A self-taught experimenter spent several weeks testing whether generative AI could produce a full video project from scratch, following claims that "vibe coding" an entire movie was now possible. The result was a two-minute satirical news program that required significant manual scripting, creative direction, and editing - reinforcing both the expanded capabilities and persistent limits of commercially available tools for creative professionals.
The experiment started from a position of deep skepticism. Image generators had seemed like toys, useful for amusement but incapable of consistent character design or close adherence to a director's vision. The output was, at best, "a springboard for further ideation." That changed after hearing a proponent claim that prompt-driven movie creation had arrived.
"I thought that was likely BS but I decided to put it to the test," the creator said. What followed was weeks of planning, test renders, and manual editing across multiple platforms.
What the tools actually do well - and where they fail
Generative AI, shorthand for large language models trained on datasets that produce output based on predictive analytics, has scaled dramatically in the past two years. The core mechanism, however, has not changed. These systems go "step by step, analyzing the context of what has been supplied and deciding what should follow based on the highest statistical probability," the tester said. "This is exactly how autocomplete works in your word processor. Gen AI is just that on steroids."
The verdict on current tools was blunt, captured in a two-part assessment: "the tools that are commercially available are far better than they were a couple of years ago" - followed immediately by "they still suck." The central finding is that generative AI excels at rendering but cannot handle creativity or problem solving. Those responsibilities remain with the human operator.
A real workflow for AI-assisted video
The workflow began with Gemini, Google's AI product, which was used to identify suitable tools based on project objectives. That led to Runway, a platform housing multiple apps for generative video projects. Runway's agent built characters, settings, and animation tests but had a tendency to improvise and deviate from the script. The reason appears tied to its clip-length restrictions of 15 or 30 seconds - if the script undershoots, the agent pads the runtime. If it runs long, dialogue gets cut or sped up.
For a news-style program with talking-head composition, the creator turned to Hey Gen, which lacks the clip-length constraint and offers granular control over voice performance. Visuals were built from screenshots of Runway-generated scenes, closing the loop between the two tools. All scripting was done manually. "100% of the script was from me, so any jokes that don't land are my fault," the creator said. Final assembly happened in DaVinci Resolve, with royalty-free music sourced from YouTube.
The finished product - a satirical children's-news-style segment running two minutes - took several weeks of work. The creator, who is not an animator, noted that all rendered and graphic elements came from the apps based on their direction.
The skepticism that didn't go away
Despite the results, broader concerns remained intact. The creator acknowledged that everything generated by these models is "built by scraping the IP of others without any compensation or attribution." Environmental impacts from processing demands and resource management issues were also flagged as unresolved, with the creator calling their own minor usage "a statistically insignificant amount" but conceding "maybe that's just a cop out."
The worry about corporate misuse also persists. AI remains "a bad substitute for human productivity," but that has not stopped executives from pursuing workforce reductions. The creative bottleneck that automation was supposed to remove stays in place because human oversight is still required to catch mistakes and hallucinations.
Why this matters for creatives
For those working in creative fields, the signal from this experiment is practical and specific: current generative AI tools can handle rendering tasks under close direction, but they are not creative partners. The human operator must plan, script, test, correct, and edit - the tools accelerate execution without replacing judgment. For anyone exploring AI for creatives, the workflow described here suggests that success depends on treating these systems as powerful but erratic rendering engines, not as collaborators with taste or narrative sense. The gap between what the tools promise and what they deliver is real, and bridging it still takes weeks of hands-on work.
Your membership also unlocks:
