Online Hate Speech in 2023

As conflicts escalate throughout the world, from Ukraine to Gaza, it is of increasing importance to observe how the online public opinion evolves.

Earlier this year, we carried out an analysis from data gathered within the framework of the European Observatory of Online Hate (EOOH) spanning from January 2023 until the middle of September 2023.

This article presents the results of the first iteration of the EOOH project that ran between 2021 and 2023.

Global Toxicity

Figure 1: Average Toxicity per Week

Toxicity reflects how undesirable an online message is, ranging from 0 to 1. A toxicity score of 0 means that a message is completely harmless and welcome, while a message with a score of 1 contains illegal hate speech such as discrimination based on protected categories (race, sexual orientation, religion, etc.), death threats and other types of harmful and dangerous content.

The first 3 quarters of 2023 have seen an increase of more than 30% in online toxicity (from an average of 0.16 to an average of 0.20). Growing political tensions as well as looser moderation practices on certain very large online platforms are contributing factors to that rise.

Topic Highlights

Figure 2: Average Toxicity per Topic

We kept track of 4 big topics known for attracting polarising debates, and often, toxic messages: Islam, Judaism, (the war in) Ukraine and Refugees. Via a combination of neutral keywords related to each theme, this gave us a collection of social media messages relevant to that topic, without introducing a bias from loaded search terms. 

Of all 4 topics, “Judaism” sported the highest average toxic score (0.28) followed by “Islam”. The fact that both religious topics appear first, as well as significantly higher than the other 2, shows the polarised and tense nature of discussions around these religions, along with substantial anti-semitism and islamophobia. 

“Ukraine” and “Refugees”, despite being lower than the religious themes, are not spared by harmful messages either amidst the ongoing war in Ukraine and recurrent public debates around migrants and refugees in the European Union.

Mainstream & Fringe Platforms

Figure 3: Proportion of messages in each toxicity bin per platform type

Different platforms and content policies mean different environments and discourses. That reality is reflected at the proportion of messages in each toxicity bin (from 0 to 1, in increments of 0.1). More than half of messages posted on mainstream platforms are completely harmless. The proportion of messages per bin gradually decreases as toxicity increases, indicating that most virulent messages are rarer. A slight exception to that trend is to be noted for the ‘0.9’ where the most extreme messages seem to be concentrated.

Contrary to the long tail exhibited by messages on mainstream platforms, the distribution for fringe platforms tells a different story for 2 reasons. Firstly, toxicity tapers off less rapidly and the more toxic bins contain a larger proportion of messages. Secondly, this distribution is much more polarised: 2 prominents spikes are to be observed on both ends (0.1 and 0.9 respectively), potentially reflecting extreme levels of discourse polarisation on those platforms.

Hate Speech Categories

Figure 4: Hate Speech Category Density per Platform Type

The figure above depicts the proportion of each hate speech category found in messages from mainstream online platforms (Facebook, X, Reddit, Instagram,TikTok) compared to fringe ones (4chan, Gab, Minds). For every category, fringe platforms showed a significantly higher proportion than mainstream ones, reflecting the overall more toxic environment of those platforms. Most notably, the 3 most prominent categories are “Politics”, “Violence” and “Racism”, reflecting an online environment crystallising mainly around those issues and their intersection. 

In conclusion, during the first 3 quarters of 2023, toxic messages (from merely undesirable to illegal hate speech) have been on the rise, most notably concerning politics, violence, racism and religion, and their intersection.

Figure 5: Toxicity over Time per Language

Over the course of 2023, 3 out of 5 of our most represented languages posted an increase in toxicity. Of those 3, Russian is the one with the steepest average increase (approximately 60%). Ongoing polarised discussions around the war in Ukraine are the major driver behind it. A spike is also observed for the month of April for English. This spike is correlated with increased activity on the (mostly) unregulated imageboard 4chan where a prominent and recurring theme for that timeframe is the “New World Order” (NWO). The “NWO” is a conspiracy theory that states that elites are trying to impose a global totalitarian regime. Messages about that theory also often feature antisemitic elements.