Intolerance in Social Media

Payal Dey

(pdey6677@gmail.com)

Social media has become one of the primary means of consuming information online. However, the barriers to creating content on social media are low, which has given rise to a wave of misinformation. As a reaction, social media platforms have pledged to contain the spread of fake news.

One common strategy is to use crowd-sourced flagging systems: users will flag news items, and those receiving more than their fair share of flags will be fact-checked by experts. This system works in theory, but the objectivity of users in evaluating the truthfulness of a news item has been called into question. Phenomena like confirmation bias echo chambers —although the magnitude of their effect has been called into question—trust in the source, or lack thereof, and information overload might pollute such objectivity for a variety of different reasons that are not necessarily only ascribed to social identity. As a result, users might be prone to give a pass to misinformation confirming their world view while at the same time being excessively zealous against news challenging their opinion. In a previous study, we developed an agent-based model where this potential differential treatment is taken into account. In such a model, the crowd-sourced flagging systems produce counterproductive results. Most flags end up being assigned to truthful and neutral news sources, while producers of polarising misinformation are barely flagged.

In this paper, we take our agent-based experiment one step further. We are interested in the effect of differential tolerance. In the original model, all users had the same level of tolerance, which determines how far from their worldview a news item must be to earn a flag from them. While it is true that both liberals and conservatives are prone to characterise opposing points of view as fake news, it is questionable whether they do so at the same level of ideological distance. In fact, there is evidence that some portions of the opinion spectrum might be less tolerant than others—e.g., one side having more and stronger reasons to tolerate opposing points of view; media on one side of the spectrum using stronger outrage language than the other; or users reacting more or less strongly when exposed to opposing points of view. We start by assigning a lower tolerance to agents on one side of the polarisation spectrum. In such a scenario, news sources on the opposite side of the intolerant users receive more flags. In an extension of the model where we allow sources to react to these flags by changing their polarity to minimise the number of flags received, we see that sources are attracted to move towards the polarisation values of the intolerant users. This result is less trivial than might appear at first sight: maximally intolerant users flag everything near their position, thus they are repulsive for everything that is not exactly conforming to their polarity. In fact, in our model, the tolerance sweet spot is different from zero.

Moreover, tolerance is not an absolute and immutable quality of an individual but can change in different contexts, for instance, by increasing when talking in abstract terms but decreasing when facing concrete examples. In other research, greater democratic activism is linked with an increase in political tolerance. In a second extension of our model, we hypothesise that users on social media might copy the low-tolerance strategy of users on the other side of the polarity spectrum in a form of retaliation—which is a classic game-theoretic strategy. There is research supporting retaliation as a realistic potential mechanism: male group members are more likely to retaliate against an outgroup if the outgroup makes them question their own identity (such as in a political debate).

Intergroup anger is a group-level emotion that predicts the desire of the individual to harm a threatening outgroup as a whole. This is a group-level example of appraisal theory, which shows that a person with a strong perception of their own self would tend to retaliate against other individuals who threatened that self. In this case, it works at a group level when group affiliation is incorporated into the image of the self, as is the case for many groups. This process of group integration in a self-image is known and studied as “self-categorization theory”. Besides literature-backing, we can find examples of opposite ends of the political spectrum copying each other’s strategies as a form of retaliation. For instance, the derogatory term “snowflake” has been widely used by conservatives to mock the variety of issues triggering a strong emotional response from liberals, but has been quickly retorted against conservatives exploiting their own triggering issues (date of access: November 16th, 2021). A similar fate occurred to the “Make America Great Again” (MAGA) meme, accusing the utgroup of having caused the downfall of a country, a fall that can be reverted by strong actions from the ingroup. Originally a campaign slogan for Donald Trump, it received a response with the same underlying message (“We just did”) from the Biden campaign. In another example, brigades of alt-right social media users weaponized the methods of cancel culture—through which many conservatives had been asked or forced to step down from their positions due to claims from liberals of racism, sexism, and homophobia—to ostracise liberal celebrities. One example is the firing of movie director James Gunn. In our models, this retaliation process creates downward tolerance spirals: a tolerant society in our model is out of equilibrium and will settle on a significantly lower average tolerance level as a result of users attempting to attract news sources on their side by exploiting the flagging system.

Again, this tolerance is not zero, confirming that there are more complex dynamics at play than simply minimising tolerance to maximise the number of flags assigned to opposing points of view. Interestingly, our model settles this equilibrium tolerance in an interval that is empirically supported: the equilibrium parameter range is included in the range that is the best at reproducing the relationship between a news source’s popularity and the number of flags it receives on Facebook, as described in previous work. This study is fully based on simulations on an agent-based model; thus, its conclusions should be verified with empirical experimentation in future works. However, ABMs have been successfully applied to social media polarisation studies in the past and have proven their usefulness. Our ABM is designed to capture the most salient characteristics of social media information consumption: echo chambers; selective exposure; confirmation bias; realistic distributions of users, source polarity, and the popularity of news sources; realistic topology for the social network among users with communities, high clustering, broad degree distributions, and the small world effect. We support our choices by identifying the relevant pieces of literature in the sections describing the model.