Who Owns the Output of an Image-to-Video Model Under Copyright Law

The most common wrong assumption about who owns AI-generated video copyright is that you automatically own the output because you typed the prompt and uploaded the source image. That is not how copyright works. Copyright vests in a human author who exercises creative control. A machine that predicts pixel motion from training data is not an author. The US Copyright Office, the UK IPO, and the EU AI Act all start from that premise. They diverge on what counts as enough human input to turn an AI video into a copyrightable work. The answer to who owns the output depends on which jurisdiction you are in, what you actually did to direct the creation, and what the terms of service say about ownership. A contract can give you rights copyright law does not.

What The March 2023 Guidance Requires

The US Copyright Office will not register a work produced entirely by artificial intelligence. That policy, published in the Copyright Registration Guidance of March 2023, states that copyright protects only the fruits of human authorship. The office rejects applications where the applicant lists the AI system as the author or where the human role is limited to providing a prompt. For image-to-video, this means the output of a model is unregistrable if your only contribution was uploading a still image and clicking generate.

How The Courts Have Ruled

The office has refused registration for AI-generated works in multiple decisions, and the federal courts have upheld that position in Thaler v Perlmutter, where Stephen Thaler challenged the refusal to register an AI-generated image and lost. The court held that human authorship is a bedrock requirement of copyright. That precedent applies directly to video output because the same principle governs all media. If you want a copyright registration for an AI video, you must show that a human made creative choices the model merely executed. Selecting a source image and writing a text prompt is not enough. The Copyright Office has said that de minimis human contribution does not satisfy the requirement.

Thaler v Perlmutter is the controlling US precedent on AI-generated works and copyright registration. The case involved an image, but its reasoning covers video because the legal question is identical: can a non-human entity be an author? The answer was no. The court affirmed the Copyright Office's position that a work must be created by a human being. For image-to-video output, you cannot register a clip produced entirely by the model. The case does not bar registration entirely. Use the output as part of a larger work where you exercise creative control. Edit the clip. Add sound. Integrate it into a film. The registration applies to the final work, not to the generated segment alone. The Copyright Office treats the generated segment as unprotectable material. You cannot claim copyright over it independent of your original contributions.

Compilation And Editing

The core question is how much human intervention turns AI output into a copyrightable derivative work. The Copyright Office and courts have not drawn a bright line, but some patterns are clear. Select and arrange multiple generated clips into a sequence with a narrative structure. That likely qualifies as human authorship of the final compilation, even if each clip individually is unregistrable.

Directing The Model

Use region mask support in a model like Pika 2.2 to animate only a selected area of the image while keeping the rest static. Or specify an end frame in Kling 2.5 to control the final composition. Both show creative direction beyond a prompt. The more control you exercise over the output, the stronger your claim to copyright in the resulting work. Prompt authorship alone, typing a description of the desired motion and accepting the first result, is insufficient. The de minimis human contribution standard means trivial inputs do not create copyright. A filmmaker who produces 50 clips, selects 8, edits them together, adds a soundtrack, and performs colour grading has a much stronger case than someone who produces one clip and posts it.

EU AI Act Ownership Silence: What The Law Does Not Cover

The EU AI Act does not address ownership of AI-generated output. Its transparency obligations under Articles 50 and 51 require deployers of AI-generated video to label the content, but the act is silent on who owns the copyright in that content. That silence means the general copyright law of each member state applies, which varies. Most EU member states follow the human authorship requirement rooted in the InfoSoc Directive and the Berne Convention. A generated video with minimal human input is unlikely to be copyrightable in Germany, France, or Italy. The EU AI Act transparency requirements apply regardless of ownership. If you publish an AI-generated video in the EU, you must label it as such, even if you believe you own the copyright.

The single thing that most often goes wrong is treating the terms of service ownership clause as a copyright registration. A broadcaster or platform that requires copyright indemnity will not accept a terms of service document as proof of ownership. They will ask for a copyright registration or a chain of title. If the video is unregistrable, you cannot provide one. Build your workflow around enough human creative control that the final work qualifies as copyrightable, even if the individual generated segments do not. Use the tools the models provide: end frame support in Kling 2.5, region masks in Pika 2.2, camera control in Runway Gen 4. Then add your own editorial and post-production work. The legal question is unsettled. Act as though the output has no copyright until you add something the law recognises as original authorship.