Article 21 in this series was called "The chatbot ate the camera." OpenAI folded Sora into ChatGPT and the dedicated video generation tool became a feature inside a conversation. A month later, Sora died entirely.
Last week, ByteDance put Seedance 2.0 inside CapCut. The pattern repeats, but the direction reversed. Sora moved into a chatbot. Seedance moved into an editing timeline.
Those are very different places to land.
A better home
CapCut has over 600 million monthly active users. Most of them are making short-form video for TikTok. The app is already where they cut footage, add effects, adjust timing, layer audio. Now it generates clips too.
The editing timeline provides something the chat bubble never could: context. When you generate a clip inside an editor, you know what comes before it and what comes after. You know the aspect ratio because the project set it. You know the duration because the gap on the timeline defined it. You know the audio because it is playing underneath. A chat interface stripped all of that context. You typed words into a vacuum and hoped. The timeline gives the generation a neighborhood.
That is structurally promising. A model generating a four-second clip to bridge two existing shots has more creative information available to it than a model generating a four-second clip from a cold prompt. The timeline is a form of ground truth, not as complete as a game engine's geometry files, but infinitely more than a chat bubble's nothing.
A smaller identity
But the timeline does something else. It makes the generated clip feel ordinary.
On the timeline, an AI-generated clip is one track among maybe forty. It sits next to phone footage, screen recordings, stock clips, text overlays, transitions. It occupies the same rectangle. It is interchangeable with any of them.
When Sora was a standalone app, generating a video was the activity. The prompt was the entire creative interface. You invested in it because it was all you had. When the generator moves into the editing timeline, the prompt becomes a function call. Fill this gap. Replace this B-roll. Generate something that looks like the shot I did not get on location. The prompt shrinks from a creative act into a convenience.
CapCut's announcement specifically highlights that creators can use "a few words" to describe a scene. That is the design direction. Not forty precise words with specified lens and movement and lighting and color science. A few. The interface encourages shorthand because the context is supposed to carry the weight. The neighborhood fills in what the prompt leaves out.
Sometimes it will. The editing timeline gives the model duration, aspect ratio, and neighboring visual context. Those are real inputs. But lighting, composition, color intent, camera behavior, and atmosphere are still the prompt's job, and "a few words" hands every one of those decisions to the model's defaults.
The restrictions are not creative
Seedance 2.0 in CapCut blocks generation from images or videos containing real faces. It blocks unauthorized intellectual property. It adds invisible watermarks to every generated frame.
None of these are aesthetic decisions. They are legal ones. The face restriction exists because ByteDance spent March settling copyright disputes with Disney and major Hollywood studios. The IP blocking exists because characters with lawyers showed up in the training data. The watermarking exists because takedown requests need metadata.
The platform's risk tolerance now shapes what you can generate. A filmmaker who wants to create an AI clip that interacts with a real person's face cannot, because the platform's legal department drew the boundary before the creative department could weigh in. CapCut did not decide face generation produces bad output. CapCut decided face generation produces lawsuits.
This is a different kind of default than the ones this series has documented. Beauty bias, center framing, safe lighting, convergent color palettes. Those are invisible opinions about what your work should look like. Face restrictions and IP blocks are visible prohibitions about what your work is allowed to contain. Different in kind. Arguably more honest, because at least you can see the wall before you walk into it. But the wall is not yours. It belongs to the platform. And the platform built it to protect its balance sheet, not your creative intent.
The geography continues to shift
CapCut launched Seedance in Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand, and Vietnam. Not the United States. Not Europe. Not Japan or South Korea.
The Higgsfield competition data showed India submitting nearly twice as many AI films as the United States. Now the most popular editing tool on the planet is giving generative AI video to Southeast Asian and Latin American creators first. The access barriers keep falling in places the industry did not expect.
But notice who is excluded. The markets with the strongest IP enforcement regimes are the markets where the tool is not yet available. The copyright disputes that paused the global launch determined the geography of the rollout. The legal boundary and the access boundary follow the same line.
The standalone tool had a short life
Sora launched as a standalone app in December 2025 and died in March 2026. Seedance launched on Dreamina, paused globally, and relaunched inside CapCut. Runway is becoming a general-purpose platform hosting competitors' models. Grok Imagine lives inside X's interface. Veo lives inside Google's expanding ecosystem.
The dedicated, standalone AI video generation tool is disappearing. Not because it failed. Because it is being absorbed. Every major model is migrating into a larger product: a chatbot, an editor, a social network, a platform, an ecosystem. This is how technology matures. The feature eats the product. The camera app ate the camera. The phone ate the camera app. CapCut is eating the generation tool.
For filmmakers with structured creative vocabulary, the absorption changes the interface, not the requirement. The same forty-word prompt with specified lens, movement, lighting, and color intent works whether you paste it into a standalone app, a chatbot window, a timeline panel, or a node graph. The model on the other end does not care where the prompt came from. It cares what the prompt says.
But the interfaces care. Chat bubbles encourage four casual words. Timeline panels encourage "a few." Standalone apps, when they existed, at least presented the prompt as the main event. Each absorption makes the prompt field smaller, more casual, more interchangeable with a click. The real estate shrinks. The expectation of brevity grows.
Structured prompting does not depend on the interface being a standalone destination. It depends on the filmmaker treating the prompt as a creative decision rather than a form field. That distinction lives in the person, not the software.
The editor ate the generator. The question is whether you still write the order or just let the kitchen decide.
Bruce Belafonte is an AI filmmaker at Light Owl. He once tried to generate a video inside an editing timeline and briefly confused it with ordering fast food at a white-tablecloth restaurant.