In recent years, artificial intelligence (AI) has been recognized as a powerful tool for monitoring and countering extremism online and on social media. From automated content filtering to predictive analytics, AI is considered promising in detecting threats and extremist content digitally (e.g. terrorism, hate speech, online radicalization, misinformation, and coordinated extremist activity) on a large scale and responding faster than human teams[1]. However, when it comes to Arabic-language content, the picture is entirely different. According to the Institute for Strategic Dialogue (ISD), extremists such as Daesh were observed exploiting gaps in Arabic AI monitoring tools between 2020 and 2021 through camouflaging, a set of tactics designed to conceal the true nature of content and accounts. These tactics include leet speak,i.e. replacing letters with numbers, symbols, or alternate characters to evade detection, alongside broken text, multilingual mixing, unusual spacing and punctuation, and embedding extremist material within seemingly neutral posts. Such techniques enable actors to bypass automated detection, mislead human moderators, and disseminate propaganda widely. Notably, ISD monitored the Global Happenings network during this period, documenting the creation of new Twitter accounts and Facebook pages in August 2021 to circulate pro-Islamic State content under the guise of legitimate media[2]. These radical groups have also long used encryption tools such as Mujahedeen Secrets, Tashfeer al-Jawwal, and Asrar al-Dardashah to evade detection online by hiding communications and coordinating off-platform[3]. These evasion techniques have evolved in reaction to AI encryption tools such as neural networks that develop their own ciphers or embed data in ordinary text and media, and more sophisticated and harder to detect, raising urgent security, social, and ethical concerns.
Arabic is not a homogeneous language, rather, it encompasses a rich network of local dialects, colloquial expressions, code-switching between Arabic and English, and the creative use of symbols and emojis. Standard AI models, which are mostly trained on English datasets or Classical/Modern Standard Arabic[4], often fail to accurately detect extremist content. This leads to high rates of false positives (classifying harmless content as a threat) and false negatives (failing to detect real threats)[5]. These gaps create a dilemma where extremists operate with near-complete freedom, while ordinary users face the risk of excessive censorship.
In 2023, for example, Meta’s automated systems removed posts using the Arabic word “shaheed” when referring to individuals on its Dangerous Organizations list. The term “shaheed” literally means “martyr” in Arabic and is commonly used in Islamic contexts to honor someone who has died in the service of their faith or country. While the word is associated with martyrdom in extremist contexts, it also carries non-violent, religious, or memorial meanings. For instance, it can appear in posts commemorating deceased family members, religious figures, or community leaders, or in discussions of historical events where individuals died defending their beliefs. The company did not count for this nuance and removed content regardless of context, resulting in the censorship of legitimate religious or memorial content that was not promoting violence. The case highlights the risks of over-reliance on automated moderation without cultural sensitivity or human oversight[6]. The gap between what moderation systems assume and what users mean in context can be large when context, dialect, cultural usage is not accounted for.
Operational shortcomings
Many technology companies like Meta invest in monitoring content in English or other languages, which results in unbalanced oversight on social media platforms. An internal review of Facebook’s Arabic content moderation system revealed significant challenges in managing the wide range of dialects spoken across the Middle East and North Africa. The report noted that “Arabic is not a single language, but rather a family of languages many of which are mutually unintelligible.” It also emphasized that the social and political context varies by country, making it even harder to accurately detect and remove hate speech and misinformation[7]. At the same time, government entities in the Arab world often lack the resources, expertise, or regional collaborative frameworks necessary to support AI-based monitoring systems.
Security and Policy perspective
These gaps have serious consequences. From a security perspective, extremists exploit the weaknesses of AI capabilities to organize their activities, recruit individuals, and openly disseminate violent ideologies on social media with minimal risk of detection. This not only amplifies their reach but also erodes public safety in both digital and physical spaces. From a policy perspective, governments and organizations may develop a false sense of security, believing that AI systems are effectively combating extremist threats, while dangerous activities go unnoticed,[8] delaying the development of more robust countermeasures and allowing harmful activities to proliferate unchecked.
Arabic dataset for evaluating large language models
In February 2024, a group of researchers from the NLP department at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) developed the Arabic Massive Multitask Language Understanding Assessment (ArabicMMLU). The first standardized Arabic dataset designed to evaluate both the performance and understanding of large language models in Arabic. This paper was designed to bridge the gap caused by the shortage of Arabic data and to ensure more accurate measurement of linguistic understanding and reasoning. ArabicMMLU evaluates the cognitive dimension, namely what the model learns and retains during training. Previously, the data available to assess the Arabic knowledge of models was limited to sources translated from English, which reduced effectiveness, as translation can introduce errors and overlook the cultural and contextual nuances specific to the Arab region[9]. The ArabicMMLU initiative can be viewed as a model for developing authentic Arabic linguistic resources to build advanced, specialized AI tools in context of countering violent extremism that go beyond translation, deeply understand the Arab context, and contribute to bridging the gap in monitoring and countering digital extremist discourse and activities. Therefore, AI models must be trained on diverse dialects, code-switched language, and culturally relevant material such as religious discourse, political rhetoric, and historical narratives to help reduce false positives by distinguishing legitimate expression from incitement. Integrating human annotation by native Arabic speakers with cultural and geopolitical awareness is also critical to improving model accuracy. While no complete solution exists yet, these strategies offer a clear and practical path forward.
