French
Arabic
English
Portuguese
Spanish
Swahili
Sacré-Coeur 3 – BP 15177 CP 10700 Dakar Fann – SENEGAL.
+221 33 827 34 91 / +221 77 637 73 15
contact@timbuktu-institute.org
Source : Sud Quotidien
Jihadists, separatists, military authorities and Russian Africa Corps: every belligerent in the Malian conflict has seized on generative artificial intelligence to boost its propaganda. A report by the Timbuktu Institute, titled “Algorithms at War: How Generative AI is Reshaping the Malian Conflict” and published on 9 September 2026, offers an unprecedented overview of these differentiated uses and warns of a major risk: the line between the real and the fabricated is growing blurrier by the day, even as terrorist groups’ recruitment capabilities keep expanding.
For more than a decade, Mali has been the theatre of a multidimensional conflict. Alliances form and dissolve according to shifting strategic interests. The tactical alliance forged in 2025 between the jihadist group JNIM, affiliated with Al-Qaeda, and the Azawad Liberation Front (FLA), a predominantly Tuareg separatist movement, is the most recent illustration of this pattern. Their coordinated assaults in April 2026 dealt a severe blow to the Malian army and to its Russian Africa Corps allies. But another transformation is under way – less visible, yet just as decisive. The information space has become a battlefield in its own right. And on this new front, generative artificial intelligence (GenAI) is starting to rewrite the rules of the game. The Dakar-based research centre Timbuktu Institute has devoted a report to this phenomenon, titled “Algorithms at War.” Published on 9 September 2026, it examines how each actor in the conflict is appropriating this technology according to its capabilities, resources and objectives.
Jnim: a still-timid experiment with ai
Among the four parties examined, JNIM shows the most limited use of generative AI. So far, the group has confined itself to producing visual material to accompany its propaganda. In June 2026, its media arm Az-Zallaqa circulated AI-generated posters announcing the release of videos commemorating the “victorious” attacks of April 2026. These posters, which depicted fighters in combat, bore the telltale marks of AI generation: buildings and vegetation blending into one another, background details reduced to blotches, and an unnaturally smoothed-out image. According to the researchers, JNIM “is still at a limited but potentially evolving experimental stage”. The example of the Islamic State, which already uses AI for multilingual translation, video production and personalised recruitment chatbots, points the way forward. JNIM could, in time, harness AI to translate its messages into some fifteen regional languages (Bambara, Fulfulde, Zarma, Hausa, Tamasheq, etc.) or to dub its propaganda videos into local languages, thereby reaching populations that have so far been excluded from jihadist discourse.
The FLA and its synthetic characters
In contrast to this centralised approach, the FLA and its supporters are pursuing a more decentralised and creative strategy. TikTok accounts run by sympathisers or members of the movement feature recurring synthetic characters. The most famous of them is “Azzghim,” a Tuareg fighter whose fiery statements against the Africa Corps are widely shared. The Timbuktu Institute scrutinised several of his videos. Pixel anomalies, unnatural body movements, strange colouring and audio inconsistencies all betray his artificial nature. Azzghim does not exist. He is a pure algorithmic creation. These AI characters offer the FLA several advantages. They help humanise the movement, tailor messages to specific audiences (a teenage character can more easily reach young people), and facilitate the spread of a narrative favourable to Azawad independence. Above all, they provide unmatched operational security: it is impossible to capture or interrogate a fighter who has no physical existence. The report identifies a network of TikTok accounts that repost and amplify one another, creating the illusion of massive grassroots support. One account, @tawri.n.azawad, with 207,000 followers, is devoted exclusively to reposting and remixing Azzghim’s content. Another, @asnan831, features a teenage boy whose still images likewise show signs of AI generation.
Russia’s africa corps: disinformation at industrial scale
At the other end of the spectrum, the use of generative AI reaches its highest degree of sophistication on the Russian side, through the Africa Corps. Backed by the resources and expertise of the Russian state, the paramilitary group has folded AI into an already well-oiled disinformation apparatus. Several investigations have documented this strategy. In 2024, OpenAI and Meta identified a Russian operation using OpenAI’s models to generate comments, articles and images in English and French, aimed at accounts targeting sub-Saharan Africa. A joint technical report by VIGINUM (France’s agency for vigilance against digital interference), the United Kingdom’s Foreign Office and the European External Action Service exposed an information-manipulation network dubbed “AI-Freak.” This apparatus, run by African Initiative – a media outlet classified as a Russian propaganda platform – relies on a network of fake news sites and AI-generated content. African Initiative is the media arm of the Africa Corps. These Russian operations are not aimed solely at spreading disinformation. As an OpenAI report notes, AI-generated images serve above all to “catch the eye, so as to make articles or social media posts more engaging”. The goal is to saturate the information space, blur the trail and make attribution difficult, thanks to a built-in capacity for plausible deniability.
