Correlated-signal extension for item-level clustering

Develop a correlated-signal extension of the sequential Bayesian cascade model that relaxes the conditional independence of users’ private impressions and accounts for the strong item-level clustering observed in the empirical data.

Background

The empirical analysis finds substantial between-item over-dispersion in agreement with the AI prediction, including rooms that cascade against the AI regardless of its correctness. Because the per-user behavioral model cannot generate this item-level clustering, the paper identifies a shared, item-level component of users’ impressions as a plausible missing mechanism.

The proposed extension would relax the model’s assumption that private impressions are conditionally independent given the story’s veracity. Such a model would allow correlated private signals and could explain residual dependence in the direction of cascades that is not captured by the current scalar public-log-odds process.

References

Per-user parameters cannot generate item-level clustering; it points to a shared, item-level component of impressions beyond the i.i.d. private-impression assumption of Section~\ref{sec:model}---a correlated-signal extension (which would relax the conditional independence that yields the scalar chain $(P_i)$) that we leave to future work and flag in Section~\ref{sec:discussion}.

— One AI Signal, Many Human Judgments: A Bayesian Cascade Analysis of AI-based Credibility Indicators in Online Information Spread  (2608.30311 - Lu et al., 31 Aug 2026) in Appendix, Section "Item-level clustering: a residual outside the per-user model"; referenced in Section 6, "Limitations and future work"