The Social Dynamics around Misinformation Spreading on Social Media

Mag. Dr. Hannah Metzler
Complexity Science Hub & Medical University of Vienna

Digital Humanism Lecture Series
13 November 2024

Slides: https://hannahmetzler.eu/digum | Recording on Youtube

Digital Humanism & social media

Shaping digital technology in line with human values & needs

Psychology & computational social science:

• Designing and using social media in ways that foster individual and societal well-being.

  • mental health
  • extremism & polarization
  • echo chambers & misinformation

• Core of DigHum for me: shaping
• In my research area of psych/CSS on the topic of social media, this means…
• You’ve probably all heard about these fears people have about social media…
• Controversial discussions

Time: If I stick to slide content (no story add ons), then 1 slide 2 minutes is fine. Duration: 30 slides, 41 minutes without questions = 82 seconds/slide (1.4 min)

Slide count: 30 = 24 section titles: -3 recommendations: - 2 Thank you: 1

Challenges for Digital Humanism in Computional Social Science

• Complex social issues with many influencing factors
  • Intuition vs. evidence
  • What are the root causes?
  • Which solutions address these?
• Social media companies:
  • Transparency & cooperation

• Talk about the challenges of “Digital Humanism” from their point of view

Crucial methods and skills

• Strong theoretical & multidisciplinary background:
  • What does solid/total evidence show?
• Strong statistics skills: Real-world data & reproducibility
• Awareness of cognitive & social biases
• Transparency & Open Science
• Big data analysis (text, video…), machine learning, LLMs etc.
• Complexity science approaches:
  • E.g. Network analysis, Agent-Based Modeling

• Complex social issues: which evidence is solid, all evidence together?
• Requires statistics skills, messy data, and many non-reproducible studies
• Because we have own intuitions, and problems are complex: biases.
• Willingness to question assumptions
• For all this, Open Science really helps: making hypotheses, analyses plans, code, data openly available.
• Methods
• Now, as an example of research from this area, let me come to your project on misinformation.

Social dynamic around
misinformation on Twitter

Why the social dynamic is key

• Most interventions focus on information & accuracy
• Beliefs are shaped by our social identity
• Emotions attract attention & increase sharing
• Group & emotion dynamics are central on social media
• Social media algorithms mirror these human patterns

why not just teach humans the facts? Add ref van bavel etc.

Our project on misinformation

Observational social media analyses of emotions

Experiments: Psychological interventions

Agent-based modelling: network & algorithms

Today, I will show you results from 3 different parts of our project

Emotions around news posts on Twitter/X

• Posts with news URLs & threads 2020-2022
• Emotions: Machine learning
• Trustworthiness: NewsGuard

\(\rightarrow\) Untrustworthy sources: More anger

• If we look in the “real world”, where misinfo occurs on social media
• there indeed observe emotions around news, in particular anger

What is the origin of emotions around (untrustworthy) news?

One main question I want to look at with you today is… - in the person (mood)? - From the news? - What about the news: wrong, or emotional/sensationalist?

Emotions: Mood vs. response to news

• Experiment with COVID-19 news headlines
• Austria 2021
• No effect of mood

• Response to false news:
  • More anger
  • Less positive

What about false news causes the emotion?

1. Causal effects of misinformation
   • untrustworthy compared to trustworthy information
2. Effects of emotions in the news
   • compared to emotionally neutral news

\(\rightarrow\) Social media data analysis by Jula Lühring

Jula image

1) Causal effects of misinformation

• Matching news posts to create 2 similar groups
  • Emotions in the news, political orientation, followers, …
• Comparison: Unique effect of untrustworthy news sources

• Tweets assigned to 2 conditions based on similarity on covariates:
  • Political orientation, followers, following, time from post to response, emotions in post with the news
• Comparing conditions: effect of misinformation on emotions

Small but causal effects of misinformation on emotions & engagement

News posts with untrustworthy vs. trustworthy URLs

• 3% more angry posts

• 0.4% more disgust posts

• 0.2% more fear posts

• 0.4% less joy posts

2) Effects of emotions in the news post

\(\rightarrow\) Emotions in discussion mirror emotions in news post

• Strongest predictor of anger in the discussion:
  • Anger in the news post (influence of external events)
• Same pattern for all other emotions

• Compared to other emotions and trustworthiness

• EU registration of the first COVID-19 vaccines

• enforcement of measures against the spread of COVID-19 in December 2020 Floods in Germany in July 2021

Anger around misinformation

• Emotions in discussions largely reflect emotions in initial post and less its trustworthiness

• More anger in the context of misinformation

Why are people angry? Does anger make people easier to manipulate?

\(\rightarrow\) Results from COVID-19 news headlines experiment

Why were people angry?

• Higher anger in people good & bad at recognizing false news
• Angry responses because most people recognize false news
• Anger arose when information contradicts existing beliefs

\(\rightarrow\) Emotion \(\neq\) Manipulation

in other words: emotions don’t make people gullible

Who believes and shares most misinformation?

