Videos circulating since late 2025 that appear to show Ukrainian soldiers weeping, apologizing to the Russian people, and surrendering en masse are authentic front-line footage
By The Conspiratory EditorsAugust 2, 2026
Where the evidence lands: ContradictedThat videos circulating from November 2025 onward, showing Ukrainian soldiers weeping, apologizing to Russians, denouncing their own government, or surrendering in large numbers, are authentic recordings of real Ukrainian servicemen, and therefore evidence of collapsing morale and mass capitulation at the front.
Contradicted. The claim rated here is that a body of videos spreading since early November 2025, showing Ukrainian soldiers crying for their mothers, blaming their own government, and surrendering to Russian forces in large numbers, is genuine footage from the front. It is not. NBC News reviewed 21 of them and published its findings on 13 December 2025. In at least half, the corner watermark of OpenAI's Sora 2 was still visible. The clips that appeared to come from other AI platforms carried subtler errors instead, notably uniforms and helmets rendered in incorrect or simplified versions of what Ukrainian soldiers wear. A third tell was obvious to Ukrainians and invisible to everyone else: most of the soldiers were speaking Russian, and only eight of the 21 videos featured personnel speaking Ukrainian. Some clips used the face of Aleksei Gubanov, a Russian livestreamer who left Russia over his criticism of Vladimir Putin and now lives in New York. Two things need separating. Ukrainian soldiers have surrendered, been captured, and died in this war, and real footage of prisoners exists; none of that is in question and none of it is diminished here. What this file rates is the authenticity of these specific clips, which is debunked. Attribution is a genuinely open question: who produced and posted them has not been established, and this file names no perpetrator.
Believed by: Circulated heavily through pro-Russian and war-commentary accounts, and picked up by ordinary viewers on autoplay feeds who had no reason to inspect a short clip of soldiers in a trench. The videos were not aimed only at Russian-speaking audiences; they spread in English-language feeds too, where uniform details that would look wrong to a Ukrainian viewer meant nothing.
Opening entry, with the state of the record as of early August 2026. No clip from this body of material has been authenticated, and no authorship has been established; the file names nobody and will not until someone publishes evidence rather than inference. The thread most worth following is what this case says about watermarking as a defense. Sora's mark was present in roughly half the circulating videos and did not stop them spreading, which points at viewer attention rather than technical provenance as the binding constraint. If provenance signals are going to work, they will have to be surfaced by the platforms rather than left in the corner of the frame for people to notice. source →
The full story
A war's worst footage, generated to order
The clips began appearing in early November 2025. Ukrainian soldiers weeping in trenches. Soldiers calling for their mothers. Soldiers apologizing to the Russian people and blaming their own government for the war. Groups of soldiers surrendering to Russian forces. The footage was ultrarealistic, short, emotionally direct, and it moved quickly across YouTube, TikTok, Facebook and X.
Much of it was made with Sora 2, OpenAI's text-to-video model, released with an invite-only app at the end of September 2025. NBC News reviewed 21 of the videos and reported on 13 December 2025that in at least half of them nobody had bothered to remove the model's own watermark.
Two tells, one loud and one quiet
The loud one barely counts as forensics. Sora stamps generated video with a small moving logo, and it was still there. That is a provenance signal working exactly as designed and failing anyway, because a mark in the corner of an autoplaying clip reads to most viewers as ordinary platform branding.
The quieter ones mattered for the clips that came from other generators. NBC reported errors most people would find hard to spot, principally uniforms and helmets rendered in incorrect or simplified forms of what Ukrainian soldiers actually wear, the characteristic failure of a model approximating equipment it has seen relatively little of.
The third signal needed no expertise at all, only the right nationality. Across the 21 videos NBC reviewed, only eight featured anyone speaking Ukrainian. The majority of these supposed Ukrainian soldiers were speaking Russian, which is obvious to a Ukrainian viewer and completely invisible to a scrolling English-speaking one. Several clips also reused the face of Aleksei Gubanov, a livestreamer who left Russia over his criticism of Putin, lives in New York, and has never served in any army. He found out because his own followers flagged it.
