From sharing propensity to diffusion control: temporal catch-up costs of delayed misinformation correction

Introduction Why do corrections underperform misinformation at the network level? We distinguish a correction-sharing disadvantage (lower willingness to transmit corrections) from a correction-timing disadvantage (corrections enter after misinformation gains a first-mover advantage). Methods Study 1 compared self-reported forwarding likelihood across misinformation, matched detailed corrections, and corrections presented after instructions to assume prior misinformation forwarding among 266 Chinese participants using three health-message pairs. Study 2 used the ratings as a relative propensity profile, not calibrated platform probabilities, in a competitive agent-based model across Barabási–Albert, Erdős–Rényi, and Watts–Strogatz networks. We varied transmission intensity, onset delay, seed supply, and placement. Temporal catch-up cost (TCC) was the smallest tested correction-to-misinformation seed ratio restoring the immediate equal-seeding cumulative-coverage reduction, minus 1. Results Under the present asymmetrically labeled conditions, direct corrections elicited higher stated forwarding likelihood than misinformation (mean difference = 1.24, 95% CI [0.52, 1.96], Hedges’ g = 0.63), whereas assumed prior forwarding did not significantly reduce correction forwarding (mean difference = 0.35, 95% CI [−0.37, 1.07]). Across three regimes, all 18 equal-seeding delay penalties were positive and increased with delay and transmission intensity. After one generation, all nine regime-topology combinations recovered the benchmark within the grid (TCC = 0.80–7.60); after two generations, outcomes ranged from finite recovery to failure within the grid and, for BA under the reference regime, model-implied unattainability. Reference-regime degree targeting especially reduced BA catch-up requirements. Discussion These hypothetical intentions are susceptible to social desirability and the intention–behavior gap; because labeling differed across conditions, they do not isolate correction content from disclosed veracity or corrective framing. Correction difficulty can therefore emerge despite higher stated correction-forwarding likelihood under disclosed labels; the thresholds are conditional model comparisons, not platform-specific resource or clock-time estimates.