A race to belief: How Evidence Accumulation shapes trust in AI and Human informants
Johan Sebastián Galindez-Acosta, Juan José Giraldo-Huertas
TL;DR
The paper investigates why people trust AI versus human informants differently across factual (epistemic) and interpersonal (social) contexts. Using a Bayesian hierarchical drift-diffusion model (HSSM), the study shows that context-dependent trust is driven primarily by drift-rate differences in evidence accumulation (negative $v$ for AI in epistemic tasks; positive $v$ for humans in social tasks), with near-neutral starting bias and consistent decision boundaries. A strong link between drift rate and moment-to-moment confidence supports the mechanistic interpretation of trust dynamics, revealing how rapid, domain-specific processing can lead to brittle AI trust. These findings have implications for AI governance, suggesting transparency and vigilance features that sustain calibrated trust without sacrificing efficiency.
Abstract
The integration of artificial intelligence into everyday decision-making has reshaped patterns of selective trust, yet the cognitive mechanisms behind context-dependent preferences for AI versus human informants remain unclear. We applied a Bayesian Hierarchical Sequential Sampling Model (HSSM) to analyze how 102 Colombian university students made trust decisions across 30 epistemic (factual) and social (interpersonal) scenarios. Results show that context-dependent trust is primarily driven by differences in drift rate (v), the rate of evidence accumulation, rather than initial bias (z) or response caution (a). Epistemic scenarios produced strong negative drift rates (mean v = -1.26), indicating rapid evidence accumulation favoring AI, whereas social scenarios yielded positive drift rates (mean v = 0.70) favoring humans. Starting points were near neutral (z = 0.52), indicating minimal prior bias. Drift rate showed a strong within-subject association with signed confidence (Fisher-z-averaged r = 0.736; 95 percent bootstrap CI 0.699 to 0.766; 97.8 percent of individual correlations positive, N = 93), suggesting that model-derived evidence accumulation closely mirrors participants' moment-to-moment confidence. These dynamics may help explain the fragility of AI trust: in epistemic domains, rapid but low-vigilance evidence processing may promote uncalibrated reliance on AI that collapses quickly after errors. Interpreted through epistemic vigilance theory, the results indicate that domain-specific vigilance mechanisms modulate evidence accumulation. The findings inform AI governance by highlighting the need for transparency features that sustain vigilance without sacrificing efficiency, offering a mechanistic account of selective trust in human-AI collaboration.