What is a fact in the age of generative AI? Fact-checking as an epistemological lens

Summary

This conceptual paper uses fact-checking as an epistemological lens to interrogate what counts as a “fact” when generative AI produces fluent, plausible, but unverified content. Dierickx and colleagues argue that the three fact categories implicit in fact-checking practice — evidence-based (positivist), interpretative (constructivist), and rule-based (institutional) — each fail to capture the epistemic status of LLM outputs, which are probabilistic recombinations without ontological referents, traceable provenance, or genuine adherence to verification norms. They propose a fourth category, emergent facts, drawn from emergence theory and complex-systems thinking, and offer four indicators — accuracy, verifiability, contextual relevance, and consistency — for evaluating such outputs relationally rather than through binary true/false judgments.

Key Contributions

• Introduces emergent facts as a new epistemic category for AI-generated content: plausible, prompt-dependent, variable, and opaque.
• Provides a four-indicator evaluative framework (accuracy, verifiability, contextual relevance, consistency) for assessing GenAI outputs.
• Distinguishes emergent facts from algorithmic truth: the former demands external validation; the latter relies on internal coherence or plausibility.
• Extends fact-checking epistemology from journalism into computational and AI contexts.
• Bridges philosophy of science, AI ethics, and media studies, reframing factuality as a sociotechnical outcome.

Methods

Purely theoretical and conceptual. The authors synthesize literature from philosophy of science (correspondence, coherence, pragmatic theories of truth), sociology of knowledge (Durkheim, Searle, Latour & Woolgar, Foucault), and journalism/fact-checking scholarship. They map each of three established fact categories against the characteristics of LLMs, then build a new conceptual framework summarized in a comparative table and a diagram.

Findings

• Evidence-based fact-checking fails for GenAI because outputs lack referents and verifiable provenance.
• Interpretative/constructivist framings illuminate negotiation of facts but cannot account for opaque probabilistic generation absent shared social practice.
• Rule-based/institutional facts depend on collective protocols that LLMs only mimic linguistically.
• Emergent facts form a distinct epistemic category: computationally constructed, context-dependent, and irreducible to brute facts.
• The four indicators provide a structured way to evaluate hallucination, bias, and contextual drift without collapsing into binary verification.
• The framework foregrounds evaluation and AI literacy over purely technical mitigation.

Connections

This piece sits among other theoretically and normatively oriented work on factuality and AI-mediated information, complementing empirical studies of LLM-driven fact-checking such as DeVerna2025-dl and Cazzamatta2026-lo, and conceptual treatments of misinformation epistemics like van-der-Linden2026-jt and Marwick2025-ov. Its concern with how generative outputs destabilize verification norms also resonates with Starbird2025-jj on contested knowledge production and with Triedman2025-uy and FitzGerald2025-nv on the epistemic risks of LLM-generated content.

Podcast

A research-radio episode discusses this paper: Listen