The malian regime: an attentive pupil
The Malian military regime, increasingly isolated on the international stage and dependent on Russian support, is gradually settling into this ecosystem. According to mapping by the Africa Center for Strategic Studies, the Alliance of Sahel States (AES) – which brings together the juntas of Mali, Burkina Faso and Niger – is the second-largest sponsor of disinformation in Africa, after Russia. Malian authorities have adopted Russian methods as their own. One TikTok account, @6tm_officiel, has amassed more than 436,000 followers by publishing AI-generated content extolling the successes of the Malian Armed Forces (FAMa) and the Africa Corps. One of its videos, showing the recapture of Kidal after the April attacks, reached 4.7 million views. Run by an individual based in Bamako, the account is representative of an “intensive production phenomenon” of content favourable to the regime. By controlling the national media space and leaning on these digital relays, the junta is seeking to shore up its legitimacy and divert attention from the deteriorating security situation.
A full-scale laboratory
Mali, the report concludes, is becoming “a veritable laboratory of AI-assisted information warfare”. Generative AI acts as a force multiplier for every actor involved. It lowers technical barriers, speeds up content-production cycles and makes attribution of messages more complex. It does not create new fault lines, but it amplifies and accelerates those that already exist – whether communal, ideological or geopolitical. The threat is not confined to propaganda. The report raises the possibility that JNIM could use chatbots to tailor its recruitment messages to the psychological profile of each target. It also mentions the use of AI-coupled drones to sharpen military targeting. The Africa Corps, for its part, could integrate autonomous weapons into its operations. The challenge facing Sahelian states and their partners is immense. Traditional strategies for preventing violent extremism, which still rely on collective awareness campaigns, risk being overtaken by AI’s capacity to produce hyper-personalised messages. The response will need to combine stronger digital literacy, enhanced detection capabilities, and prevention strategies adapted to this new technological reality. In a saturated information space, the ability to tell the real from the fabricated is itself becoming a matter of security. And for Malian populations, caught in the crossfire of manipulated narratives, truth has never been more of a weapon of war.
Alioune BA
AI Ethics Specialist
Timbuktu Institute – Septembre 2026
Télécharger l'intégralité du rapport
L’intelligence artificielle générative pourrait bien ouvrir un nouveau front dans la bataille informationnelle au Mali. Encore limité, l’usage de cette technologie par le JNIM révèle déjà les possibilités qu’elle offre à sa propagande : produire plus rapidement des contenus, les décliner dans les langues locales, adapter les récits aux différents publics et à terme, possiblement personnaliser le recrutement. Si le groupe n’en est encore qu’aux prémices de cette appropriation, la rapidité avec laquelle les barrières techniques s’abaissent pose une question plus large : les dispositifs de prévention de l’extrémisme violent sont-ils prêts à répondre à une propagande capable, elle aussi, de changer d’échelle ?
Le conflit malien, marqué depuis plus d’une décennie par l’enchevêtrement des dynamiques jihadistes, séparatistes et étatiques, pourrait être à l’aube d’une nouvelle mutation. La démocratisation de l’intelligence artificielle générative ouvre en effet aux différents protagonistes du conflit de nouvelles possibilités dans le domaine informationnel. Un paramètre mis en lumière par le rapport du Timbuktu Institute sur les fonctions de l’IA générative chez les différents acteurs du conflit, notamment l’appropriation progressive de cette technologie par le Jama'at Nusrat al-Islam wal-Muslimin (JNIM) à des fins de propagande. Un usage encore limité, certes, mais dont les implications potentielles méritent d’être suivies de près.
Le premier signe de cette évolution est apparu en juin 2026, lorsque la branche médiatique du JNIM, Az-Zallaqa, a diffusé une affiche générée par IA célébrant les attaques meurtrières menées à la fin du mois d’avril 2026, qui avaient coûté la vie au ministre malien de la Défense. Le visuel représentait des combattants du groupe dans différentes situations de combat et présentait plusieurs marqueurs caractéristiques de la génération par IA, notamment des bâtiments et une végétation qui se confondent ainsi qu’un lissage excessif des textures. Si l’utilisation de cette technologie demeure pour l’heure cantonnée à des supports promotionnels, elle n’est pas sans rappeler une trajectoire déjà empruntée par l’État islamique. Dès 2023, celui-ci avait développé un guide consacré à l’usage sécurisé de l’IA générative, avant d’intégrer progressivement des chatbots et des vidéos automatisées à ses campagnes de recrutement. L’expérience de l’EI laisse ainsi entrevoir une évolution possible : lorsqu’une organisation extrémiste acquiert la capacité technique et la volonté organisationnelle d’intégrer l’IA générative à un pan de son appareil de propagande, son usage tend ensuite à s’étendre à d’autres fonctions.
