Researchers have adapted epidemic models to study how misinformation moves online. A new study compared two such models using tweets about Mpox, finding that a model with an explicit stage for exposure before sharing can represent a delay that a simpler model leaves out. The authors present the work as a proof of concept, however, because the underlying data are too brief and uneven to support broad conclusions about misinformation dynamics.
Udoka Duruaku, Tolulope O. Agboola and Olalekan Ajao analyzed a public dataset of 5,787 Mpox-related tweets posted from July 7 to 13, 2022. Most records came from a single day, July 9. The dataset was assembled as an educational resource, and its misinformation labels were assigned by the dataset contributor through manual review, not by a panel of trained fact-checkers.
The team compared an SIR model, which groups people as susceptible, actively sharing, or no longer sharing, with an SEIR model that adds an “exposed” stage for people who have encountered misinformation but have not yet shared it. In this framework, “infectious” describes active sharing; the metaphor does not mean misinformation spreads biologically. The extra stage gives the model a way to represent a possible lag between seeing a post and sharing it.
The authors report that SIR captured rapid initial dissemination, while SEIR represented the exposure-to-sharing delay more fully. That is a difference in what the models can express, rather than direct evidence that users in this dataset followed a particular path. Passive exposure is difficult to observe in tweet records, and the study inferred it indirectly. The authors also note that the SIR model’s predicted peak lay beyond the seven-day observation period, so its simulation showed a rising phase rather than a measured full rise and decline.
That distinction matters because a compartment model can organize questions about online behavior without demonstrating that information moves like an infection. People can see a claim repeatedly, ignore it, or decide to share it; exposure is not equivalent to transmission. A model that includes a waiting stage can make that difference explicit, but a short record of posted tweets cannot show all the exposures or decisions behind those posts.
The dataset imposes further limits. It was a convenience sample, not a census of Mpox discussion on Twitter, and the concentration of activity on one day makes the apparent time trend difficult to interpret. The authors warn that model parameters are weakly identified: multiple parameter combinations can fit the sparse observations. They also used a homogeneous-mixing assumption, treating users as if they were equally likely to encounter and share content, though real social media networks are uneven and clustered.
The study’s value is therefore methodological. It illustrates how researchers might choose between a simpler model and one that represents a delay, while showing why longer, more balanced data and models that account for network structure are needed. The paper does not establish when an intervention should occur, predict a reliable misinformation peak, or show that one model is generally superior. Those questions need stronger evidence across time and outbreaks.