Synthetic Reality and the Erosion of Epistemic Trust: Deepfake Misinformation, Confirmation Bias, and Democratic Cynicism Among Indonesian Voters

Abstract

Synthetic media generated by artificial intelligence (deepfakes) threatens democratic consolidation, yet most evidence is Western and video-centred and rarely models how such media shapes citizens' political attitudes. Grounded in epistemic-trust theory, motivated-reasoning theory, and the Liar's Dividend thesis, this study examined how the illusion of authenticity, confirmation-bias acceleration, and perception of the Liar's Dividend relate to democratic cynicism among Indonesian voters, and whether confirmation bias mediates and synthetic-media literacy moderates these relationships. A nested mixed-methods design combined computational extraction of 4,200 synthetic-media items (detector F1 = 0.92), a cross-sectional survey of 3,000 adults recruited by proportional random sampling across three urban centres in Indonesia, and digital ethnography with 50 key informants. The novel Epistemic Trust Erosion Scale demonstrated strong reliability (total alpha = 0.893; subscales 0.77 to 0.85) and sampling adequacy (KMO = 0.93). Cloned-audio deepfakes were trusted by 76.2% of respondents but consciously detected by only 11.5%, inverting the global assumption of video primacy. Illusion of authenticity (beta = 0.256), confirmation bias (beta = 0.269), and Liar's Dividend perception (beta = 0.212) each predicted democratic cynicism (all p < 0.001), explaining 33.0% of its variance (F(3, 2996) = 492.57, p < 0.001, f-squared = 0.493). Confirmation bias partially mediated the authenticity-cynicism pathway (indirect effect = 0.153, 95% CI [0.135, 0.172]). Synthetic-media literacy lowered cynicism but did not buffer the pathway. The findings extend deepfake scholarship to oral, messaging-centred Global-South publics and inform electoral-integrity policy.