
How can you recognise AI-generated videos used in Russian propaganda? Check the source and context, look for consistency in faces and details between frames, compare the sound with lip movements, and seek independent corroboration. No single ‘magic’ clue or detector is enough. Russian propaganda increasingly uses generative AI to produce emotional videos about the war in Ukraine. Ukrainian soldiers supposedly surrender or complain about their commanders, while children cry over their ‘fathers’ mobilisation’. Such content often aims to provoke fear, exhaustion, distrust of the state and a sense that defeat is inevitable.
This expanded article draws on journalist Yevhen Vakulenko’s report for Free Radio on Russian fabricated stories and identifying AI-generated videos, published on 7 July 2026. Its main explanations came from Andrii Yurychko, a journalist, media trainer, Candidate of Philological Sciences and lecturer at Taras Shevchenko National University of Kyiv. He has worked with the Academy of Ukrainian Press for many years on media literacy, fact-checking, critical thinking and responsible AI use.
Generative models have reduced the cost and time needed to produce false content. Creating dozens of videos no longer requires actors, a film crew or a real location. A prompt, a generated scene, an added voice and a network of accounts can be enough.
In Russian information operations, quantity sometimes matters more than quality. Even imperfect videos fill feeds with recurring images: ‘the Ukrainian army is demoralised’, ‘the state has abandoned people’, ‘mobilisation destroys families’ or ‘Western media supposedly acknowledge Ukraine’s defeat’. Seeing the same narrative repeatedly in different forms can make it feel familiar and therefore plausible.
Yurychko draws attention to this psychological dimension. A poor-quality video may not convince viewers literally, yet still create the emotional atmosphere its makers want. Verification therefore needs to examine the intended effect as well as the pixels: what reaction is being sought, and who benefits?
Generated scenes may show groups in uniform supposedly laying down their weapons or addressing commanders. These videos seek to demoralise soldiers and civilians, exaggerate Russian successes and create an illusion that the defence is collapsing.
A fictional character’s face and voice supply supposed ‘insider’ testimony. Viewers receive an emotional monologue rather than a verified fact, easily shared without a date, location, unit or other verifiable details.
Scenes involving children deliberately trigger compassion, anxiety and anger. The characters often inhabit an excessively clean, almost advertising-like environment, repeat similar phrases and have no verifiable history. Emotion is meant to precede the questions: who filmed this, where did it first appear, and do these people exist?
Propagandists copy newsrooms’ logos, captions, fonts and design, and create pages with similar names and web addresses. Visual resemblance is intended to borrow a genuine source’s reputation. A logo on screen is not evidence: look for the report on the outlet’s official website or verified account.
Yurychko considers the search for visual artefacts only the first level of verification. Generative models rapidly improve their rendering of faces, hands, lettering and movement, so lists of ‘typical errors’ keep becoming outdated. Questions about source, context, purpose and emotion remain useful regardless of the technology’s version.
Treat a suspicious clip as a possible element of an information operation: who published it, what picture of events does it impose, and what reaction does it expect? A technically imperfect fabrication can still work if people share it out of fear or outrage. The strongest defence is not guessing from one detail, but refusing to let emotion outrun verification.
Who first published the video? When was the account created? Does it have a history of genuine posts, contact details and a confirmed organisational connection? Similar names, extra characters or numbers in an address can disguise an imitation page.
Are the location, date, unit or people identified? Have official institutions and independent media reported the event? A sensational clip without verifiable details is a reason to pause, not to share.
As the head turns, the shape of the nose, lips or ears, the position of moles or the hairline may change. What matters is consistency throughout the clip, not one attractive frame.
Unusually infrequent blinking, excessively wide eyes, indistinct lip outlines or teeth becoming a smooth white strip during speech may suggest generation. Also check whether lip movements precisely match the sound.
Fingers, earrings, glasses, watches and wedding rings should remain consistent between frames. In hastily generated clips, they may disappear, change shape or merge unnaturally with skin or clothing.
Camouflage may ‘flow’ during movement, insignia may change and lettering on vehicles, buildings or clothes may consist of meaningless characters. Models are improving rapidly, but mass-produced propaganda often cuts corners on quality.
People in the background should not appear the same size as those nearer the camera. Unnaturally large hands, disproportionate shoulders or heads, and objects that do not fit their surroundings warrant closer examination.
Glasses, bottles, windows and mirrors should reflect surrounding objects correctly. A reflection inconsistent with the scene, a shadow pointing in an impossible direction or an object suddenly changing transparency may be a generation artefact.
Perfectly symmetrical faces, smooth skin, new clothes and an absence of dust or exhaustion in a supposedly frontline scene do not prove fabrication. But a mismatch between the claimed context and people’s appearance should raise questions.
Clues that expose generation today may disappear after tomorrow’s model update. Conversely, compression, poor lighting, editing or an unstable connection can create defects in genuine footage. An odd hand or blurred teeth are not a verdict, merely one part of the investigation.
Automatic detectors require similar caution. They can help formulate a hypothesis, but their percentages do not prove a file’s origin. AUP discusses their limitations in Why AI Text Detectors Cannot Be Reliable. The same principle applies to video: conclusions must rest on combined clues and sources.
There is no universal clue. The most reliable approach combines examination of the source, context, frame sequence, sound and independent corroboration.
They can serve as supporting tools, not conclusive evidence. Results depend on the model, file quality, editing and the service’s training data.
Do not share it immediately. Strong emotion is often part of an information operation’s design. Pause and check the author, original source and corroborating evidence.
Mass repetition of a narrative can create an atmosphere of distrust and hopelessness even when individual videos are not regarded as entirely truthful.
The Academy of Ukrainian Press works systematically with journalists, educators, students and trainers, combining fact-checking, critical thinking, digital security and practical AI use.
At Media Literacy 2.0: Using AI for Media Literacy and Education, participants studied AI basics, its relationship with disinformation, ethical rules and practical tools. Through Artificial Intelligence and Media Literacy: Tools and Recommendations, AUP helped educators understand new technologies as tools with specific capabilities and limitations, rather than magic.
AUP’s project for media literacy and counter-disinformation trainers prepared specialists to pass on these skills. Within the project, AUP President Valeriy Ivanov stressed that Russian information policy works through confusion and deception; AUP therefore equips participants to identify propaganda and apply critical thinking in everyday life.
Another large-scale example was the media education forum for 850 educators from eastern Ukraine, where AUP explained verification procedures and information hygiene. Such events turn individual pieces of advice into systematic skills that educators, journalists and trainers share with their communities.
The guiding principle is that technologies change quickly, but the habit of asking questions, checking sources and refusing to let emotion determine what we share remains our most durable protection.
About the original source. This is an independent, expanded AUP compilation based on Yevhen Vakulenko’s Free Radio article of 7 July 2026. Read the original Free Radio article in full. The outlet is linked in the second paragraph in accordance with its republication rules.
The Ukrainian text was structured and linguistically edited using AI tools. Its facts, links and final wording were checked and edited by a human.
Translation note: This is a literary translation into British English, produced and proofread using a language model. Read the Ukrainian original.