. Second, and independently, even where content does exist, it may fail to achieve knowledge integration — semiotic, domain, organisational, and human — across the boundary that would make it available to discourse at all, a precondition Habermas's framework requires but does not itself supply. Third, a logged empirical instance of context silently crossing a boundary between two distinct AI systems shows why automating the verification of one AI system's output by another cannot close either gap: it relocates the same unredeemed facticity one level up, dressed as oversight. This mismatch, it is argued, is not a drafting failure correctable by better rules of the same kind; it is the collapse of a specific, load-bearing co-originality thesis that governance theory silently depends on, compounded by an unexamined precondition that theory has never named, and resolvable only by requiring findable, answerable human agency rather than more verification. Each governance camp is arguing as though facticity and validity still cohere and as though the content behind AI output is automatically integration-ready; neither assumption holds, and neither failure has yet been named as the actual site of disagreement.