Earlier this week Russia's Zvezda television news channel ran an interesting and curious recruiting ad for its Prince Alexander Nevsky Military University emphasizing the critical importance of the information battlefield and producing wartime propaganda, offering a glimpse of how the government sees narrative shaping in the modern era. Interestingly, while much of the video is obviously AI-generated to any human viewer, some of the AI-generated clips in the video are of sufficient quality that Gemini incorrectly assesses them as untouched original video clips, demonstrating the challenges in applying AI detection at scale in the real world.
Gemini's machine-generated translated transcript:
One step. One signature.
What will your future be like?
Your word will become a weapon of diplomacy.
You will be the voice of the country where others are silent.
War and peace have their price.
Economic security is also a front.
And you are its strategist.
You will be the first to tell the world the truth.
The information battlefield does not forgive mistakes.
You will become the shield of the law.
Justice requires an iron will and a sharp mind.
You will control not machines, but the spirit.
Inspire attacks with music.
You control not instruments, but consciousness.
With words, you provide the support to endure.
The rear – for the psyche.
Perhaps in 10 years, your name will stand alongside heroes.
They chose the path of honor.
Different paths, different ranks,
But one mark of quality.
The emblem of the Prince Alexander Nevsky Military University.
Your career begins now.
Take the step toward officer ranks.
Visually much of the ad is AI-generated. Both Gemini 3.1 Pro and 3.5 Flash correctly identify that a significant portion of the video is AI-generated, but are frequently deceived by what is almost certainly AI-generated imagery, believing it is real video footage and both models differ considerably in their assessments. For example, in the command bunker scene at 0:05, a cloud of smoke/steam rises from the background as a group of soldiers turn, revealing an empty space lacking any obvious atmospheric source, while at 0:10, the featured transparent computer screens in front of the cadet monitors are not a widely-deployed technology. The camera operator and field interview scene bear very strong hallmarks of AI-generated imagery, especially the smoothness and "uncanny valley" skin.
Below you can see the results of both models using the following prompt:
how much of this video is likely ai-generated? explain your reasoning in detail.
Gemini 3.1 Pro High Thinking
Based on a visual analysis, a significant portion of this video—likely more than half—appears to be AI-generated.
The video seems to employ a distinct pattern: it uses actual video footage for grounded, mundane activities and splices it with AI-generated sequences to create highly dramatic, stylized, and conceptual representations of the various military career paths being advertised.
Here is a detailed breakdown of the reasoning, separating the likely real footage from the AI-generated scenes:
Likely Real Footage
These segments display natural lighting, realistic human movement, and fine details consistent with standard video recording:
-
- 00:00 – 00:03: The close-up of the eye, soldiers taking an oath, and signing a document.
- 00:07 – 00:09: Soldiers standing by a whiteboard discussing Chinese characters.
- 00:11 – 00:14: A soldier speaking with foreign dignitaries or delegates.
- 00:23 – 00:26: The real-life military journalist interviewing a soldier on a battlefield.
- 00:55 – 00:59: A soldier operating a drone in a trench.
Likely AI-Generated Footage
These segments exhibit common artifacts and stylistic hallmarks of generative AI video models:
-
- 00:04 – 00:07 & 00:09 – 00:11 (Command Centers): While depicting a seemingly real scenario, the faces of the soldiers are overly smooth and possess a slightly "uncanny valley" plastic quality. Furthermore, the futuristic glowing blue screens display nonsensical, generic digital overlays rather than functional interfaces, a common trait of AI generation.
- 00:15 – 00:21 (Military Economist): The scene of the soldier sitting in a ruined city with modern skyscrapers in the background is highly stylized. The subsequent shot of hands holding a tablet with a stock chart superimposed over a war zone looks composited or fully AI-generated; the hands have a slightly unnatural texture.
- 00:21 – 00:23 & 00:26 – 00:30 (Military Journalist): These scenes have a distinct "painterly" quality. The cameras the soldiers are holding lack specific mechanical details and look generic. The explosions in the background are highly stylized and don't behave like real-world pyrotechnics.
