Zhuoran Lu, Weilong Wang, Yangyang Yu, Xinru Wang, Zhuoyan Li, Zhiwei Liu, Sophia Ananiadou
22 pages, 20 figures. Accepted at HCOMP 2026 (2026 ACM Conference on Human-AI Complementarity and Alignment). Supplementary material included as appendices
When an AI credibility signal is shown to every user in a sequential judgment setting, crowd agreement reflects dependence on a shared AI signal rather than independent human evidence, producing a preservation-correction trade-off where stronger AI reliance preserves correct predictions but also locks in wrong ones by blocking private corrective information.
The authors extend the classical Bikhchandani-Hirshleifer-Welch Bayesian cascade model by adding a shared AI signal, deriving a 'Gateway' condition comparing AI accuracy (alpha) to human private-impression accuracy (q). They calibrate the model on human-subject data from a prior news-veracity judgment experiment where the AI outperformed humans, fitting both fully Bayesian benchmarks and behavioral weights, plus mixture models across subjects and simulation sweeps over true vs. perceived AI accuracy.
The average user weights the AI signal equivalently to 3 to 6 peer judgments, placing her in the learning regime (below the Gateway, alpha < q behaviorally), so private impressions still enter the record. A two-component mixture shows a majority who under-weight the AI and never cascade on it alone, and a cascade-prone minority who over-weight it enough to cascade immediately. Simulations indicate over-reliance on a weak AI is more damaging than under-reliance on a strong one, and that diversifying AI signals across users is more robust than expert-review triggers.
The authors concede the calibration reuses a prior human-subject dataset rather than new experiments, the model assumes binary veracity, symmetric priors, independent private signals, and a fixed single AI prediction, and the intervention results (diversification, expert review) come only from simulations beyond the calibration data. Unstated limits: the behavioral model is fitted on one experimental setting with news headlines, real platform dynamics (votes, shares, algorithmic ranking) are not modeled, and the AI's accuracy and its perception by users were controlled inputs rather than organic conditions.
This is a stylized theory paper calibrated on one lab experiment with a small subject pool, not evidence that real misinformation dynamics on platforms behave this way. The Bayesian cascade framework's assumptions (independent signals, common knowledge of accuracies, binary states) are strong, and the behavioral estimates could be prompt- or interface-specific. The claimed policy lever (diversifying AI signals) is demonstrated only in simulation with no field validation.
Little or nothing yet for AI economics. The paper speaks to platform trust-and-safety design and misinformation governance; if platform-deployed AI indicators become widespread, its warning that a shared AI signal can suppress crowd independence could eventually matter for how platforms justify AI moderation and how regulators assess it, but that is a design implication, not a measurable economic effect in the next 12 to 24 months.
Societal & Cultural: Anyone worried about misinformation ecosystems should care that a single AI label can convert apparent crowd consensus into dependence on one model, weakening the crowd's corrective capacity precisely when the model is wrong.
Regulatory & Legal: Regulators considering mandates or standards for AI-generated content labels face a new trade-off here: the same intervention that protects correct judgments also entrenches wrong ones, so design rules (e.g., signal diversification) may matter more than accuracy targets.
Technical Capability: Deployers of AI credibility systems learn that accuracy gains alone do not fix the failure mode; the cascade structure means even a strong AI's errors become harder for the crowd to correct as reliance rises.