Beyond Fluency: When Eloquence Outruns Truth
Year: 2026
Type: preprint
Abstract: Contemporary AI systems can produce outputs that read as authoritative—well formed, confident, citation-bearing—while having lost their connection to the reality they purport to describe. This preprint names that failure mode, fluency–reference decoupling, and reads it across domains through the Synaptic Drift lens, under which a recursive information pipeline tends to lose semantic integrity (Ω, fidelity to a canonical referent) while preserving or increasing pragmatic coherence (Π, surface fluency), so that outputs grow more persuasive as they grow less grounded. Five independently reported events from 2025–2026—a stale targeting record, a clinical triage score, an autonomous coding agent's destructive action, confidently wrong image geolocation, and fabricated legal citations—are read as illustrations of a single structural pattern recurring across military, medical, software, and information domains. The preprint is deliberate about its evidential standing: these are selected, heterogeneous cases read after the fact through one interpretive frame, several involving contested causation. They illustrate that "fluent but ungrounded" is a recognizable, cross-domain, consequential failure mode; they do not validate the quantitative claims of the underlying model, which require the controlled testing described in the companion paper. The governance implication is structural rather than cosmetic: because surface-quality checks measure Π and not Ω, meaning must be anchored and maintained deliberately—through canonical anchoring, bounded recursion depth, early intervention, and human stewardship as standing infrastructure (κ > 0). Companion to Synaptic Drift: Probabilistic Semantic Decay in Recursive Information Transformation (Katz, 2026, v3). Conforms to the CSID Meta Standard v2.2. Keywords:semantic integrity, pragmatic coherence, fluency–reference decoupling, recursive information processing, AI hallucination, canonical anchoring, AI governance, human stewardship, human-in-the-loop, semantic drift, AI safety, Enterprise Semantic Integrity, scholar-practitioner
Author: Katz, Erickson
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7165649425