Produire davantage, plus vite
C’est peut-être dans cette perspective qu’il faut envisager la première implication directe de l’IA générative pour le JNIM : l’accélération de sa production de propagande. Le groupe repose actuellement sur des équipes internes et des procédés largement manuels, qui peuvent nécessiter des semaines, voire des mois de travail pour produire une vidéo de qualité professionnelle. Avec les outils d’IA générative désormais disponibles, cette temporalité pourrait être considérablement raccourcie. À partir d’un simple descriptif textuel, une demi-douzaine de visuels professionnels pourraient être générés en quelques minutes, avant d’être assemblés en vidéo avec animation, musique et texte. Une telle évolution aurait une conséquence assez immédiate : la démultiplication du volume de contenus produits et, potentiellement, une présence beaucoup plus virale du groupe dans l’espace informationnel régional. Autrement dit, l’enjeu ne résiderait plus seulement dans la capacité du JNIM à produire des contenus sophistiqués, mais dans celle d’en produire beaucoup plus rapidement.
Changer de langue, changer de public
Parallèlement, une seconde évolution pourrait concerner la localisation linguistique et culturelle du recrutement, dans un contexte d’endogénéisation croissante du jihadisme. Le JNIM diffuse aujourd’hui principalement ses messages en arabe classique, ce qui limite nécessairement leur réceptivité auprès de populations pour lesquelles l’arabe n’est pas la langue de communication quotidienne. La traduction neuronale pourrait cependant modifier cette équation. Elle permettrait de générer presque instantanément différentes versions d’un même message dans une quinzaine de langues régionales, parmi lesquelles le hassania, le fulfuldé, le bambara, le moore, le tamasheq, le songhaï ou encore le wolof.
Mais le changement serait plus marqué qu’une simple traduction linguistique. Cette capacité permettrait au groupe d’adapter ses narratifs aux différents publics qu’il chercherait à atteindre. Une version destinée aux pasteurs peuls pourrait ainsi mettre l’accent sur la protection des intérêts pastoraux et la défense contre la marginalisation gouvernementale. Un autre message, destiné aux jeunes urbains marginalisés, pourrait davantage insister sur la dignité, l’insertion économique et le statut social. À destination de jeunesses religieuses idéalistes, le discours pourrait, quant à lui, privilégier la spiritualité et la résistance à l’impérialisme occidental. Chaque narration pourrait ainsi être produite rapidement et simultanément, avec pour effet potentiel de maximiser l’efficacité du recrutement auprès de chaque segment de population. Le doublage par IA ouvrirait également un autre champ. En permettant de produire des vidéos de recrutement doublées en bambara, avec une voix synthétisée native, cette technologie pourrait donner accès à des populations jusqu’ici largement exclues du marché de la propagande écrite. Diffusées via Telegram, WhatsApp ou la radio informelle, de telles productions pourraient atteindre plus efficacement des jeunes analphabètes ou semi-alphabètes au Mali, au Burkina Faso et au Niger.
De la propagande de masse au recrutement sur mesure
L’évolution la plus préoccupante serait toutefois celle d’un recrutement hyper-personnalisé. Là encore, l’État islamique a déjà ouvert une voie. Le groupe avait développé des chatbots utilisant l’IA pour personnaliser ses messages de propagande en fonction des vulnérabilités psychologiques individuelles des personnes ciblées. Les possibilités désormais offertes par l’IA générative pourraient permettre au JNIM d’emprunter une trajectoire similaire. Un chatbot déployé sur Telegram pourrait ainsi être capable d’évaluer progressivement les aspirations, les frustrations et les convictions d’un jeune, avant d’adapter dynamiquement son discours afin de maximiser son impact psychologique et de favoriser son orientation vers des groupes extrémistes violents. Le changement d’échelle serait considérable. Là où la propagande traditionnelle s’adresse à des catégories relativement larges de population, l’IA permettrait potentiellement d’aller vers un discours individualisé, construit en fonction des caractéristiques propres à chaque interlocuteur.
