Is AI-Generated Video Copyrighted? Three Questions People Keep Merging Into One

"Is AI-generated video copyrighted?" is three questions in a trench coat, and the reason the debate never resolves is that people answer a different one than the one being asked.

Split them and it gets clear fast.

Question 1 — Who owns it? Depends on how much of it you decided

Jurisdictions differ, and this is genuinely unsettled law. But one principle runs through most of the practice so far: protection attaches to human creative contribution, not to the act of generating.

So two things both called "an AI video" can sit at opposite ends:

• One prompt, raw output, posted as-is. Minimal human authorship. Weakest possible claim.
• Written script, shot list, hundreds of generations reviewed and rejected, sequenced, cut to a rhythm, scored, colour-matched. Every one of those is a human choice. That is what a claim is built from.

What decides whether you own it isn't whether you used AI. It's how much of the finished thing you actually decided. Which — awkwardly — is the same test platforms use to decide whether to distribute it.

That convergence isn't a coincidence, and it's why YouTube's inauthentic content language reads like a copyright test written by a distribution team: templated, no commentary, nobody home.

Question 2 — Does it infringe? Completely independent of Question 1

This is the merge that costs people money: owning a video and that video not infringing anyone are unrelated facts.

A piece you unquestionably authored can still infringe on four separate tracks:

• Underlying work — the plot, characters or structure came from someone's novel, comic or film.
• Likeness — the generated face retains identifiable features of a real person.
• Voice — the synthetic voice is recognisably a specific performer's.
• Assets — music, stock footage or reference images without commercial terms you can produce on demand.

Each runs on its own chain of permissions. Clearing one clears exactly one.

Question 3 — Who's liable? The publisher

Searches for this often phrase it as "individual responsibility or platform responsibility." The workaday answer: you are responsible for what you publish. That's the standing arrangement in platform terms of service across the industry.

Generation tools carry their own, different obligations — compliance, labelling, preventing obviously infringing outputs. Those obligations sit with the tool. They do not migrate onto the question of who answers for the video on your account. When a takedown lands, it lands on your post; when a strike lands, it lands on your channel.

The August 2026 development: guardrails moved upstream

On 17 August 2026, the Motion Picture Association and ByteDance announced a Memorandum of Understanding establishing a framework for intellectual property safeguards in generative AI video and image models — explicitly covering Seedance and Seedream and AI tools across TikTok, CapCut and Dreamina. It follows an MPA cease-and-desist issued that February over earlier model versions.

MPA chief Charles Rivkin: "copyright is a cornerstone of the film and television industry," stressing the need for "meaningful guardrails." ByteDance general counsel John Rogovin: "responsible innovation in AI goes hand in hand with meaningful protections for rightsholders."

How to read it:

• ✅ Reasonable inference: mainstream generation models will get harder to steer into obviously infringing characters and scenes.
• ❌ Not an inference: "so generated content is safe now." The specific technical safeguards weren't disclosed, and no upstream agreement covers the reference image you uploaded or the novel you adapted.
• ❌ Not an inference: "ByteDance admitted infringement." It's a cooperation framework, not a finding.

The four clearance tracks, and how they actually fail

Since infringement is the question that costs money, it's worth walking each track and naming the failure that actually happens rather than the theoretical one.

Underlying work. The failure here is almost never "I knowingly adapted a novel without asking." It's chain-of-title: someone sold you rights they didn't hold, or held only in part. Adaptation, distribution and audiovisual rights can sit with different parties, and a contract with the wrong one is expensive paper. Ask for the chain back to the author, and be suspicious of any counterparty who finds that question unreasonable.

Likeness. The failure is drift. You start with a reference image "just for the vibe," iterate, and the output keeps enough identifiable structure to be recognisable. Nobody makes a decision to infringe; the decision was made twenty generations earlier and never revisited. The clean rule is to not use real-person references at all for character work — it removes the judgment call entirely, which is the point.

Voice. The failure is treating "it just sounds similar" as a defence. If listeners can identify a specific person, similarity is the problem rather than the excuse. This track is also the cheapest to avoid: general-purpose synthetic voices come with published commercial terms you can save a copy of, and they cost roughly the same as the risky option.

Assets. The failure is record-keeping, not licensing. Most people do license their music and stock — they just can't prove it eighteen months later when the file has moved through three editors and two hard drives. A single folder per project containing the licence text, the date, and the URL closes this permanently. If you produce at volume, this is the same discipline that makes reviewing AI content at scale tractable rather than theatrical.

Three of these four failures are administrative, not creative. That should be encouraging: most infringement exposure in AI video is solved by a folder and a habit, not by legal spend.

What to actually do

1. Keep process evidence from day one. Prompt versions, shot lists, project files, selection notes, asset receipts. In a dispute this outperforms almost anything else you could have done.
2. Clear the four tracks separately. Underlying work, likeness, voice, assets. A checklist, not a vibe.
3. Don't confuse labelling with licensing. Disclosing that something is AI-generated satisfies a disclosure rule and nothing else — see AI content disclosure rules by platform.
4. Assume the record is durable. Deleting a post doesn't unwind a claim, and increasingly doesn't unwind the platform's record either.

One artefact resolves all three questions at once. Ownership, infringement and liability all bottom out in the same place: a record of how the thing was made.

Ownership needs it to show human authorship. Infringement needs it to show where every element came from. Liability needs it because you're the one who has to answer. One folder does all three jobs, and it costs nothing but the discipline of not deleting things.

The trap is timing. This evidence can only be collected while you don't need it. Eighteen months later, when a claim arrives, the project file has been overwritten, the licence page has changed, the freelancer who sourced the music has moved on, and the prompt history lives in a tool you no longer subscribe to. Nobody reconstructs that. They just settle.

So there isn't a "start doing this once I'm big enough" option. The content most likely to generate a claim is your earliest, least careful work — which means the habit has to start before the work is worth defending, or it starts too late to defend the work that needs it.

FAQ

Can I register copyright in an AI-generated video?

Registration practice varies by jurisdiction and is evolving, and registration evidences rights rather than creating them. Preserved process evidence is usually the more practical investment — and it's free.

Someone reposted my AI video. Can I do anything?

It depends on whether you can show sufficient human creative contribution and that you created it first. Both reduce to the same thing: your project files. Keep them from the first draft, not from the moment a dispute starts.

Does labelling my video as AI-generated protect me?

No. Disclosure rules and copyright rules are separate regimes with separate tests. You have to satisfy both, and satisfying one has no bearing on the other.

General information only, not legal advice; copyright treatment of AI outputs varies significantly by jurisdiction — consult a qualified lawyer for your situation. MOU details: TikTok Newsroom official announcement, 2026-08-17 (retrieved 2026-08-22).