Artificial Intelligence, Disengagement, and Terrorism Prevention: Opportunities and Challenges - Global Network on Extremism and Technology

Artificial Intelligence, Disengagement, and Terrorism Prevention: Opportunities and Challenges

This Insight explores the use of Artificial Intelligence (AI) as a tool in the fight against terrorism, particularly jihadist terrorism. While AI is often seen as a weapon for terrorists – whether for propaganda or the manufacturing of explosive devices – it is rarely viewed as a tool for counterterrorism. 

However, some companies have developed AI-based technological solutions to combat radicalisation. This Insight looks at two companies – Mythos Lab and CAPRI – with whom interviews were conducted to gather some of the data detailed below with executives from these companies, who kindly provided information on how their tools work and on figures from internal studies. 

Beyond technological considerations, effectively addressing terrorism requires clearer conceptual frameworks for understanding it. The concepts of radicalisation, prevention, and disengagement remain debated, with no clear consensus in the academic literature or among practitioners. While prevention aims to limit entry into radicalisation processes, disengagement refers to moving away from violence once individuals are already involved. In practice, however, these dimensions are closely linked.

The fight against terrorism can be understood as a continuum between prevention and disengagement: individuals may disengage from violence at one point, only to re-engage or re-radicalise later. This interdependence suggests that both approaches must be considered jointly rather than in isolation, an assumption that structures the analysis developed in this article.

AI in Reintegration Support for Former Prisoners

Companies have developed AI tools to support professionals working to disengage jihadist militants, particularly those who joined terrorist groups such as ISIS in the 2010s, as they are released from prison. 

One such company is Mythos Labs, founded by Priyank Mathur, a former policy advisor and intelligence analyst with the US Department of Homeland Security (DHS). 

Mythos Lab AI’s tool is called Aldous, which its creators describe as the world’s first AI-based system for terrorist rehabilitation and risk assessment. Based on a large language model (LLM), it is designed to generate context-sensitive and personalised interactions with users. In 2025, Aldous was deployed to help rehabilitate 52 former terrorists belonging to the Islamic State, Jemaah Islamiyah (JI), and Jamaah Ansharut Daulah (JAD) in Indonesia and the Philippines. 

In practical terms, former terrorists participating in local rehabilitation programs had the opportunity to interact with the Aldous chatbot via WhatsApp and Facebook Messenger. The chatbot is capable of engaging in personalised and contextualised conversations with them, aimed at addressing the cognitive, psychological, and emotional drivers that contributed to their radicalisation. The primary goal is therefore disengagement from violence. 

Analyses of overall risks and psychological profiles, automatically generated by the tool, are used by practitioners in the field who work with individuals convicted of terrorism.

According to internal data, after three months of conversations with Aldous, 61% of former terrorists showed lower risk levels and 0% showed higher risk levels. In addition, 92% of practitioners using Aldous reported that the information from Aldous’ risk assessments enabled them to provide more effective and targeted advice to former terrorists. Analysis of the conversations showed the ability of the Aldous chatbot to steer former terrorists away from violent fantasies, directing them toward constructive goals, transforming resentment into resilience, and shifting the dynamics of self-accusation toward forms of self-compassion. Also, it should be noted that they were informed that everything they say to the Aldous chatbot will be monitored and shared with the program administrators. Despite this monitoring requirement, 80% chose to speak with it.

According to Mythos Lab, former terrorists who interacted with Aldous regularly highlighted its accessibility and empathy. One former Islamic State fighter said, “I chat with him when I’m resting, as if he were my friend.” Another explained, ”What I like most about Aldous is that when you write to him, he responds immediately and his answers are very clear.”

Mythos Labs plans to expand the deployment of Aldous to other regions in partnership with local authorities.

AI as a Tool to Help Prevent Radicalisation

While AI can serve as a tool to support the deradicalisation process, it can also be used to intervene before radicalisation takes hold. In France, certain organisations are experimenting with the use of LLM-based conversational tools to facilitate access to information and direct people to the appropriate services. The Centre for Action and Prevention against the Radicalisation of Individuals (CAPRI) is an innovative example of this. 

This service was initially designed as a listening platform for families concerned about the possible radicalisation of a loved one, often a child. CAPRI’s missions are based on the intervention of educators and psychologists with the goal of reestablishing dialogue within the family environment. This support helps prevent a breakdown in family ties, which could fuel a process of radicalisation.

Since its creation in 2015, the platform has gradually built up a database that provides a better understanding of the concerns expressed by families. The Chatbot CAPRI relies on the use of anonymised data from requests processed by the platform, combined with research, institutional guides, and various educational resources. These elements enable the training of a language model (LLM) capable of providing contextualised responses, regularly adjusted according to changes in requests and issues encountered in the field. The tool thus constitutes an initial information portal that can refer users to specialised services when the situation calls for it. The aim is not to replace professional intervention, but to facilitate care by screening requests.

According to CAPRI, approximately 50 users interact with the chatbot each month. Of these, about 20% have a genuine need for help. Since the tool’s launch, requests have increasingly involved individuals adhering to an ideology considered to fall under far-right extremism, while the majority of individuals – about 60% – are linked to the political-religious jihadist ideology.

The CAPRI system has built in several safeguards. The chatbot does not make independent diagnoses and systematically refers users to the appropriate services for any situation requiring professional assessment. In addition, interactions are anonymised in order to preserve user confidentiality and maintain the relationship of trust that is central to this type of system.

In practice, the chatbot seems to play a role in guiding and filtering requests. Some users are simply looking for information or advice when faced with situations they perceive as ambiguous, delicate, or difficult to understand, while a small proportion correspond to situations requiring in-depth support. 

