AI video chain of title: who owns a generated commercial in 2026

Why the AI video chain of title has three links, not one

An AI video chain of title is the documented path from prompt to published cut that shows which expressive elements a human contributed and which a model produced. In 2026 that path decides three separate questions: whether you may use the footage commercially, whether you can stop a competitor copying it, and who pays if it infringes. Brand teams routinely collapse the three into one, and the gap surfaces the first time a platform, a freelancer and a legal reviewer disagree.

Each link has a different owner. Commercial use is settled by the generator's terms of service, which is a contract. Protection is settled by copyright law. Liability is settled by indemnification clauses and by who ultimately published the ad. A tool can hand you full commercial ownership of an output that copyright law refuses to protect, which leaves you free to run it everywhere and powerless to stop anyone copying it.

Stock footage used to arrive with a tidier paper trail. A generated cut has no equivalent of {{link}}, so the three answers arrive from three different documents that nobody has read together.

That fragmentation is not a paperwork annoyance. It decides what a brand actually owns at the end of a campaign, and it decides what a competitor is allowed to lift.

A generated cut has no equivalent of a single stock licence that answers use, resale and liability at once, so the three answers arrive from three different documents that nobody has read together.

What the Copyright Office settled before the 2026 commercial cycle

The U.S. Copyright Office has been consistent since its March 2023 registration guidance, and its AI report series reinforces the position: copyright does not extend to purely AI-generated material, or to works where a human lacks sufficient control over the expressive elements. Part 1 of that report, published in July 2024, covered digital replicas. Part 2, published in January 2025, addressed the copyrightability of outputs created with generative AI. Neither part carves out a special lane for video advertising.

The prompting question is where most production teams lose the argument. The Office treats a prompt as an instruction rather than an expression, closer to directions handed to a commissioned artist, because the system rather than the user determines the specific expressive elements. Refining a prompt a hundred times does not convert selection into authorship; the Office compares it to re-rolling a roulette wheel and then choosing a number. One wrinkle is worth knowing. The prompt itself may be protectable as a piece of writing, but that protection does not extend to the video it produced.

The bar has held. In March 2026 the U.S. Supreme Court declined to take up the question of whether an AI system alone can create a copyrightable work, and the Office has said existing legal principles are flexible enough to address generative AI without new legislation.

The human contribution ladder, rung by rung

How much of a generated commercial is defensible depends on how much a human shaped the expression. The practical ladder runs from nothing to a normally protectable work. Raw prompt-to-output used as delivered is unprotected. Generating dozens of variations and picking one is still weak, because selection alone is not treated as a creative act. Editing and sequencing generated clips into a timeline protects the arrangement even when the individual clips are not. Mixing generated elements with human-authored material protects the human parts. AI used as a background tool inside a human-directed production is protectable as ordinary creative work.

The Office's own framing adds three specific routes to protection: creative selection and arrangement, significant human modification, and expressive human inputs that remain perceptible in the final output. The register has granted and withheld protection along exactly those lines. In the Zarya of the Dawn registration, the human-written text and the arrangement of the images were protected, while the individual generated images were not.

For a video team the operational translation is unusually concrete. The edit, the grade, the sound pass and the compositing decisions form the protectable layer. The raw clips underneath do not, and a competitor who generates a similar clip is not copying anything the brand owns.

Where the chain snaps: vendor plan tiers and the free-account gap

The ownership link lives entirely in vendor terms, and those terms are not uniform. OpenAI assigns to the customer all its right, title and interest in the output, which supports full commercial use on paid plans. Runway grants commercial rights on every plan, including free. Midjourney assigns ownership only to paid subscribers, which is where the free-tier gap turns from a curiosity into a production risk.

The failure mode is quiet. A freelancer generates a batch of clips on a free account, delivers them as a finished job, and the brand assumes it holds commercial rights. It does not. The tool's contract was with the creator, not with the advertiser, so a broken upstream term leaves downstream ownership resting on nothing. Two rules close most of that hole: get commercial rights confirmed in writing by whoever delivers the asset rather than by the vendor's marketing page, and record which model and which plan tier produced the final file.

