Michael Chernoff

Video Artist | Researcher | Educator


Indestruction (AI Video Project) 2025


Indestruction, 3:35 min (2025) Generated Using SORA by OpenAI


Above the short film titled Indestruction featuring a series of video climps, sampled, extended, and compiled into video editing software, with stock audio effects added. The clips themselves are AI generated videos made using SORA, a web based model for making still and motion image content. 

This short film amusingly depicts a building undergoes waves of destructive attacks yet never falls down. The building never takes damage eventhough extreme elements explode from the windows and roof. This film however, was not intenionally sought but was coincidentally made when I noticed how SORA could not fulfill my original prompt. Rather than look at the generated video as a failure and move on to keep further refining, what I saw was a programmed limit on an AI model. The theme of an indestructible building was not made by me but by the AI model itself. 

Below I summarize this process of critical engagement with an AI gen model through play, repetition, and prompt that function as tests rather than effects for finished scenes. 


Prompting & Cutting


Often when working with generative AI video, I resort to prompts that have a destructive description.  Initially I wanted to see an impossible thing happen. I wanted to view a building being destroyed with dynamite and then endlessly keep collapsing. I imagined a waterfall of falling concrete, metal, glass, steel, and wires cascading to the ground. Yet, this phsically impossible sight was also a codified as an impossibility for the SORA AI model to make happen. 








Only the top of the building was animated to mimic destructive impact, but when the dust and debris vanish and the building remained standing. I found it interesting and theorized that the building could not be destroyed. The idea is similarl to how AI models (or at least SORA) cannot depict harm or violent actions due to restrictions on disinformation. Below is another example where my prompt did not allow a human to hurt an elephant, resulting in awkward and unrealistic movements.






I decided to push the animation of the building explostion further by using the Re-Cut tool to extend the generated video clip. I grabbed and minimized the clip shape to be a single frame. This 1 frame becomes the basis for the proceeding generated video, allowing a user to get more video length out of multiple clips, which at this time OpenAI limits 480p videos to 10 seconds apiece. 

I dragged the frame to the start of the timeline, and left the prompt window blank. SORA will still generate a result without textual input, and I wanted to see what kind moving falling debris would SORA would produce. I was pleased to see that it followed through with clean enough motion from the previous clip sample.



Step 2, Re-Cut, Blank Text Prompt

From the 2nd round clip, I generated another Re-Cut. The 3rd round gen result was the building standing undisturbed. I then continued using the Re-Cut method several times. Repeating the process, I began inputing text of destructive details such as smoke, water, boulders, lava, and fire to come out from the building. Over some time the sky becomes streaked with colors in stead of realistic clouds and the building still stands with some fuzzy details. 

Since the building could not fall down, I decided to use rocks to encase and cover the building up. I still played along with the narrative of mass destruction happening from within the structure but also sought to transform the building into something else as way to collapse it. Over time the building disappears as rocks creep in, eventually forming a rock tower. Then the tower gradually turns into a pile of rocks. Then a prompt to explode the rocks was entered and visually occured. Only until the anthropomorphized constructed building was replaced by unpoliticized natural materials, could the building be destroyed incidentally.

While generating clips, I also imported the downloaded videos to my computer and assembled a complete video timeline of all the clips in Adobe Premiere, with some filter effects applied and audio sourced from online. Each clip was trimmed and fitted to be seamless. The prompting and editing came to an end with rocks gone and a distant flame erupted in a black void.


Conclusion



Rightfully, AI are programmed to be limited photorealistically. Destruction of concieveably real places, people, and things would generate more disinformation with fake murders and mass destruction. There are plenty of video recordings that document high rise buildings being demolished by inwardly toppling, and collapsing level by level.

If SORA had actually created an endlessly collapsing building, it would be easily recognizable as fantasy. One guess I have, is that perhaps the time limits for generating a video clip would result in showing a building just being dynamited. The video would appear normal and believable, which is unacceptable because such video content is useful for a fake news story of an apartment or office building killing many people.

What we see in this compilation of adjusted and enlongated video sequences are the rules that limit AI imagery. Erroneous details like bad movement and disappearing objects create unbelievable things which inidicate one kind of limitations SORA has at this time. Some of those limitations could be because of copywrite or training data that is lacking. Cleary anthropic symbols of human society can be inhibitors for making convincingly real looking content . Apparently only unoffensive raw matter and objects can be damaged, while human inventions are rendered invincible to hellish phenomena. Humans, animals, and civilization are indestructible until they are transformed through prompted instructions. It seems that AI have been instructed by developers to cherish and protect certain subjects which in turn shows just how unreal it is despite it’s rich access to image and video data. 

To just be sure if this is the case, I’ve attached a single video generated on the day of publishing this AI video project.




Generated 8/27/2025 SORA by OpenAI