C’est précisément ce qui pose un nouveau défi aux dispositifs de prévention de l’extrémisme violent. Les stratégies actuelles (sensibilisation communautaire, contre-discours religieux ou encore programmes d’insertion économique) sont pour l’essentiel conçues à une échelle collective et générique. Le recrutement assisté par IA pourrait, à l’inverse, s’opérer au niveau individuel et hyper-personnalisé. Dès lors, plusieurs questions se posent : comment contre-programmer un discours doublé en mooré, malinké, wolof ou peul sans disposer de capacités comparables de production multilingue par IA ? Comment identifier des chatbots de propagande sans disposer d’une expertise en détection algorithmique ? Et comment atteindre des populations analphabètes avec des messages de prévention essentiellement écrits ? Autant de questions auxquelles les acteurs de terrain devront répondre dans les prochaines années, sous peine de voir se creuser l’écart entre la sophistication croissante des dispositifs de recrutement et la stagnation relative des capacités de résilience régionales.
Une menace encore émergente, mais déjà à surveiller
Il convient toutefois de ne pas surestimer, à ce stade, l’ampleur du phénomène. Le rapport du Timbuktu Institute relève qu’il existe pour l’heure peu de preuves publiques d’une adoption officielle de l’IA générative par le JNIM au-delà de ses supports promotionnels. Cette prudence est d’autant plus nécessaire qu’un rapport du SITE Intelligence Group, publié en mai 2024, soulignait déjà la difficulté à mesurer précisément l’ampleur de l’aubaine que représente l’IA pour les groupes terroristes, compte tenu de leur dépendance persistante aux procédés classiques de production de contenu. Il n’en demeure pas moins que l’abaissement continu des barrières techniques d’accès à ces outils rend de plus en plus probable une expérimentation accrue du JNIM. Génération vidéo, traduction automatique, doublage multilingue ou recrutement ciblé pourraient progressivement entrer dans son répertoire d’outils informationnels. Pour l’heure, le phénomène reste donc émergent. Mais son intérêt tient précisément à cette phase précoce : les capacités technologiques progressent plus rapidement que les dispositifs destinés à en anticiper les usages. Si le JNIM venait à franchir un nouveau seuil dans l’intégration de l’IA générative à son appareil de propagande, la question ne serait alors plus seulement celle de la sophistication de ses contenus, mais de sa capacité à industrialiser, localiser et personnaliser son discours. Ce faisant, il apparaît que la montée en puissance de ces technologies appelle dès maintenant une vigilance accrue et une adaptation rapide des dispositifs de prévention de l’extrémisme violent.
Timbuktu Institute – September 2026
Generative artificial intelligence could well open a new front in the information battle in the Mali. Still limited, JNIM’s use of this technology already reveals the possibilities it offers for its propaganda: producing content faster, adapting it into local languages, tailoring narratives to different audiences and, eventually, potentially personalising recruitment. While the group is still only at the early stages of this appropriation, the speed at which technical barriers are falling raises a broader question: are violent extremism prevention mechanisms ready to respond to propaganda that is itself capable of changing scale?
The Malian conflict, marked for more than a decade by the entanglement of jihadist, separatist and state dynamics, may be on the verge of a new shift. The democratization of generative artificial intelligence is indeed opening up new possibilities in the information domain for the conflict’s various protagonists. This is a factor highlighted by the Timbuktu Institute’s report on the functions of generative AI among the conflict’s different actors, in particular the gradual appropriation of this technology by Jama’at Nusrat al-Islam wal-Muslimin (JNIM) for propaganda purposes. Its use remains limited for now, certainly, but its potential implications deserve close monitoring.
The first sign of this evolution appeared in June 2026, when JNIM’s media wing, Az-Zallaqa, released an AI-generated poster celebrating the deadly attacks carried out in late April 2026, which had claimed the life of Mali’s Minister of Defence. The visual depicted the group’s fighters in various combat scenes and displayed several markers typical of AI generation, notably buildings and vegetation blending into one another and excessive smoothing of textures. While the use of this technology remains confined for now to promotional material, it is reminiscent of a trajectory already followed by the Islamic State. As early as 2023, that group had developed a guide dedicated to the secure use of generative AI, before gradually incorporating chatbots and automated videos into its recruitment campaigns. The Islamic State’s experience thus offers a glimpse of a possible evolution: once an extremist organisation acquires both the technical capability and the organizational will to integrate generative AI into one part of its propaganda apparatus, its use then tends to spread to other functions.
