Analyzing Post-Riot Behavior: Tracking Far-Right Radicalization Through 51,000 Facebook Messages

Over 1,100 individuals have faced charges related to the summer 2024 riots, with a small fraction being prosecuted for crimes associated with their online conduct.

Sentences varied from 12 weeks to seven years, igniting a surge of online backlash. The individuals behind the posts were varied; one notable case is that of I defended, who emerged as a cause célèbre and was labeled a “political prisoner.” Their posts were minimized and mischaracterized; their prosecution was framed as an infringement on free speech, despite the majority of online-related charges involving allegations of inciting racial hatred.

The posts did not predominantly surface in mainstream social media platforms like X, Instagram, or Facebook, but rather in niche online spaces commonly linked to fringe ideologies like Telegram, Parler, GetTr, 4Chan, and 8Kun. While many of these posts were on personal profiles, some appeared in public group forums.

This raised questions: What online communities did these individuals engage with, and who were their advocates? What type of content was circulating in these environments? It seemed that within these circles, views were so normalized that individuals felt emboldened to share content that was considered criminal by British authorities and the judiciary.

As a starting point, we utilized publicly accessible resources (police records and news reports) to track Facebook accounts of those implicated in previous investigations. Out of approximately 20 individuals charged with online offenses related to the summer 2024 riots, we followed five to three public Facebook groups. We also discovered visually similar or replicated posts defending those referenced in these groups.

This led to the mapping of a broader network of other Facebook groups, connected through shared memberships and group moderators and administrators.

In this exploration, we uncovered vibrant ecosystems characterized by a profound distrust of government and its institutions, alongside online communities preoccupied with anti-immigrant sentiments, naturalism, conspiracy theories, and misinformation.

Additionally, we found individuals who expressed genuine concerns about the society they belong to, alongside those who are deeply disillusioned and believe their freedom of expression is at risk.

Identification of Groups

Why focus on these groups?

Three groups were selected for the primary analysis because they included one or more current or former members charged in connection with the summer 2024 riots, or individuals involved in the riots who made comments either in person or online.

We established links between these and 13 additional groups, with all but three being public. These groups play significant roles, as moderators can oversee memberships, approve requests, and issue bans, with the authority to delete posts and comments. Administrators have even broader permissions, including the ability to modify group settings, update descriptions, and appoint additional moderators or administrators.

Which posts were analyzed?

To understand the type of content shared within these groups, we aimed to capture all posts made by the three largest groups from their inception until mid-May 2025.

We collected links and text from a total of 123,000 posts. However, due to the classification process (outlined below), the analysis was ultimately focused on 51,000 text-based posts.

What was the group membership size?

We did not record the names of individual group members (aside from moderators, administrators, and prominent posters). Therefore, when discussing combined memberships across groups, it is likely that individuals who belong to multiple groups were counted more than once.

Classification

First, we verified that the posts contained far-right content using established academic methods and categorized them through specific keywords indicating radicalization. We supplemented this with an AI tool that became available to data teams due to recent changes in editorial policies surrounding its journalistic use, classifying content as anti-establishment, anti-immigrant, migrant demonization, naturalism, and far-right identity/denial.

For categorizing the 51,000 social media posts, we employed ChatGPT 4.1 via OpenAI’s API. The prompts underwent rigorous testing across a random sample of 12 iterations, ensuring that at least two reviewers concurred until a consensus of over 90% was reached between the model and three human reviewers.

We are confident in the model’s reliability in small batches, supporting our broader evaluation based on a statistically determined sample of posts which achieved 93% agreement between human reviewers and the AI model.

The final analysis involved a statistically validated number of posts reviewed by the same annotator.

Testing concluded that the model performed exceptionally well, matching or even exceeding human reviewer consistency across most categories.

• Accuracy (Percentage of correctly classified instances): 94.7%.

• Precision (Percentage of correctly assigned true label counts by GPT): 79.5%.

• Recall (Percentage of instances classified as true by humans and also classified as true by GPT): 86.1%.

• F1 Score (A single percentage combining accuracy and recall, with higher values indicating better classification): 82.6%.

The model’s performance was evaluated by an internal statistical analyst, concluding its results were strong, benchmarked against similar academic studies.

Despite the model’s impressive performance, misclassifications in the analysis are inevitable.

We believe the classification process employing OpenAI’s API is thorough, transparent, defensible, and bolsters rigorous journalism.