Fake Footage, Real Fear: AI Is Rewriting the Visual History of the Ukraine War
Russia’s full-scale invasion of Ukraine, launched in February 2022, was accompanied from the very first hours by an information war. That war has never stopped. It is fought not only on the front lines of Donbas and the southern steppe, but also on Telegram channels, TikTok feeds, Facebook pages, and television screens across Europe. In recent months, according to monitoring by the European Digital Media Observatory (EDMO), a network of independent fact-checkers and researchers supported by the European Union, Ukraine has consistently been the main target of disinformation circulating in the EU, alongside migration. The narratives are recurrent and familiar: Ukraine’s government, military and population are portrayed as broadly pro-Nazi; Ukrainian refugees are depicted as violent or parasitic; European countries are accused of treating them better than their own citizens; President Volodymyr Zelenskyy and his entourage are presented as drug users and corrupt officials; the war is said to be NATO’s fault; and the West, not Russia, is accused of seeking escalation. These stories are not random. War propaganda has existed for as long as war itself, and its function is straightforward: to bolster the morale of one’s own side and to undermine that of the adversary. Both sides in any conflict use it. What has changed in the Ukraine war is not the existence of propaganda, but the technology available to manufacture it.
Over the past two years, the most striking development has been the growing use of Artificial Intelligence to create short videos, usually lasting between five and fifteen seconds. These clips are engineered to look like genuine eyewitness footage, even though many of them are not technically deepfakes. A classic deepfake normally involves replacing the face or voice of a real person in an existing video. Many of these AI-generated videos are different: they are entirely synthetic scenes, built from scratch by generative models, but designed to depict precisely what propaganda would like the world to believe. The short format is no accident. Five to fifteen seconds is long enough to deliver a powerful, emotional image and short enough to be shared rapidly before anyone has time to scrutinize it. In late 2025, EDMO detected a significant wave of such videos in connection with the Russian offensive on Pokrovsk. Numerous pro-Russian accounts spread AI-generated clips showing Ukrainian soldiers surrendering or crying. In September, a new wave was identified, this time focusing on the mobilization of Ukrainian soldiers to the front. Several videos and images showed soldiers in handcuffs on buses, being forcibly taken away. The timing was probably not coincidental. Ukrainian armed forces had just achieved a series of successes in the Donbas area, in particular the liberation of several towns north of Lyman as part of Operation Vivaldi. At the same time, reports were intensifying of a possible new mobilization in Russia. The videos offered a ready-made counter-narrative: Ukraine, not Russia, was the country bleeding, collapsing and forcing its own men into a hopeless war.
Russian disinformation is not aimed at a single audience. The same images and clips can be calibrated to serve different purposes in different places. For the Russian public, they confirm the official narrative that Ukraine is weak, morally rotten and nearing defeat. For the Ukrainian public, they are a psychological weapon, designed to spread fear, to demoralize soldiers and their families, and to discourage young men from complying with mobilization orders. For public opinion in European countries allied with Kyiv, they are meant to undermine the image of Ukraine as a deserving victim and to suggest that Western aid is propping up a corrupt and coercive regime. This multi-audience strategy explains why the videos are being recirculated in several European countries with subtitles in different languages. The goal is not merely to deceive; it is to create a parallel reality in which the moral roles are reversed. AI is not the only tool in this campaign. The old technique of taking real videos or images and distorting their meaning with false captions remains a central part of the effort. A particularly telling example involved a protest in Kyiv against the cutting down of trees. Pro-Russian disinformation circulated footage of that protest with the false claim that it showed Ukrainian civilians being taken away by soldiers because they had tried to disrupt a recruitment operation. The original protest was real; the military context was invented. This combination of authentic images and fabricated framing is often more dangerous than a fully synthetic video, because it is harder for casual viewers to identify and easier for fact-checkers to miss in the flood of content.
The most concerning element, according to the EDMO monitoring, is that these propaganda and disinformation operations are still relatively low quality and limited in coordination. They combine generative AI with older manipulation techniques, but they have not yet reached the level of a seamless, centralized, industrial-scale machine. That should not be a reason for comfort. Even at this relatively primitive stage, the content is realistic enough to be believed by the more vulnerable segments of the public. The potential for technical and tactical improvement is enormous. Today’s AI-generated clips are often short and contain visible flaws in hands, faces, lighting, or sound. Tomorrow’s clips may be longer, more coherent, and virtually impossible to distinguish from authentic footage. If this type of content is already capable of causing harm by being accepted as genuine, it is chilling to think what it could do in the future. A few years ago, producing a realistic fake video required sophisticated skills and expensive equipment. Now it can be done with publicly available tools in a matter of minutes. As the technology improves, the cost falls and the quality rises. The barrier to entry for creating convincing synthetic evidence is disappearing, and the platforms where these videos circulate are often poorly equipped to detect content that is not a classic deepfake but a fully invented scene.
The broader danger is cumulative. A single fake video can be debunked, but a continuous flood of fake videos has a different effect. It creates a fog of uncertainty in which nothing can be fully trusted. Audiences may begin to doubt genuine footage of real events because they have been misled too many times. That erosion of trust is itself a strategic objective. If European citizens can no longer distinguish between real and fabricated images from Ukraine, their support for Kyiv may weaken, not because they have made a conscious political choice, but because they no longer know what to believe. The consequences could extend far beyond this war. The same generative tools can be used in future conflicts to manufacture false evidence of atrocities, fabricate statements by political leaders, or stage events that never took place. The next generation of AI propaganda will not need to be perfect. It will only need to be plausible enough to seed confusion, to slow down fact-checkers, and to arrive faster than platforms and regulators can respond. Fact-checkers can debunk individual videos, but they cannot debunk an entire ecosystem of deception. The battle for truth is becoming a battle of speed and scale, and the current defenses are not built for it.
For now, the most visible AI-generated videos are still relatively crude, but the direction of travel is unmistakable. The war in Ukraine is being fought not only with missiles, drones and artillery, but with images, algorithms and synthetic realities. The same techniques that are now being used to show Ukrainian soldiers surrendering or being handcuffed on buses could one day be used to create evidence of events that never happened, to blackmail leaders, or to tip an election at a moment of crisis. Countering this threat will require much more than after-the-fact fact-checking. It will require investment in detection technologies, closer cooperation between platforms and independent researchers, stronger media literacy programs for citizens, and a political recognition that the information space is part of the battlefield. There is no single solution. Technical tools can be defeated by better AI. Platform policies can be circumvented by new accounts and encrypted channels. Media literacy is essential, but it takes time and cannot keep pace with every new trick. Governments and civil society need to work together to expose the methods of disinformation, not just the content. The EDMO monitoring is a warning, not a solution. It shows that the phenomenon is already here, and that the operations behind it are learning. The next wave of synthetic disinformation is likely to be more convincing, more widespread, and more damaging than anything seen so far. Democracies need to prepare for that wave before it breaks. By the time a convincing fake is everywhere, it may already be too late to persuade the public of what is real.
