Diverse Misinformation: Impacts of Human Biases on Detection of Deepfakes on Networks

TL;DR

The paper tackles how human biases shape susceptibility to deepfake misinformation and the capacity of crowds to self-correct on networks. It combines an observational survey (N=2,016) with non-primed viewing of deepfake videos and a stylized heterogeneous network model with duped-rate $\lambda_i$ and correction-rate $\gamma$ to project population-level dynamics, using an MMSBM-inspired structure. Key findings show near-random detection without priming ($MCC$ around $0.334$ overall) but clear demographic-specific patterns (e.g., homophily effects where matching demographics improve accuracy), suggesting herd correction can emerge in diverse networks. The work highlights implications for interventions that leverage social diversity and human judgment alongside automated detection, and it provides a framework for further empirical validation and network-level studies of diverse misinformation.

Abstract

Social media platforms often assume that users can self-correct against misinformation. However, social media users are not equally susceptible to all misinformation as their biases influence what types of misinformation might thrive and who might be at risk. We call "diverse misinformation" the complex relationships between human biases and demographics represented in misinformation. To investigate how users' biases impact their susceptibility and their ability to correct each other, we analyze classification of deepfakes as a type of diverse misinformation. We chose deepfakes as a case study for three reasons: 1) their classification as misinformation is more objective; 2) we can control the demographics of the personas presented; 3) deepfakes are a real-world concern with associated harms that must be better understood. Our paper presents an observational survey (N=2,016) where participants are exposed to videos and asked questions about their attributes, not knowing some might be deepfakes. Our analysis investigates the extent to which different users are duped and which perceived demographics of deepfake personas tend to mislead. We find that accuracy varies by demographics, and participants are generally better at classifying videos that match them. We extrapolate from these results to understand the potential population-level impacts of these biases using a mathematical model of the interplay between diverse misinformation and crowd correction. Our model suggests that diverse contacts might provide "herd correction" where friends can protect each other. Altogether, human biases and the attributes of misinformation matter greatly, but having a diverse social group may help reduce susceptibility to misinformation.

Figures (6)

• Figure 1: Illustration of the problem considered in this work. Populations are made of individuals with diverse demographic features (e.g., age, gender, race; here represented by colors), and misinformation is likewise made of different elements based on the topics they represent (here shown as pathogens). Through their biases, certain individuals are more susceptible to certain kinds of misinformation. The cartoon represents a situation where misinformation is more successful when it matches an individual's demographic. Red pathogens spread more readily around red users with red neighbors, thereby creating a misinformed echo chamber whose members can not correct each other. In reality, the nature of these biases is still unclear, and so are their impacts on online social networks and on the so-called "self-correcting crowd".
• Figure 2: A confusion matrix showing our participant guesses about the state of the videos vs. the real state of the video. Participants in our study watched two videos followed by a questionnaire and a debriefing on deepfakes. They were then asked to guess whether the videos were deepfakes or real. Out of 2,016 participants and 4,032 total videos watched, 1,429 videos duped our participants, meaning they saw a fake video they thought was real. The top right panel shows the participants who were duped by deepfakes. The confusion matrix is defined by the number of true positives in the top left, false negatives in the top right, false positives in the bottom left, and true negatives in the bottom right.
• Figure 3: Question where survey participants are asked after the debrief of the survey if they think the videos they watched are real or fake. The performance metric we use to measure participant accuracy is the ratio of the correct guesses to the entire pool of guesses where accuracy = (True Positive (TP) + True Negative (TN))/(True Positive (TP) + False Positive (FP) + False Negative (FN) + True Negative (TN)).
• Figure 4: Bootstrap samples from observed confusion matrices to compare MCC scores of user and video feature pairs. Categories that satisfy a threshold of credibility above 95% are as follows (all bootstrap samples can be seen in the Supplementary Information. (a) White users were found to have a homophily bias and are better at classifying videos of a persona they perceive as white. (b) Consequently, videos of personas of color are more accurately classified by participants of color. (c) Similarly, videos of male personas are better identified by male users. Across multiple age classes, we find that participants aged 18-28 years old are better at identifying videos that match them than older participants (panels (d) and (e)) or even better at classifying videos of persona perceived as 30-49 years old than participants from that same demographic (panel (f)). In addition we reproduce our findings from bootstrapping and conduct a Bayesian logistic regression to explore the effects of matching demographics on the detection accuracy which can be seen in our Supplementary Information
• Figure 5: Spread of diverse deepfakes with heterogeneous transmission rates $\lambda_i$ across demographic types 1 and 2 (in-group density is set to $Q=0.75$, degree heterogeneity to $\alpha = 3$). Other parameters are given in the figure, with panels (b) and (c) using the correction rate highlighted in (a) at 1.7. Panel (c) shows how high degree nodes can be protected if they have a diverse set of neighbors.