Recommendations
Addressing these challenges requires a multidimensional and ethically conscious approach:
Invest in dedicated Arabic models: Specialized AI models tailored to Arabic should be developed through the creation of large, annotated datasets that capture and adapt to changes in dialects, colloquial language, and extremist terminology, ensuring more accurate monitoring tools. While AI can handle the scale and speed of detection, human analysts remain essential for contextual review, interpreting subtle signals and cultural nuances, and continuously updating and feeding the models. Analysts, researchers, engineers, policymakers, and NGOs must rely more on subject-matter experts who can identify new trends in the space. These proactive findings can then inform the development of more up-to-date and responsive leet speak dictionaries, which can be used to better train machine learning tools that search for this content at scale. While this approach would require greater human resources in the short term, it will significantly reduce the amount of extremist content that makes it through.
Enhance regional and international cooperation: Governments, technology companies, and research institutions in the Arab world should work together to share data, develop best practices, and establish unified protocols for monitoring and responding to extremist content digitally. A practical model for this can be seen in international initiatives like Tech Against Terrorism and its Terrorist Content Analytics Platform (TCAP), which provides smaller tech platforms with tools and real-time alerts to detect and remove terrorist content including in Arabic. The Middle East can utilize these models by establishing a regional Arabic-language extremism monitoring tool, creating cross-border data-sharing agreements, and forming dedicated partnerships with major tech companies such as Meta, Telegram, and YouTube. This would enable the development of culturally and linguistically tailored AI moderation tools and unified takedown protocols.
Training programs for regional researchers and content moderators: By aligning with global efforts and developing region-specific strategies, Arab countries can take a proactive, coordinated approach to build local capacity to counter radical narratives more effectively.
Conclusion
Despite AI’s growing role in countering extremism online, significant gaps remain, especially in Arabic-language monitoring. Extremist actors continue to exploit linguistic, technical, and operational weaknesses to evade detection online, while automated systems often over-censor legitimate expression due to lack of cultural and contextual understanding. Bridging this gap requires investment in Arabic-specific AI models, deeper human oversight, and stronger regional cooperation. By combining advanced technology with localized expertise and ethical safeguards, the Arab world can build more accurate, responsive, and fair systems to counter digital extremism effectively.
Disclaimer:
The views and opinions expressed in the INSIGHTS publication series are those of the individual contributors and do not necessarily reflect the official policy or position of Rabdan Security & Defense Institute, its affiliated organizations, or any government entity. The content published is intended for informational purposes and reflects the personal perspectives of the authors on various security and defence-related topics.
References
AI and the Arabic language: Preserving cultural heritage and enabling future discovery. (2024, Decemeber 18). Retrieved from MBZUAI: https://mbzuai.ac.ae/news/ai-and-the-arabic-language-preserving-cultural-heritage-and-enabling-future-discovery/
Dark Motives Online: An Analysis of Overlapping Technologies Used by Cybercriminals and Terrorist Organizations. (2016, May 3). Retrieved from Trendmicro: https://www.trendmicro.com/vinfo/us/security/news/cybercrime-and-digital-threats/overlapping-technologies-cybercriminals-and-terrorist-organizations
fajri.koto, h. s. (2024, February 20). ArabicMMLU: Assessing Massive Multitask Language Understanding in Arabic. Retrieved from Arxiv: https://arxiv.org/html/2402.12840v1#Sx1.SS0.SSS0.Px2
Iyengar, R. (2021, October 26). Facebook has language blind spots around the world that allow hate speech to flourish. Retrieved from CNN: https://edition.cnn.com/2021/10/26/tech/facebook-papers-language-hate-speech-international
Mohamed, N. (2025, April 30). Artificial intelligence and machine learning in cybersecurity: a deep dive into state-of-the-art techniques and future paradigms. Spriger, 46. Retrieved from https://link.springer.com/article/10.1007/s10115-025-02429-y
Moustafa Ayad, N. K.-T. (2022, October 31). The Terror Times: The Depth and Breadth of the Islamic State Alternative News Outlet Ecosystem Online. Institute For Strategic Dialogue, 16-18. Retrieved from Global Network on Extremisim and Technology.
PAO on the treatment of the term “shaheed” (شهيد). (2024, July 2). Retrieved from Meta: https://transparency.meta.com/pt-br/oversight/oversight-board-cases/shaheed-pao
الذكاء الاصطناعي في مواجهة الإرهاب.. فرص وتحديات. (2021, October 27). Retrieved from مركز المعلومات ودعم اتخاذ القرار: https://www.idsc.gov.eg/Article/details/6259
باستخدام تقنيات الرصد والتحليل والتنبؤ - توظيف الذكاء الاصطناعي للحدِّ من التطرف والإرهاب. (2024, March 28). Retrieved from Islamic Military Counter Terrorism Coalation: https://www.imctc.org/ar/eLibrary/Articles/Pages/article28032024.aspx
[1] (الذكاء الاصطناعي في مواجهة الإرهاب.. فرص وتحديات, 2021)
[2] (Moustafa Ayad, 2022)
[3] (Dark Motives Online: An Analysis of Overlapping Technologies Used by Cybercriminals and Terrorist Organizations, 2016)
[4] (AI and the Arabic language: Preserving cultural heritage and enabling future discovery, 2024)
[5] (Mohamed, 2025)
[6] (PAO on the treatment of the term “shaheed” (شهيد), 2024)
[7] (Iyengar, 2021)
[8] (باستخدام تقنيات الرصد والتحليل والتنبؤ - توظيف الذكاء الاصطناعي للحدِّ من التطرف والإرهاب, 2024)
[9] (fajri.koto, 2024)