Vendor terms also diverge on scope. Free and consumer tiers frequently restrict commercial use outright or bake in a watermark, while paid and API tiers license the output but almost universally disclaim any warranty that it does not infringe. That is why {{link}}, because a platform whose terms assign output rights on the plan you actually pay for removes a link from the chain before anyone starts generating.

Rosters amplify it. Teams that {{link}} multiply the exposure, because every engine added to the roster brings its own rights-transfer language, its own tier restrictions and its own indemnity carve-outs.

That is why rights clarity now decides which generative video platform a team picks, because a platform whose terms assign output rights on the plan you actually pay for removes a link from the chain before anyone starts generating.

Teams that route each job to the model that wins it multiply the exposure, because every engine added to the roster brings its own rights-transfer language, its own tier restrictions and its own indemnity carve-outs.

Indemnification covers less than the invoice suggests

Indemnification is the link brand teams check last and value most, and it is the narrowest. In practice enterprise agreements tend to indemnify against output-similarity claims while explicitly excluding training-data claims and right-of-publicity claims, so the clause that sounds like cover for a generated ad may not touch either of the risks video actually creates.

Availability differs by tier as well. OpenAI extends intellectual-property indemnification to API, business and enterprise customers rather than to consumer plans, while Runway's cleaner commercial terms come with no indemnification at all if an output triggers a copyright claim. Video also adds a track that still images and instrumental music do not: a moving, talking likeness. Tennessee's ELVIS Act extends publicity protection to AI-generated voice and likeness specifically, and several other states are considering similar bills.

Some vendors now require customers to attest that they hold consent for any real person's likeness, and breaching that attestation can void indemnification even where it otherwise exists. Nor does indemnification answer {{link}}, because that exposure belongs to whoever published the ad.

The practical reading is to treat indemnity as a partial discount on risk rather than a transfer of it, and to price the remainder deliberately.

Nor does indemnification answer the consent chain behind a synthetic performer, because that exposure belongs to whoever published the ad.

The documentation spine that makes authorship provable

Because ownership turns on demonstrated human control, the record becomes an asset in its own right. Guidance on human authorship and creative control is blunt about the takeaway for production teams: document the creative process, because the clearer the evidence that human choices in selection, arrangement, modification and expressive input shaped the final work, the stronger the authorship claim.

A working spine for a commercial cut is short and boring, which is the point. Keep the prompt and its version, the model and plan tier behind each approved frame, the reference and keyframe approvals, the edit and grade decisions with timestamps, and the name of the person who signed off. Compliance guidance for AI video already lists prompt logs and generation records as an ongoing obligation, precisely because good-faith use is hard to reconstruct after the fact. Three months later, when a client asks who owns a shot, those records answer in minutes instead of arguments.

This is where the chain holds or fails. Terms assign rights, law decides protection, and the record decides whether either claim survives scrutiny.

Performance creative and hero assets need different decisions

Not every asset deserves the same attention. A fifteen-second performance variant is disposable by design, tested and replaced, and its value sits in the performance data rather than in exclusive ownership of the pixels. Legal effort spent there is largely wasted.

Hero assets are the opposite case: brand films, mascots, signature sequences and anything reused across years. Those are the cuts where climbing the ladder pays, where the human-authored layer is worth registering, and where the vendor tier and indemnity carve-outs deserve a read before the first frame is generated.

In regulated categories the record has a second audience. For finance, health and pharma work {{link}} is the natural home for the prompt log, because a compliance reviewer will ask for the same evidence a copyright question would.

Run the decision in order, because the ordering matters more than the analysis. Confirm the commercial right in writing, decide whether the asset is worth protecting, then settle who pays if the output turns out to be unsafe to run. Most teams examine the second question obsessively and skip the first and the third.

For finance, health and pharma work the compliance layer finance and pharma brands already maintain is the natural home for the prompt log, because a compliance reviewer will ask for the same evidence a copyright question would.