A
Tuareg fighter calling himself Azzghim has become a familiar face in Mali’s online war. He turns up in military dress in videos shared by accounts sympathetic to, or run by members of, the Azawad Liberation Front (FLA), and he spends most of his screen time attacking Russia’s Africa Corps.
He also appears not to exist. Researchers who studied the clips found the usual marks of synthetic video, in the way his body moves and in glitches in the color and the sound, and concluded that he had been generated rather than filmed. His following has not suffered for it. One account given over largely to reposting and remixing Azzghim material had 207,000 followers when the researchers looked at it. Another account in the same orbit used a synthetic teenage boy.
A face helps. An anonymous account can post the same message, but a man in fatigues talking into the camera gives it a body and a voice, and unlike the FLA’s real fighters, he can’t be captured or made to talk.
Azzghim is the most striking product of a wider experiment in Mali, and every side is taking part. In June, the media arm of Jama’at Nusrat al-Islam wal-Muslimin, the al Qaeda affiliate better known as JNIM, put out AI-generated posters to promote videos of its April attacks. Pro-government accounts have published synthetic clips celebrating the Malian army and its Russian partners. Russia-linked networks aimed at African audiences have used the same tools to write articles, translate them and put synthetic presenters in front of a camera.
It would be easy to read all this as the familiar story in which cheap AI hands powerful propaganda tools to anyone who wants them. What stands out in Mali is how differently the same software performs depending on who is holding it. A jihadist media office making posters, a few separatist sympathizers building a character on TikTok and a Russian operation running websites in several languages are doing very different jobs, and in each of them the software is the least expensive part.
A September report by the Dakar-based Timbuktu Institute compares how JNIM, FLA-linked accounts, networks supporting Mali’s military authorities and Russian-linked operations use generative AI. Of the four, JNIM’s documented use is the thinnest; the researchers describe it as “still at a limited but potentially evolving experimental stage.” The June posters remain the clearest example. The battlefield footage promoted by Az-Zallaqa, the group’s media arm, was real. The advertising wrapped around it was not, and it carried the usual giveaways, with blurred backgrounds and buildings melting into the vegetation behind them.
That modesty should not be mistaken for incapacity. JNIM already has what most influence campaigns spend years building: a media office, a back catalog of battlefield footage and channels that reliably put it in front of people. Generative tools could speed all of that up, turning a statement into Fulfulde or Tamasheq within the hour, dubbing a video for a different audience, or recutting the same claim several ways before anyone on the other side has drafted a reply. Nobody has yet shown that JNIM does any of this, and having the tools is a long way from knowing how to use them to any effect.
The pro-government side, meanwhile, has found an audience. One Bamako-based TikTok account identified by the Timbuktu Institute built up more than 436,000 followers with AI-generated material praising the army and the Africa Corps. After the April fighting it posted a video presenting Kidal as a city the government had recaptured. It drew 4.7 million views, far more than anything publicly documented from JNIM’s experiments.
That video came out of the fighting that began on April 25, when JNIM and the FLA launched coordinated attacks across Mali. Defense Minister Sadio Camara was fatally wounded in Kati after a suicide attacker drove an explosives-laden vehicle into his residence and a firefight followed. In Kidal, rival versions were circulating within hours.
The FLA said it had taken the city; the army called its departure a tactical repositioning, and Russian accounts dwelt on the Africa Corps’ resistance. Armies have always rebranded retreats and rebels have always inflated their gains, and an old photograph with a fresh caption can still travel a remarkable distance. What has changed is the volume and the speed. A single clip can now be recut, given a new voiceover and translated for a second audience before anyone has checked where it was filmed, and a small team can sustain an output that once needed a newsroom.
Russia’s operation sits in a different category. Moscow was feeding anti-French and anti-Western narratives into Malian politics long before generative AI arrived, and the outlets it built for that purpose have since absorbed the new tools. In 2025, France’s VIGINUM agency mapped the network around African Initiative and traced a related operation, AI-Freak, probably linked to it, which mixed AI-generated text, images and video with fake news sites, inauthentic accounts and search manipulation. A model can turn out a script or a translation in minutes. Deciding who should see it, which accounts should carry it and where it will look credible is still work done by people.
The output dwarfs the audience. Between September 2023 and April 2025, African Initiative’s website published 18,127 articles in several languages. VIGINUM, citing SimilarWeb, counted 104,818 unique visits between December 2024 and February 2025, around 35,000 a month. Its six official Telegram channels had 69,720 subscribers between them.
OpenAI’s encounter with Stop News, another Russian operation, is starker still. OpenAI disrupted it in 2024, after its operators used AI to produce articles, comments and images in English, French and Russian for sites posing as news outlets in Africa and Britain. Few operations have leaned so heavily on synthetic imagery, and most of its accounts attracted almost no engagement. It came back the following year with more ambition, including YouTube and TikTok channels fronted by a synthetic newsreader who praised Russia and the Africa Corps and attacked France, the United States and Ukraine.
By August 2025, Newstop Africa’s main account on X had 172 followers, and its TikTok and YouTube channels about 1,900 each. One TikTok video reached 63,300 views; another reached 87. OpenAI found little sign that the network had spread beyond its own accounts and placed it in Category 2 on its six-point Breakout Scale.
Mali’s audiences are not blank screens. They come to these videos with years of war behind them and settled views about France, Russia and the junta, and they judge a message partly by who sends it. Some players entered this contest with large followings and voices people already recognized, able to push a clip across several platforms in an afternoon. Others started with almost nothing. A shaky battlefield video from an account people trust will usually travel further than a polished synthetic image from a source nobody knows.
Checking is slow in Mali and getting slower. Access to conflict zones has narrowed. RFI and France 24 have been off the air since 2022, when the authorities made their suspension “definitive.” In June, the authorities detained the journalists Chahana Takiou and Abdrahamane Keïta under the cybercrime law, Takiou after criticizing its use against the press and Keïta after saying on television that JNIM controlled Kidal.
Between an event and an independent account of it there is now a gap of days, sometimes longer, and the army, the armed groups, Russian-linked outlets and their partisans all rush to fill it. Generative AI shortens the time it takes them to get there. The state, for its part, has been working on other ways of finding out what Malians say to one another.
Recommended
Persuasion, in other words, is slow and unreliable work, and the Malian authorities have not bet everything on it. In September, Anthropic reported that a Claude user it assessed as likely to be an independent consultant in Bamako, working for the state intelligence service, had relied on the model as the main engineering resource for a platform called Lakana 360.
The system was designed to monitor data linked to roughly 25 million SIM cards across all three national mobile operators, collecting call records, messages and voice traffic, linking users across SIM cards and generating dossiers on individual numbers. Claude helped write the software; the surveillance was left to the people running it. Anthropic banned the account and said the ban interrupted the developer’s work without disabling the system. Azzghim, for all his 207,000 followers, still has to win people over. A system built to listen to 25 million SIM cards does not.
Stay informed by subscribing to
our weekly newsletter.