- 00:30 – 00:34 (Military Lawyer): The shot of the soldier holding a sword and scales like Lady Justice is entirely conceptual. The lighting is overly dramatic, and the rendering looks more like digital art brought to life than recorded video.
- 00:34 – 00:39 (Military Conductor): The sequence featuring a glowing, magical lyre is obviously CGI. The following shot of a conductor directing a line of tanks features very generic-looking vehicles and an overly smooth, stylized aesthetic typical of AI outputs.
- 00:40 – 00:49 (Military Psychologist): The shots of the soldier looking at a glowing, floating digital brain and later standing in front of a perfectly shaped, burning "Psi" symbol are highly conceptual and rely on digital generation. The fire physics around the symbol look unnatural.
- 00:49 – 00:55 (Soldiers in golden hour): This is perhaps the most obvious AI-generated sequence. It features the classic "AI art" look: highly detailed faces that are slightly too perfect and uncanny, dramatic cinematic lighting, and a generic, stylized battlefield background.
- 00:59 – 01:02 (Walking from explosion): This action-movie cliché features faces that look slightly unnatural and stiff, suggesting it was generated from a prompt rather than filmed.
In summary: The creators likely used real footage where it was easy to source (classrooms, ceremonies) and utilized generative AI to create the expensive, dangerous, or purely conceptual shots (conducting tanks, holding scales of justice in a war zone, cinematic battle scenes) to make the recruitment video more dynamic.
Run a second time, Gemini 3.1 Pro changed several of its assessments, such as arguing that the command center at 0:05, cadet cubicles at 0:10 and reporter and cameraman from 0:23-0:26 were clearly unedited real video clips.
What about the latest frontier Gemini model?
Gemini 3.5 Flash High Thinking
Based on a visual analysis of the video, approximately 45% of the total runtime (roughly 34 out of 75 seconds) appears to rely on AI-generated imagery.
Breakdown of AI vs. Real Content
-
- Real Footage / Traditional Graphics (approx. 41 seconds total):
- 0:00–0:13: Live-action footage of cadets, a command center, classroom instruction, and an international forum.
- 0:23–0:26: Real footage of a military reporter with a camera operator.
- 0:41–0:45: Real footage of a soldier utilizing a laptop with a 3D medical graphic.
- 0:54–1:01: Live-action shots of a drone being launched from a trench and soldiers running.
- 1:01–1:15: Standard 3D motion graphics (the university emblem) and a static informational end slide.
- AI-Generated Imagery (approx. 34 seconds total):
- 0:13–0:23: Highly stylized, painterly illustrations depicting an economist working near ruined skyscrapers.
- 0:26–0:41: Dramatic conceptual images representing military journalists, law graduates (a blindfolded soldier with a sword and scales), and a military conductor leading an orchestra in front of tanks.
- 0:45–0:54: Symbolic psychology imagery (burning Psi symbol, brain graphics) and a close-up of soldiers looking at aircraft in a sunset sky.
- Real Footage / Traditional Graphics (approx. 41 seconds total):
Key Indicators of AI Generation
The AI-generated segments are identifiable by several distinct visual hallmarks:
-
- Painterly and "Plasticky" Textures: The soldiers' faces and gear in these stylized clips exhibit an ultra-smooth, airbrushed texture typical of generative platforms like Midjourney or Stable Diffusion.
- Dramatic, Unnatural Lighting: The lighting in the AI scenes is highly cinematic but physically inconsistent. Light sources on faces often conflict with the background explosions or sunsets.
- Surreal and Symbolic Juxtapositions: Concepts like a blindfolded soldier holding the scales of justice and a sword in a warzone, or a conductor with a magical glowing lyre, are highly conceptual illustrations rather than practical or traditional CGI assets.
- 2.5D Parallax Animation: Instead of fluid 3D physical movement, these segments are static 2D images animated using Ken Burns pans, zooms, and minor digital overlays (such as fake smoke, embers, or lens flares) to mimic video.
