When and how does hateful online speech lead to real-world violence? Many violent events that dominate the news are preceded by online discussions. The uprisings of the 'Arab Spring' were called 'Twitter revolutions', the storming of the US Capitol was inspired by 'QAnon' online posts, and propaganda on Facebook contributed to the 2017 Rohingya genocide in Myanmar.
However, this does not mean every hateful thing said on the Internet leads to actual violence. In fact, hate speech is everywhere, while violence is rare. To understand how online speech leads to offline harm, we need a closer look.
ODAS provides insights by studying Hindu-Muslim tensions in India both online and in the real world. The project is composed of four work packages (WPs).
WP 1 researches how social media can contribute to real-world violence: Does social media simply help coordinate riots and demonstrations, or does it change people’s perceptions in ways that make violence more likely? We have designed a survey for Indian social media users to get to these effects. We present respondents with randomized social media posts, so we can learn about their effects in an experimental framework. Both the pilot and the full survey with 4,000 participants have been conducted, and the results have been publishes in Comparative Political Studies. We find some evidence for social media being used to coordinate violence, and more surprisingly, we find that negative sentiments against out-groups can emerge from seemingly harmless posts. This is important, because censorship of hate speech or disinformation will not be able to address these effects. One more research article is expected to draw on the survey, by researching the effects of factual corrections in debunking online rumors.
WP 2 has developed software to automatically collect online posts from social media in India. We have collected a total of 33 million posts, reposts, and comments, following legal and ethical guidelines from the now defunct network kooapp.com. Koo (an Indian Twitter equivalent) has been scraped before, and we followed the established procedure of using an undocumented programming interface for downloading data over longer time period.
Additionally, we have obtained activist records of attacks on religious minorities in India and data on internet outages. We find that certain hashtags conveying Hindu-nationalist ideas precede violent attacks. However, we also find that other tags praising religious unity are followed by decreases in violent attacks. Both of these effects disappear during times of internet outage, suggesting that online access needs to be in place for the statistical connection to hold. We have published our results in PNAS Nexus. Two more studies are under development, using similar data. One explores how elites can use political tensions to generate support and the other one researches the effects of government endorsed speech on social media discourse.
For extracting variables form social media speech, we have developed a programming library, drawing in Large Language Models. This allowed us to extract machine-readable variables from more than three hundred thousand Koo posts in less than 24 hours. The programming package and an associated data paper will also be published very soon.
Scholars from Norway, Sweden, India, Germany, and the US have formed the project team.
The main outcome of this project is improved knowledge about how online dangerous speech leads to real-world violence. Although our focus is on communal conflict in India, the research has implications for how group violence in other contexts can be driven by online speech. Moreover, our research sheds light on the mechanisms underlying online speech resulting in violence, which messages are dangerous, and the role of context conditions. To this end, we have found that seemingly benign references to in-group symbols and online rumors affect hostile emotions and even predict violence again religious out-groups. Moreover, we have identified ways to counter these effects, in terms of references to inter-religious cooperation and fact-checking of hostile rumors. Insights from ODAS could lead to an improved ability to limit communal conflict without censorship on social media. We have disseminated results through open access channel, increasing the potential impact these findings have.
Does hateful online speech lead to real-world violence between religious communities? This question remains unresolved and presents a puzzle: online hate speech is everywhere, while physical violence is rare.
Previous research has focused on the role of social media in individual “hate crimes” or in uprisings against authoritarian rule. However, no large-scale study has yet researched the role of online speech in recent deadly confrontations between Hindus and Muslims in India.
ODAS will leverage insights from Dangerous Speech Theory (DST) to shed much-needed light on the consequences of online speech for offline harm. Following DST, the project will research the mechanisms connecting speakers and audiences in large-scale surveys. Additionally, a novel “Speech Event Dataset” will be coded, using both automated pre-selection and manual coding. Finally, rich context information on settlement locations, demographics, and conflict history will be assembled from existing sources. Combining these data, ODAS will be able to research the effects of online speech on offline harm accounting for context conditions. Beyond causal connections, ODAS will produce actionable predictions of regions at risk of future communal clashes.
Theoretically, ODAS constitutes the first large-scale test of DST. Empirically, ODAS will deliver the first large-scale dataset of dangerous speech events and provide a rigorous test of their consequences for real-world violence. Methodologically, ODAS will break new ground: an innovative setup for coding speech events will be developed featuring automated processing and manual verification. Local conflict history and group demographics will be coded for all of India, benefitting a wider research program on communal conflicts. Key insights will be widely disseminated toward academic audiences, social media providers, stakeholders in India, and online audiences.
