Exploring Cognitive Bias Triggers in COVID-19 Misinformation Tweets: A Bot vs. Human Perspective

TL;DR

This study investigates how cognitive bias triggers are used in COVID-19 misinformation on Twitter by comparing Bot-authored and Human-authored tweets. It introduces the Misinfo Dataset, assembling $1,768,838$ Bot tweets and $1,782,122$ Human tweets from July 2020–July 2021, and a computational pipeline to detect eight bias triggers across three judgment heuristics. The analysis shows Bots employ bias triggers more extensively and across a wider set of biases, with Availability Bias, Cognitive Dissonance, and Illusory/Confirmation biases most prominent, and with certain triggers boosting Bot engagement while others suppress it; Human engagement shows weaker and less consistent associations. These findings highlight how automation leverages bias cues to amplify misinformation and offer a methodological basis for countermeasures, including bias-trigger detection and Bot identification, to mitigate misinformation spread in public health contexts. The work underscores the need to account for platform-driven amplification and informs strategies for misinformation management and digital literacy.

Abstract

During the COVID-19 pandemic, the proliferation of misinformation on social media has been rapidly increasing. Automated Bot authors are believed to be significant contributors of this surge. It is hypothesized that Bot authors deliberately craft online misinformation aimed at triggering and exploiting human cognitive biases, thereby enhancing tweet engagement and persuasive influence. This study investigates this hypothesis by studying triggers of biases embedded in Bot-authored misinformation and comparing them with their counterparts, Human-authored misinformation. We complied a Misinfo Dataset that contains COVID-19 vaccine-related misinformation tweets annotated by author identities, Bots vs Humans, from Twitter during the vaccination period from July 2020 to July 2021. We developed an algorithm to computationally automate the extraction of triggers for eight cognitive biase. Our analysis revealed that the Availability Bias, Cognitive Dissonance, and Confirmation Bias were most commonly present in misinformation, with Bot-authored tweets exhibiting a greater prevalence, with distinct patterns in utilizing bias triggers between Humans and Bots. We further linked these bias triggers with engagement metrics, inferring their potential influence on tweet engagement and persuasiveness. Overall, our findings indicate that bias-triggering tactics have been more influential on Bot-authored tweets than Human-authored tweets. While certain bias triggers boosted engagement for Bot-authored tweets, some other bias triggers unexpectedly decreased it. Conversely, triggers of most biases appeared to be unrelated to the engagement of Human-authored tweets. Our work sheds light on the differential utilization and effect of persuasion strategies between Bot-authored and Human-authored misinformation from the lens of human biases, offering insights for the development of effective counter-measures.

Figures (6)

• Figure 1: Overview of Dataset Formation Methodology. This figure illustrates the pipeline of data collection and filtering. The data was collected when the platform was named "Twitter".
• Figure 2: Distribution of the Bias Triggers in Tweets. This illustrates the percentage of tweets that attempted to trigger cognitive biases by two different user types. For example, 62.08% of Bot tweets triggered Availability Bias, and 34.41% of them triggered Cognitive Dissonance
• Figure 3: Co-Occurrence of Bias Triggers in Misinfo Tweets. The heatmap is colored according to the prevalence of triggers of two biases occurring in the same misinformation tweet.
• Figure 4: Engagement by Bias Triggers. This presents the proportion of engagement from tweets associated with triggers of a given bias over the total engagement from all tweets authored by Bots and Humans, respectively. For example, Bot tweets with Affect/Negativity Bias triggers receive 29.95% of the total favorite counts in the Misinfo Dataset.
• Figure 5: Relation between the Number of Biases Triggered and Engagement This figure illustrates the relationship between the number of biases triggered and the engagement of tweet.