The strongest objection: the war is real
The best argument for these videos is not that any particular clip is authentic. It is that they depict something true in outline. Soldiers on both sides have surrendered. Prisoners are taken and exchanged. Morale at the end of a long war is not a fiction, and neither is exhaustion, fear, or a man who wants to go home.
All of that is correct, and it is precisely what makes a fabrication effective. But a true background does not authenticate a specific document. The claim these clips actually carry is narrower and stronger than "war is terrible": it is mass capitulation and collective apology to Russia, and it is supported by nothing but generated footage. Accepting synthetic video as rough evidence for a real trend delivers exactly the result its maker wanted, at no cost to them.
Who made them is not known, and saying so matters
The content aligns with Russian messaging, and NBC described the videos as a disinformation salvo aimed at warping perceptions of the war. That is a fair reading of their function. It is not a finding about authorship, and NBC said as much: it is not clear who created or posted them.
This file holds that line. Generative video is available to state-run information operations, to freelance propagandists working unpaid, and to engagement farms indifferent to the politics and interested in the views. Any of those could produce this material, and different clips may well have different origins. Naming a culprit because the content suits their interests is the same reasoning error the videos themselves rely on: it substitutes what fits for what is shown.
The case that gets cited afterwards
By January 2026 the episode had become a reference point. In NBC News reporting on what researchers describe as a collapse of trust online, these videos appear alongside the AI-generated material that surrounded the US operation in Venezuela and the arrival of AI-generated evidence in courtrooms, cited as a case where fabrications had fooled officials rather than only the public.
That is the part with consequences beyond this war. A fake that fools an official is not merely a fake that spread widely; it is one that entered a decision. And each such case makes the reverse move cheaper, so that a genuine video of a genuine surrender becomes easier for anyone to wave away.
Keep the war and the footage separate
Two things, held apart. The war is real, the prisoners are real, the dead are real, and none of that is softened by anything on this page. Evidence about the front comes from reporters, prisoner exchange records, and international monitors, and it always did.
The claim this file rates is that these particular clips are footage. They are not; they were generated, half of them still wearing the label that says so. And the honest gap is left open: nobody has established who made them, so this file names nobody. A debunk that overreaches into an accusation would be making the same move it is criticising.
What's still unexplained
- Who produced and seeded the videos is unresolved. Reporting has not established authorship, and this file makes no attribution; the possibility that different clips have different origins has not been excluded either.
- How much of the material was ever believed, as opposed to knowingly shared by accounts pushing a line, is unmeasured. Reach figures do not distinguish between persuasion and participation.
- Whether visible watermarking helps in practice is genuinely in doubt after this case. The mark was present in about half the sample and still did not stop circulation, which suggests the constraint is viewer attention rather than technical provenance.
- Whether platform enforcement improved after the episode has not been independently assessed. The clips spread across four major platforms simultaneously, and no public accounting of what was removed, or how quickly, has been published.
Point by point
The claim: The footage looks completely real, so it must have been filmed at the front.
What the record shows: Photorealism stopped being evidence of provenance in 2025. These clips came from Sora 2, released at the end of September 2025 with a consumer app, and their realism is the product rather than a sign of authenticity. The decisive point is that the model labelled its own work: NBC News found the Sora watermark still present in at least half the circulating videos. A fabrication can be indistinguishable from footage at a glance and still be identifiable by where it came from.
The claim: The watermark argument only covers the ones with a watermark. The rest could be real.
What the record shows: The unwatermarked clips were not left unexamined; they simply failed different checks. NBC found that videos which appeared to come from other AI platforms carried consistent errors in the kit, uniforms and helmets rendered in incorrect or simplified versions of what Ukrainian forces wear, the sort of approximation a generator produces for equipment it has seen little of. A second inconsistency ran across the whole sample: most of the soldiers were speaking Russian, and only eight of the 21 videos featured anyone speaking Ukrainian, which is a strange property for footage of the Ukrainian army. Several clips also reused the face of a Russian livestreamer who has never served in any military. A batch sharing a period, a distribution pattern and a set of characteristic errors is assessed as a batch, not clip by clip in isolation.
The claim: Ukrainian soldiers really do surrender and really are captured, so these videos are showing something true even if some clips are synthetic.