This type of experiment illustrates the role that AI can play in “hyper-prevention” strategies. Rather than replacing human actors, these technologies can help strengthen existing systems by facilitating access to information, guiding users, and supporting the work of professionals.

Limitations and Empirical Challenges: AI and the Problem of Terrorist Recidivism

While AI tools appear promising for preventing radicalisation and supporting disengagement, their actual effectiveness must be evaluated with caution. Several empirical and methodological limitations must be taken into account.

First, the evaluation of these systems faces a structural difficulty. Research on terrorism shows that recidivism is rare. Indeed, empirical research shows that individuals convicted of terrorism-related offences have significantly lower recidivism rates than those observed in ordinary crime. 

For example, analysis of a longitudinal database of 629 individuals convicted of terrorism in the United States between 2001 and 2020 indicates that approximately 3.1% of individuals committed offences – 20 in total – with only 11 occurring after release from prison. However, none of these offences were classified as terrorist acts. In any case, these results show that former prisoners convicted of terrorism reoffend much less frequently than common criminals, for whom recidivism rates can exceed 30 to 60% within three to five years of release. 

Another study of perpetrators of jihadist attacks in Europe between 2014 and 2022 reached similar conclusions regarding the rarity of terrorist recidivism. Researchers analysed a database of 105 attacks involving 133 perpetrators. They identified approximately 9% of repeat offenders who had already been convicted of terrorist offences before participating in a new violent act. However, the researchers emphasise that these results should not be interpreted as a general recidivism rate, but rather as the proportion of repeat offenders among the perpetrators of attacks identified in their sample.

This low prevalence makes the evaluation of disengagement programs particularly complex. When a phenomenon is rare, it becomes difficult to attribute any reduction in risk to the effectiveness of a specific measure, whether it be a socio-judicial monitoring program or a technological tool. In other words, the absence of recidivism in terrorism is not necessarily proof of effectiveness.

Furthermore, the results announced by certain companies or organisations developing AI-based solutions must be interpreted with caution. In the case of some experimental programs, evaluations are often based on small sample sizes, relatively short observation periods, and methodologies that are difficult for independent researchers to verify. Under these conditions, it remains difficult to establish a causal link between the use of the tool and changes in participants’ attitudes or behaviours.

The issue of recidivism is a major challenge for disengagement issues. Indeed, the majority of convictions following a conviction for terrorism relate to common law offences and crimes that are apolitical and without ideological motivations (p.61). This suggests that the issue of ideological disengagement is only one aspect among others in the processes of reintegration or social reaffiliation.

Conclusion and Recommendations

The examples presented in this Insight show that AI can be used to positively support individuals at different stages of radicalisation, helping prevent violent engagement or re-engagement. It can thus support disengagement from terrorist violence motivated by political and/or religious ideology. Conversational tools and data analysis systems can thus help to improve the identification of risk situations, the monitoring of beneficiaries, and the effectiveness of interventions by practitioners in the field. However, these technologies should be viewed as support tools rather than standalone solutions.

Several limitations of using AI tools must be taken into account. 

First, we know that the use of conversational agents can give rise to what the literature refers to as the “ELIZA effect.” This emotional and psychological transfer refers to the tendency of users to attribute to a computer program capacities for understanding or empathy specific to human interactions. 

In some cases, this anthropomorphisation can foster an attachment to the digital interface and encourage intensive use, or even forms of dependence. In an area as sensitive as preventing radicalisation or supporting disengagement, these dynamics call for particular vigilance regarding the role these technologies play in support mechanisms.

The use of these tools also raises evaluation issues. The low statistical prevalence of terrorist recidivism already makes it difficult to assess the effectiveness of disengagement programs. This difficulty is compounded by the possibility of concealment strategies by certain individuals. In certain ideological contexts, these strategies may be associated with concealment practices such as taqiyya, or more broadly with strategic adaptation behaviours aimed at meeting institutional expectations without necessarily reflecting a lasting change in beliefs. The evaluation of this strategy requires a human touch. 

Furthermore, the use of digital tools to monitor individuals convicted of terrorism raises data security issues. Systems based on sensitive information may be exposed to the risk of cyberattacks or data breaches, which require high levels of governance and information protection.

In this context, it seems essential to develop independent, external evaluations of AI-based systems in order to better measure their long-term effects. Longitudinal and comparative research is needed to determine the extent to which these technologies effectively contribute to preventing radicalisation or disengagement from violence.

Finally, research on disengagement trajectories shows that engagement in violent extremism is often accompanied by processes of social disaffiliation, marked by breaks with institutions, family, or friends. Rehabilitation programs specifically seek to promote processes of reaffiliation, based on rebuilding social ties, psychosocial support, and integration into ordinary social networks. These relational dynamics are at the heart of rehabilitation programs and cannot be replicated by automated technologies.

AI can thus enhance certain analytical, guidance, and monitoring capabilities. However, in the field of disengagement from violence, it cannot replace the human interactions that make it possible to rebuild social ties and achieve the sustainable reintegration of individuals.

–

Hugo Champion: After studying political science at Sciences Po Aix and having conducted academic research on women’s involvement in jihad, I worked for a French organisation commissioned by the Ministry of Justice. As an intercultural and religious mediator, my role was to help individuals convicted of terrorism renounce violence in a post-incarceration setting. In addition to these projects, I founded an online media outlet – Le Chronoscope – which explores contemporary political violence, particularly jihadism and the far right. I also contribute occasionally to “Security and Defence Magazine,” which covers topics related to international security, geopolitics, cybersecurity, and the defence industry.

–

Are you a tech company interested in strengthening your capacity to counter terrorist and violent extremist activity online? Apply for GIFCT membership to join over 30 other tech platforms working together to prevent terrorists and violent extremists from exploiting online platforms by leveraging technology, expertise, and cross-sector partnerships.