Producing more, faster
It is perhaps in this light that JNIM’s first direct use of generative AI should be understood: the acceleration of its propaganda output. The group currently relies on internal teams and largely manual processes, which can take weeks or even months of work to produce a professional-quality video. With generative AI tools now available, this timeframe could be considerably shortened. From a simple text description, half a dozen professional-grade visuals could be generated within minutes, before being assembled into a video with animation, music and text. Such a shift would have a fairly immediate consequence: a sharp increase in the volume of content produced and, potentially, a far more viral presence for the group across the regional information space. In other words, the stakes would no longer lie solely in JNIM’s ability to produce sophisticated content, but in its ability to produce far more of it, far more quickly.
Changing language, changing audience
At the same time, a second shift could concern the linguistic and cultural localization of recruitment, against a backdrop of growing endogenisation of jihadism. JNIM today disseminates its messages mainly in classical Arabic, which necessarily limits their reach among populations for whom Arabic is not the language of daily communication. Neural translation, however, could change this equation. It would make it possible to generate, almost instantly, different versions of the same message in some fifteen regional languages, including Hassaniya, Fulfulde, Bambara, Mooré, Tamasheq, Songhai and Wolof.
But the change would run deeper than simple linguistic translation. This capability would allow the group to adapt its narratives to the different audiences it seeks to reach. A version aimed at Fulani herders could emphasize the protection of pastoral interests and defence against government marginalization. Another message, aimed at marginalized urban youth, could instead focus on dignity, economic inclusion and social status. For idealistic religious youth, the discourse could, in turn, foreground spirituality and resistance to Western imperialism. Each narrative could thus be produced quickly and simultaneously, with the potential effect of maximizing recruitment effectiveness for each population segment. AI dubbing would also open up another avenue. By making it possible to produce recruitment videos dubbed in Bambara, with a native-sounding synthetic voice, this technology could grant access to populations that have so far been largely excluded from the written propaganda market. Distributed via Telegram, WhatsApp or informal radio, such productions could more effectively reach illiterate or semi-literate young people in Mali, Burkina Faso and Niger.
From mass propaganda to tailor-made recruitment
The most concerning development, however, would be that of hyper-personalized recruitment. Here too, the Islamic State has already blazed a trail. The group had developed AI-driven chatbots to tailor its propaganda messages to the individual psychological vulnerabilities of targeted persons. The possibilities now offered by generative AI could allow JNIM to follow a similar path. A chatbot deployed on Telegram could gradually assess a young person’s aspirations, frustrations and convictions, before dynamically adapting its discourse to maximise psychological impact and encourage their move towards violent extremist groups. The change of scale would be considerable. Whereas traditional propaganda addresses relatively broad categories of the population, AI could potentially enable a shift towards individualized discourse, built around the specific characteristics of each interlocutor.
This is precisely what poses a new challenge for violent extremism prevention mechanisms. Current strategies – community outreach, religious counter-messaging, or economic inclusion programmes – are for the most part designed at a collective, generic scale. AI-assisted recruitment, by contrast, could operate at an individual and hyper-personalized level. This raises several questions: how can one counter a message dubbed in Mooré, Malinké, Wolof or Fulfulde without comparable AI-based multilingual production capacities? How can propaganda chatbots be identified without expertise in algorithmic detection? And how can illiterate populations be reached with prevention messages that are essentially written? These are all questions that field actors will need to answer in the coming years, or risk seeing the gap widen between the growing sophistication of recruitment mechanisms and the relative stagnation of regional resilience capacities.
A still-emerging threat, but one to watch closely
At this stage, however, the scale of the phenomenon should not be overstated. The Timbuktu Institute’s report notes that there is currently little public evidence of official adoption of generative AI by JNIM beyond its promotional material. Such caution is all the more warranted given that a report by the SITE Intelligence Group, published in May 2024, had already highlighted the difficulty of precisely gauging the scale of the opportunity AI represents for terrorist groups, given their continued reliance on conventional content-production methods. That said, the continuous lowering of technical barriers to accessing these tools makes increased experimentation by JNIM increasingly likely. Video generation, machine translation, multilingual dubbing or targeted recruitment could gradually enter its information toolkit. For now, then, the phenomenon remains emergent. But its significance lies precisely in this early stage: technological capabilities are advancing faster than the mechanisms meant to anticipate their uses. Should JNIM cross a new threshold in integrating generative AI into its propaganda apparatus, the question would no longer be simply one of the sophistication of its content, but of its capacity to industrialize, localize and personalize its discourse. In light of this, the rise of these technologies calls, even now, for heightened vigilance and rapid adaptation of violent extremism prevention mechanisms in the region.