What the record shows: The premise is correct and the conclusion does not follow. Soldiers on both sides have surrendered, been captured, and been killed; that is documented through conventional reporting, prisoner-exchange records, and international monitoring. A real underlying phenomenon is what makes a fabrication credible, not what makes it true. The specific claim these videos carry, mass capitulation and collective apology to Russia, is not established by clips that were generated, and treating a synthetic clip as approximate evidence of a real trend hands the fabricator exactly the outcome they wanted.
The claim: Calling them fake is itself propaganda; every inconvenient video now gets dismissed as AI.
What the record shows: That reflex is real and this site covers it as the liar's dividend, so the objection deserves a straight answer rather than a dismissal. The answer is that these clips were not waved away on vibes. They carry the generating model's own watermark in half the sample, contain a consistent class of equipment errors, and in places reuse identifiable faces from elsewhere. That is positive evidence of fabrication, not an absence of evidence for authenticity. The difference between a debunk and a dismissal is whether anything was checked.
The claim: If Russia made these, someone would have proved it by now.
What the record shows: Nobody has, and this file does not assert it. The content is plainly aligned with Russian messaging aims, and NBC frames it as a disinformation salvo intended to warp perceptions of the war, but alignment of interest is not evidence of authorship. Generative tools are available to states, to freelance propagandists, and to engagement farmers who care only about views. The honest statement is that the videos are fabricated and their origin is unestablished, and no amount of confidence about who benefits converts into a finding about who acted.
Timeline
- 2022-03The technique has a precedent early in the full-scale invasion. A crude deepfake of President Volodymyr Zelenskyy appearing to tell Ukrainian troops to lay down their arms circulates and is debunked within hours; it is poorly made and widely mocked. The idea of manufacturing a surrender is established years before the tools can execute it convincingly.
- 2025-09-30OpenAI releases Sora 2 alongside an invite-only iOS app, sharply lowering the cost and skill needed to produce short, photorealistic video from a text prompt. Downloaded video carries a visible moving watermark plus C2PA provenance metadata.
- 2025-11The videos appear. Clips spread showing Ukrainian soldiers weeping, calling for their mothers, apologizing to the Russian people, and blaming Kyiv for the war, alongside footage of groups appearing to surrender to Russian forces. They circulate on YouTube, TikTok, Facebook and X.
- 2025-11One of the faces recognizes itself. Aleksei Gubanov, a Russian livestreamer who left Russia over his criticism of Vladimir Putin and now lives in New York, sees his own likeness in uniform, speaking lines he never said. He has never served in any army. 'At first I didn't take it seriously,' he later tells NBC News, 'but soon I noticed how quickly the view count was growing and, even worse, how many people believed these videos were real.'
- 2025-12-13NBC News publishes a review of 21 of the videos, and the findings are specific. In at least half, the Sora 2 watermark is still in the corner. Those that appear to come from other AI platforms carry subtler errors instead, notably incorrect or simplified versions of the uniforms and helmets Ukrainian soldiers wear. And most of the soldiers are speaking Russian: only eight of the 21 clips feature personnel speaking Ukrainian, an inconsistency obvious to a Ukrainian viewer and invisible to a foreign one.
- 2025-12-13Attribution stalls, and stays stalled. The reporting establishes what the videos are and where they spread, but not who generated or seeded them. NBC states plainly that it is not clear who created or posted them, and this file adopts that position rather than filling the gap.
- 2026-01-09The episode is cited as a benchmark case. In an NBC News report on what experts describe as a collapse of trust online, the Ukrainian surrender videos are listed among the AI fabrications that had fooled officials, alongside AI-generated material around the US operation in Venezuela and AI-generated evidence reaching courtrooms.
- 2026-08State of the record. No clip in this body of material has been authenticated as genuine footage. The war continues and real prisoners and real casualties are separately documented through conventional reporting, which is where evidence about the front actually comes from.
