Year: 2026
Type: article
Abstract: ABSTRACT We model the LLM generated text as a complex phasor , with a grounded‐truth component and a generative‐fluidity component, drawing on the AC power identity . The resulting semantic power factor , , organizes hallucination and fabrication into four geometric quadrants and gives a single, phase‐derived readout for how a passage balances truth and fluency. We extend the framework in three directions: a two‐reference account of (external reality vs. locally‐entailed context ) that separates prompt‐originated error from generation‐time error; a sustained, device‐32‐style (reverse‐power relay) monitoring proposal for citation‐bearing spans, where fabrication is structurally well‐formed and thus falls in the same quadrant as deliberate genre fiction; and an instantaneous, device‐50‐style (overcurrent relay) check applied to prompt premises before generation, illustrated with a false‐arithmetic premise. Throughout, we present the model as an untested conceptual proposition, explicitly flag several unresolved interpretive tensions as open empirical questions and scope our strongest claims to reference‐checkable domains.
DOI: 10.1002/sres.70149
Author: Chaiya Tantisukarom
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7204114265