Online Library

Over the past decade, two major phenomena have developed in the digital realm. On the one hand, extremism has grown massively on the Internet, with sprawling online ecosystems hosting a wide range of radical subcultures and communities associated with both ‘stochastic terrorism’ and the ‘mainstreaming of extremism’. On the other hand, Artificial Intelligence (AI) has undergone exponential improvement: from ChatGPT to video deepfakes, from autonomous vehicles to face-recognition CCTV systems, an array of AI technologies has abruptly entered our everyday lives. This report examines ‘AI extremism’, the toxic encounter of these two evolutions – each worrying in its own right. Like past technological progress, AI will indeed be – in fact already is – used in various ways to bolster extremist agendas. Identifying the many opportunities for action that come with a range of AI models, and linking them with different types of extremist actors, we offer a clear overview of the numerous facets of AI extremism. Building on the nascent academic and government literature on the issue as well as on our own empirical and theoretical work, we provide new typologies and concepts to help us organize our understanding of AI extremism, systematically chart its instantiations, and highlight thinking points for stakeholders in countering violent extremism.

VOX-Pol Virtual Roundtable: Artificial Intelligence and Extremism, Wednesday 13 December.

(Please note, some slides are blurred due to sensitivities around ongoing research)

This roundtable is organized by the VOX-Pol Network of Excellence on online extremism in the context of the ExID, Extreme Identities, project funded by the EU Horizon2020 NORFACE programme, Democratic Governance in Turbulent Times. Over the past decade, two major phenomena have developed on the digital realm: the extensive growth of extremism on the internet and exponential improvements of Artificial Intelligence (AI) models. This VOX-Pol Roundtable examines the toxic encounter of these two evolutions, from five distinct perspectives offered by its chair and expert panellists, offering cutting-edge insights on issues such as AI-powered detection of radical content or deepfake use by extremist groups.

Chair

Dr Stephane J. BAELE is Professor of International Relations at the Catholic University of Louvain (Belgium) and Honorary Associate Professor of Security and Political Violence at the University of Exeter (UK). The editor of ISIS Propaganda (Oxford University Press), his multidisciplinary work on extremist and violent political actors’ communications has appeared in leading journals such as the Journal of Conflict Resolution or Studies in Conflict & Terrorism. He recently published a CREST Guide on “Artificial Intelligence and Extremism: The Threat of Language Models for Propaganda Purposes” (2022).

Panelists

Dr Lewys BRACE is Senior Lecturer in Computational Social Science at the University of Exeter (UK), where he is the Co-Director of the Q-Step Centre for quantitative social sciences. His work, published in journals like Terrorism & Political Violence or the Journal of Policing, Intelligence & Counter-Terrorism, harnesses data science, Open-Source Intelligence (OSINT), and artificial intelligence methods to analyse extremism, terrorism, cybercrime, and security.

Dr Thomas HEGGHAMMER is Senior Fellow of All Souls College, University of Oxford (UK). A political scientist and historian specializing in militant Islamism and jihadism (e.g., Jihadi Culture, Jihad in Saudi Arabia, or recently Abdullah Azzam’s biography The Caravan) he has recently deployed computational/AI methods for the study of extremist corpora in Arabic (e.g., “daiR: an R package for OCR with Google Document AI” in the Journal of Open Source Software, and “OCR with Tesseract, Amazon Textract, and Google Document AI: a benchmarking experiment” in the Journal Computational Social Sciences).

Dr Miron LAKOMY is Professor at the Institute of Political Sciences of the University of Silesia (Poland) and currently visiting fellow at the ITSTIME research centre in Milan (Italy). He has held visiting positions at, among others, the University of Oxford, the University of Cambridge, the University of Milan, and the European University Institute. His research primarily focuses on various aspects of political violence and terrorism on the Internet, as well as military conflicts. He is, among others, the author of the recent article “Artificial Intelligence as a Terrorism Enabler?” (published in Studies in Conflict & Terrorism).

Ms Marie SCHRÖTER is since August the Country Director of the Friedrich-Ebert-Stiftung (FES) in the Philippines (after working for FES in Tel Aviv). Before that, she focused on the role of AI in counterterrorism as a Mercator Fellow with the United Nations Office of Counter Terrorism in New York, the International Centre for the Study of Radicalization (ICSR) at the Kings College London, and AlgorithmWatch in Berlin. She is the co-founder (in 2020) of Codetekt – a civil society organization against online disinformation. She is, among others, the author of the report “Artificial Intelligence and Countering Violent Extremism: A Primer” (published by GNET in 2020).

Anhand der Einzelfallstudie des Radikalisierungsprozesses einer jihadistischen Straftäterin möchte dieser Beitrag die besondere Rolle des Internets bei Radikalisierungen von Extremisten aufzeigen. Die Straftäterin war in den Jahren 2007 bis 2010 einePropagandistin, die über das Internet mehrere terroristische Vereinigungen unterstützte, indem sie deren Videobotschaften online verbreitete. In der Einzelfallstudie wird aufgezeigt, wie sich die Jihadistin durch virtuelle Kontakte, Propagandaaktivitäten sowie ihre Zusammenarbeit mit zahlreichen Jihadisten im Ausland über das Internet radikalisierte.

The Role of the Internet on Radicalization of a German Jihadist Offender: A Case Study
Based on the analysis of a case study this article raises the question about the role of the internet in accelerating the radicalization of a woman who joined the jihadist milieu in southern Germany in the mid 2000s. From 2007 to 2010 she was a jihadist activist on the internet who supported several terrorist groups by publishing videos and jihadi messages in blogs, video channels and web forums. The case study shows how the woman was radicalized through virtual contacts, interactions in virtual communities, as well as through her support of several jihadists abroad.

Hate speech detection on Twitter is critical for applications like controversial event extraction, building AI chatterbots, content recommendation, and sentiment analysis. We define this task as being able to classify a tweet as racist, sexist or neither. The complexity of the natural language constructs makes this task very challenging. We perform extensive experiments with multiple deep learning architectures to learn semantic word embeddings to handle this complexity. Our experiments on a benchmark dataset of 16K annotated tweets show that such deep learning methods outperform state-of-the-art char/word n-gram methods by ~18 F1 points.

With the ever-increasing cases of hate spread on social media platforms, it is critical to design abuse detection mechanisms to pro-actively avoid and control such incidents. While there exist methods for hate speech detection, they stereotype words and hence suffer from inherently biased training. Bias removal has been traditionally studied for structured datasets, but we aim at bias mitigation from unstructured text data. In this paper, we make two important contributions. First, we systematically design methods to quantify the bias for any model and propose algorithms for identifying the set of words which the model stereotypes. Second, we propose novel methods leveraging knowledge-based generalizations for bias-free learning. Knowledge-based generalization provides an effective way to encode knowledge because the abstraction they provide not only generalizes content but also facilitates retraction of information from the hate speech detection classifier, thereby reducing the imbalance. We experiment with multiple knowledge generalization policies and analyze their effect on general performance and in mitigating bias. Our experiments with two real-world datasets, a Wikipedia Talk Pages dataset (WikiDetox) of size ~ 96k and a Twitter dataset of size ~ 24k, show that the use of knowledge-based generalizations results in better performance by forcing the classifier to learn from generalized content. Our methods utilize existing knowledge-bases and can easily be extended to other tasks.

VOX-Pol Blog post.