Contradicted. The claim rated here is that a body of videos spreading since early November 2025, showing Ukrainian soldiers crying for their mothers, blaming their own government, and surrendering to Russian forces in large numbers, is genuine footage from the front. It is not. NBC News reviewed 21 of them and published its findings on 13 December 2025. In at least half, the corner watermark of OpenAI's Sora 2 was still visible. The clips that appeared to come from other AI platforms carried subtler errors instead, notably uniforms and helmets rendered in incorrect or simplified versions of what Ukrainian soldiers wear. A third tell was obvious to Ukrainians and invisible to everyone else: most of the soldiers were speaking Russian, and only eight of the 21 videos featured personnel speaking Ukrainian. Some clips used the face of Aleksei Gubanov, a Russian livestreamer who left Russia over his criticism of Vladimir Putin and now lives in New York. Two things need separating. Ukrainian soldiers have surrendered, been captured, and died in this war, and real footage of prisoners exists; none of that is in question and none of it is diminished here. What this file rates is the authenticity of these specific clips, which is debunked. Attribution is a genuinely open question: who produced and posted them has not been established, and this file names no perpetrator.
Reviewed by The Conspiratory Editors · Last reviewed August 2, 2026 · How we rate
Common questions
Is Ukrainian 'surrender' videos true?
Contradicted. The claim rated here is that a body of videos spreading since early November 2025, showing Ukrainian soldiers crying for their mothers, blaming their own government, and surrendering to Russian forces in large numbers, is genuine footage from the front. It is not. NBC News reviewed 21 of them and published its findings on 13 December 2025. In at least half, the corner watermark of OpenAI's Sora 2 was still visible. The clips that appeared to come from other AI platforms carried subtler errors instead, notably uniforms and helmets rendered in incorrect or simplified versions of what Ukrainian soldiers wear. A third tell was obvious to Ukrainians and invisible to everyone else: most of the soldiers were speaking Russian, and only eight of the 21 videos featured personnel speaking Ukrainian. Some clips used the face of Aleksei Gubanov, a Russian livestreamer who left Russia over his criticism of Vladimir Putin and now lives in New York. Two things need separating. Ukrainian soldiers have surrendered, been captured, and died in this war, and real footage of prisoners exists; none of that is in question and none of it is diminished here. What this file rates is the authenticity of these specific clips, which is debunked. Attribution is a genuinely open question: who produced and posted them has not been established, and this file names no perpetrator.
What is Ukrainian 'surrender' videos?
In early November 2025, videos began circulating that appeared to show Ukrainian soldiers at their lowest: weeping, calling for their mothers, apologizing to the Russian people, blaming their own government for the war, and in some clips surrendering to Russian forces in groups. The footage was ultrarealistic and it traveled fast across…
What does the evidence show?
Photorealism stopped being evidence of provenance in 2025. These clips came from Sora 2, released at the end of September 2025 with a consumer app, and their realism is the product rather than a sign of authenticity. The decisive point is that the model labelled its own work: NBC News found the Sora watermark still present in at least ha…
Why do people believe it?
The war is real, long, and grinding, and exhaustion at the front is genuine and widely reported. A clip showing an exhausted soldier is not asking a viewer to believe anything implausible.
What is still unresolved?
Who produced and seeded the videos is unresolved. Reporting has not established authorship, and this file makes no attribution; the possibility that different clips have different origins has not been excluded either.
Sources
- 1.As war with Russia drags on, ultrarealistic AI videos attempt to portray Ukrainian soldiers in peril, NBC News (2025)
- 2.AI is intensifying a 'collapse' of trust online, experts say, NBC News (2026)
- 3.AI videos of Ukrainian soldiers surrendering to Russia is the war's latest disinformation salvo, AOL (NBC News syndication) (2025)
- 4.Those videos of Ukrainian soldiers crying for their mothers and apologizing? They're AI-generated deepfakes, Cybernews (2025)
- 5.As war with Russia drags on, ultrarealistic AI videos attempt to portray Ukrainian soldiers in peril, Yahoo News (NBC News syndication) (2025)
Embed this verdict on your site
Paste this snippet to show our sourced verdict as a small card, with a link back to the full case file. Free to use.
<iframe src="https://theconspiratory.com/embed/ukraine-surrender-deepfakes" title="The Conspiratory verdict" width="520" height="190" style="border:0;max-width:100%" loading="lazy"></iframe>Help us investigate
This is a living case file. If you spot an error or know evidence we missed, tell us, and weigh in on where you land.
Where do you land?
Cast your read on this one.
Comments
Add your take. Comments are read and approved by a human before they appear, so keep it on topic and civil. Please do not accuse named, living people of crimes.