An investigation into LLMs’ abilities in discerning between true and false information

Abstract

Large language models (LLMs) may judge whether information appears credible without necessarily establishing whether it corresponds to external facts. This study investigates this distinction by operationally separating Semantic Truth (ST), defined as the correspondence of textual claims with external states of affairs, from Epistemic Truth (ET), defined as the credibility or justification conveyed by a text through coherence, plausibility, evidential presentation, and consistency. The dataset comprised 274 source texts, including newer BBC and CNN articles, older CNN articles, and historical articles, from which controlled variants differing in factual accuracy and presentation were generated. In Phase 1, a single LLM reliability score remained relatively high even as factual accuracy decreased, with completely fabricated texts still receiving mean scores above 3 on a 1–5 scale. In some cases, the model also assigned high numerical reliability despite identifying substantial factual problems in its written justification. Phase 2 separately evaluated ST and ET across 548 authentic and fabricated observations. ST provided stronger discrimination between authentic and fabricated texts than ET, achieving an overall AUC of 0.801, sensitivity of 0.766, and specificity of 0.810. However, semantic discrimination varied markedly with information familiarity, ranging from near-chance performance for newer CNN articles to nearly perfect discrimination for older and historical material. These findings demonstrate that targeted semantic prompting improves factual discrimination but does not fully separate semantic correspondence from information familiarity, plausibility, and other non-factual textual cues. More broadly, the ST–ET framework exposes a potentially important form of truth inflation: epistemic credibility may remain high as semantic correspondence is progressively degraded through increasing fabrication. This provides a basis for future studies to determine how far epistemic credibility can be sustained or inflated as factual grounding deteriorates, thereby defining and quantifying an LLM’s tolerance for increasingly plausible fabrication.