Abstract Most debates on AI-generated language focus on accuracy and hallucinations, implicitly treating fluent and plausible statements as if they already qualified as knowledge. This paper offers a different diagnosis: AI-generated statements typically remain in a pre-verificational state, not false knowledge, but unfinished claims that have not yet entered the epistemic conditions under which knowledge can be tested and held accountable. Knowledge is not constituted solely by truth value but also by traceability of liability: when a claim fails, it must be localizable to sources, reasons, and responsibility-bearing agents. Because AI systems lack legal and moral personhood, their outputs prior to human intervention create an accountability vacuum. I argue that sentences are the minimal unit at which claims crystallize and responsibility can attach; paragraph-level evaluation invites semantic smuggling, where unfalsifiable bias or causal insinuation hides inside otherwise acceptable prose. While recent work has advanced sentence-level attribution, these techniques are best understood as enabling infrastructure rather than substitutes for human verification. I conclude by defining Traceability-First Language Models as an epistemic design principle that preserves human judgment by making verification possible at the sentence level. Key words : AI epistemology; Pre-verificational state; Sentence-level traceability; Knowledge verification; Accountability in AI; Explainable AI
Isabella Langwei (Wed,) studied this question.