This brings me to another crucial question: …

• In our study: Educated, left-leaning sample
• majority did not believe COVID-19 false news
• But other studies? Social media?

Motivated minority with extreme opinions

• Minority of 0.1%, 1%, 2%, 2107 US voters in 2020 election

• Reads and shares ~80% of misinformation

• Supersharers on Facebook USA, 2020:

• 65% republicans
• 60% women
• ~60 years old
  
• 90% white

Misinformation as symptom of polarization

• Trust in institutions:
  • Not feeling represented
  • Frustration with real-world societal problems
  • Corruption => more conspiracies
• Polarization of elites plays an important role (2, 3, 4)

So how much misinformation is there?

• Definition: News from untrustworthy outlets
• Small part of news diet on social media
  • 1-6% in 5 studies from 2016-2019
  • COVID-19 pandemic 2020-2022 \(\rightarrow\)
    Lühring, Garcia, Waldherr, Shetty, & Metzler (in prep.)
  • Posted by US politicians 2010-2023:
    3.92/1.25% (Rep/Dem)
    Shetty, Simmerdinger, Lühring, Garcia, Walherr
    & Metzler (in prep.)

What can we do?

• People not gullible & beliefs socially motivated
• Motivated minorities: hard to reach & convince

Our approach

1. Psychological interventions as prevention: less polarized beliefs & groups
2. Algorithmic interventions: misinformation by politicians

Results for number 1 Ongoing research for number 2

Changing attitudes via social identity

• Intervention that associates a (misinformed) belief with an incongruent social identity
• Complementary & Alternative Medicine (CAM) vs. right-wing extremism
• How: Intervention provides
  • historical information
  • news articles

Intervention effect: Exp. vs. Control group

• Pre-registered study with pre-post design (n=817)

• Highlighting inconsistency of beliefs with social identity influences attitudes

control: history & diversity of a plant, origin of cabbage/broccoli/cauliflower receipes

Algorithmic interventions

• Targeting political elites:
  • Even small effects can have large effects on exposure.
• Agent-Based Model = Computer simulation
• Model of US politicians’ retweeting behavior
  • news trustworthiness
  • emotions
  • social groups (parties)

Our model matches retweeting behavior

Can algorithmic interventions reduce/boost the sharing of untrustworthy/trustworthy news by politicians and partisans?

Digital humanism, social media & misinformation?

• Small effects in the right direction
• Psychological interventions: prevention
• Algorithmic interventions
  • Boost trustworthy sources
  • Making moderate opinions more visible:
    Agreement between groups (1, 2)
• Design for entertainment vs. nuanced political discussion

• Anonymität depolarisiert: ideas before identities
• Colour buttons to show extremity of others, to change distortion

Conclusions

• Emotions do not generally increase irrationality
• Emotions around news have a social function
• Misinformation sharing is socially motivated, not the result of gullibility
• Misinformation is a symptom of real-world societal problems that lead to polarization
• Social media reflects and distorts this polarization, but is likely not the main cause
• Solving root causes of polarization in the real-world

For the curious…

• The New Yorker: Don’t believe what they’re telling you about misinformation
• Kurzgesagt Video
  
• Spektrum der Wissenschaften: Mythos Fehlinformation - Warum Social Media weniger gefährlich sind als wir denken

• ORF Dok 1
  
• DerStandard Podcast

Book recommendations

Mercier, 2020: Cognitive Science of misinformation & propaganda

Bail, 2021:

Polarization & social media

Thank you & time for questions!

Project website: https://hannahmetzler.eu/emomis

Appendix

Emotions & COVID-misinformation

• Actual true & false COVID-19 headlines from fact-checking websites & mainstream news sources in Austria in 2021
• Accuracy ratings (discernment)
• Emotions increase gulliblity to political news in the US (Martel et al. 2020 )

Emotions predict emotions in Twitter discussions

Why were people angry?

First thoughts after seeing COVID-19 news headlines

ABM Set-up

• from all politicians in our final dataset, random politicians are selected in each timestep, who post original tweets
• the tweets‘ characteristics correspond to actual tweets in our dataset
• the tweet spreads to the politician network of the posting politician
• the platform ranking algorithm then determines its ranking (we only model this when modelling interventions). In the base model, it is a reverse-chronological ranking only
• if the post makes it to the top n posts on the follower timeline, the following politician makes a retweeting decision. This is implemented through the logistic regression model
• if the post gets retweeted, it spreads to the retweeting politician‘s politician network and so on
• record the total number of politician retweets over the simulation run

Extended version - retweets are increased by a factor to approximate total retweets (i.e., from general public), by increasing the retweet count of 1 by the politician‘s follower number - introducing randomness through drawing from a binomial distribution - this follower number is first reduced by overlaps with followers of previous tweeting and retweeting politicians of the same post