Dierickx, L., Opdahl, A. L., Bjerknes, F., & Lindén, C. (2026). What is a fact in the age of generative AI? Fact-checking as an epistemological lens. Information, Communication & Society, 1–18. https://doi.org/10.1080/1369118x.2026.2630697
Summary
This theoretical paper asks what “a fact” means when generative AI produces text that is plausible yet untethered from empirical evidence. Using fact-checking as an epistemological lens, Dierickx and colleagues examine three established categories of facts — evidence-based (positivist), interpretative (constructivist), and rule-based (institutional) — and argue that none can adequately account for AI-generated content, which arises from probabilistic recombination rather than retrieval of verified information. To fill the gap they propose a fourth category, emergent facts, drawn from emergence theory and complex-systems thinking, and offer four indicators (accuracy, verifiability, contextual relevance, consistency) for evaluating such outputs. The argument reframes factuality as a dynamic sociotechnical outcome rather than a stable property of statements.
Key Contributions
- Introduces “emergent facts” as a novel epistemic category for GenAI outputs — computationally constructed, prompt-dependent, variable across models, and epistemically opaque.
- Provides a conceptual framework with four evaluative indicators: accuracy, verifiability, contextual relevance, and consistency.
- Distinguishes emergent facts from “algorithmic truth,” insisting on external validation over coherence or plausibility alone.
- Extends fact-checking epistemology from human-centered journalism into computational/AI contexts.
- Bridges philosophy of science, AI ethics, and media/information studies by treating factuality as a relational, dynamic outcome.
Methods
Conceptual and theoretical rather than empirical. The authors synthesise philosophy of science (correspondence, coherence, and pragmatic theories of truth), the sociology of knowledge (Durkheim, Searle, Latour & Woolgar, Foucault), and journalism/fact-checking scholarship. They map the three conventional fact categories against the characteristics of large language models, then construct the “emergent facts” framework and its four indicators, summarised in a table and figure.
Findings
- Positivist/evidence-based fact-checking fails for GenAI because outputs lack ontological referents and traceable provenance.
- Constructivist/interpretative approaches capture the negotiated nature of facts, but GenAI outputs emerge from opaque probabilistic processes rather than shared social practices.
- Institutional/rule-based facts rely on collective agreements and verification protocols that GenAI mimics linguistically without genuinely following.
- Emergent facts occupy a distinct category: plausible but unverified, context-dependent, highly variable, and not reducible to brute facts, thus requiring relational rather than binary true/false assessment.
- The four indicators offer a structured way to evaluate outputs while addressing hallucination and bias.
- The framework prioritises evaluation and AI literacy over technical intervention.
Connections
This paper’s epistemological reframing of factuality complements empirical and design work on generative AI in fact-checking and verification, such as Gilardi2026-hw and Suau_Martinez2026-lv, which examine how AI is actually deployed in fact-sensitive contexts. Its emphasis on AI literacy and evaluating opaque GenAI outputs also speaks to concerns about AI-generated persuasion and disinformation raised in Hackenburg2025-dj and DeVerna2025-dl. The others under these topics address related but distinct empirical questions rather than the epistemology of factuality per se.
Podcast
A research-radio episode discusses this paper: 🎧 MP3 · Spotify · Apple Podcasts