Generative AI breaks the ownership chain at the moment of creation. Copyright has one load-bearing instant — when a work is made and a right attaches to it — and a generative model now sits inside it, able to take a clip it was never licensed to use and restyle, re-cut, or regenerate something close to it in seconds. That breaks the chain at exactly the link an enforcement question depends on. The answer is not to guess where the law will land. It is to separate what is settled from what is contested, and to build the record that holds up under either answer.
Start with the part that is not in doubt. United States copyright protects original works of authorship, and the subject-matter grant in 17 U.S.C. § 102 has always been read to require a human author. The Copyright Office states the rule plainly in its Compendium: it will register a work only if it was created by a human being, and it will refuse material produced by a machine or a mechanical process that operates without creative input or intervention from a human author. That requirement predates generative models by more than a century. What the models did was put a machine in the authoring seat and force the question into the open.
The clearest line in this area is also the oldest. A work needs a human author to be protected, and a court and the Copyright Office have both applied that rule directly to machine output. In Thaler v. Perlmutter, the D.C. Circuit affirmed in 2025 that a work generated autonomously by a machine, with no human author named, cannot be registered under the Copyright Act. The applicant had listed the system itself as the author; the court held the statute requires a human one. Separately, in its Zarya of the Dawn decision, the Office allowed registration of a comic’s human-authored text and the human selection and arrangement of its images, while refusing protection for the individual images a generative tool produced from prompts. Two different bodies, one consistent rule: protection follows human authorship, and it stops where the machine’s unguided output begins.
For a rights holder, the practical reading is narrow and useful. A clip a person filmed, framed, and edited is squarely the kind of human-authored work the system protects. The settled rule is not a threat to that footage. It is a threat to the assumption that whatever a model produces is automatically someone’s property — and that assumption is the one that quietly breaks the ownership chain.
The law here is unsettled, and any flat claim to the contrary is a position, not a holding. Whether training a generative model on copyrighted video is an infringement, and whether any such use qualifies as fair use, is the subject of active litigation across multiple courts, and those questions have not been resolved. The Copyright Office has studied the area in its multi-part Copyright and Artificial Intelligence report — its training report walks through the competing arguments without declaring a winner, because the fair-use analysis turns on facts that differ from model to model and use to use. Neither side is established — that is what makes it contested rather than decided.
The same uncertainty surrounds the output side. When a model regenerates something close to a human-created clip, the questions stack: was a protected work copied in training, is the new output substantially similar to a protected work, and who, if anyone, authored the result. The responsibility of the platforms that host this output and the providers that build the models is being worked out alongside it, and no particular claim against any of them is foreclosed today. Because these questions are being litigated on different facts in different courts, a rights holder cannot bank on a single ruling clearing the field; the only safe assumption is that the answer will turn on the specific record in front of the court.
Between the settled floor and the contested questions sits a genuinely open one: how much human authorship a work needs before protection attaches, and how a mixed human-and-machine work gets divided. The Copyright Office addressed part of this in its copyrightability report, which takes the view that protection depends on the degree of human creative control over the expression, assessed case by case rather than by a single bright line. A creator who substantially arranges, edits, and shapes a sequence built partly with a generative tool is in a very different position from someone who typed a prompt and accepted the first result. Where exactly the line falls between those two is not fixed.
That ambiguity matters for ownership because the right does not attach to the machine output; it attaches to whatever a human actually authored. A creator who can show their own selection, arrangement, framing, and editing decisions has a protectable layer even inside a partly machine-assisted piece. A creator who cannot show that input may hold less than they assume. The chain of title that 17 U.S.C. § 201 governs starts from that first question of authorship, and a murky first link makes every later transfer and license harder to stand on.
The durable question is not abstract ownership. It is what the record shows. That point held in recent secondary-liability precedent we unpack in our rights-holder playbook, and it holds again here. When a generative step sits in the creation moment, an owner cannot resolve the open and contested questions on their own. What they can control completely is the evidence of their own human authorship: the raw capture, the edit decisions, the timeline, the provenance of the source material. Those are the facts that survive whichever way the unsettled questions go.
The reason this works is structural. The settled rule rewards demonstrable human authorship. The open question about degree is answered with evidence of creative control. Even the contested training and output questions, when they reach a court, will be argued against a claimant’s ability to show what they made and how. Across all three layers, the same raw material does the work: a clean, contemporaneous record of human creation and clear title. The owner who has it is prepared for every branch. The owner who assumed the machine output was simply theirs is exposed at the exact link that breaks first.
Documentation does not wait on the law settling. It is the same discipline that makes any ownership claim hold, applied to the place AI puts pressure on it. Three records decide whether a claim survives.
Capture the human authorship. Keep the originals, the project files, and the record of creative decisions — what was filmed, what was selected, what was arranged and edited. This is the evidence that a protectable work exists and that a person, not a machine, authored it. It is also what separates a creator’s real contribution from any machine-assisted layer inside the same piece.
Document provenance over time. A time-stamped, contemporaneous record of where a clip came from and how it moved is worth more than a reconstructed timeline assembled after a dispute starts. Provenance technology is becoming part of how platforms and tools mark machine-involved content; the EU AI Act’s Article 50 requires that AI-generated or manipulated audio, image, and video be marked in a machine-readable format, which pushes a labeling signal into the file itself. Cryptographic content credentials can turn a later dispute from a similarity argument into a record check.
Clear the title before you need it. The open authorship question and the contested training questions both stall on murky title. A claim built on footage whose human authorship and chain of ownership are clear clears that threshold; a claim built on output nobody can show a person authored does not. Knowing exactly what you hold and in what order it matters is the difference between a position that survives scrutiny and one that unravels under it — the record is the asset, not the count of clips.
Generative AI did not rewrite copyright. It applied pressure to the weakest link in the chain — the creation moment — and turned a long-quiet authorship requirement into a daily one. How the contested and open questions resolve is not within any single owner’s control. The one thing that is: the evidence of human authorship and the clarity of title that every branch of this still-developing law will reward. Build that, and the question of who owns what stops being a guess about the future and becomes a fact you can show.
This article is general information about copyright law and public agency guidance, not legal advice. Verights is the rights-enforcement brand of SocialCoaster Inc.; it is not a law firm, and reading it creates no attorney-client relationship. It describes public guidance and decided cases, takes no position on any specific party, platform, dispute, or pending matter, and is not a party to any matter referenced. Several questions in this area are actively contested and unresolved; nothing here is a prediction of any future result. Consult qualified counsel about your situation.
Verights helps rights holders build the evidence of human authorship and clear title that an ownership claim rests on — before a generative step puts it in question.
Talk to the Verights team