Fair Use Boundaries for Reference Footage in AI Video Projects

Reference footage in an AI video project is a third, separate legal layer, distinct from training data lawsuits and from who owns the finished clip, one most commercial producers never examine on its own terms.

Why reference footage occupies a distinct legal layer

Most creators run three different copyright questions together as if they were one problem. The first is who owns the AI output once it's made. The second is whether that output infringes someone else's work. The third is whether the reference material fed into the workflow, the images and clips used to guide the generation, was ever cleared to be used that way. Each of these questions has its own rules, its own case law, and its own exposure. Treating them as a single question is where commercial risk actually gets in.

The loudest public fight right now is about training data: studios and stock footage providers suing the companies that build AI models over what went into the training set. If a model is later found to have trained on footage it never had the right to use, commercial users of that model can inherit exposure they didn't sign up for.

Ownership, meanwhile, is largely settled. Perlmutter, leaving the lower courts' ruling in place: an AI system cannot hold copyright, and copyright protection only attaches to work with real human authorship behind it. That closes one open question.

What's left in between is the reference layer: the image a producer uploads to hold a character's face, the clip used to teach a model a camera move, the still frame locked in before a video model generates motion. This layer gets the least attention of the three, and it's also the one a producer has full control over before a single frame renders. It's the layer a producer must get right, because it's the one nobody else can clear on the producer's behalf.

The four-factor fair use test and reference footage

Fair use doesn't resolve with one clean rule. It runs on four factors, weighed together: the purpose and character of the use, the nature of the copyrighted work being referenced, the amount of that work used, and the effect on the market for the original. Those four factors behave differently when footage is used as a reference to condition an AI generation than when it's cut directly into a traditional edit, and understanding that difference matters before a producer ever opens a generation tool.

Purpose and character, the first factor, turns on how different the final output is from any single input example. The four factors as a whole form a lens for judging how exposed a given use actually is, and the answer changes case by case.

Market effect, the fourth factor, is carrying the most weight in current litigation. The U.S. Copyright Office's Part 3 report found that AI use crosses out of fair use territory specifically where the generated content competes with the original work in the market. Applied to reference footage, the clearest infringement scenario centers on whether the resulting clip ends up substituting for the original in the marketplace.

The amount-used factor also works differently here than it does for sampling or straight reproduction. But courts don't stop at that distinction. They look at how much the output ends up resembling the source, regardless of how the input was technically handled. The style they want to borrow is often the style that's hardest to borrow safely.

What "transformative" means for AI models conditioned on reference footage

A finished AI clip can look nothing like the reference footage that shaped it and still fail the transformation test. That's the part creators get backwards most often. Transformation is measured by what the output does to the market for the original work, not by how different the output looks from the input. A wildly novel-looking image can still fail that test if it does the thing the original was made to do.

Reference-conditioning is now standard professional practice, not a workaround. The common path is image-to-video: lock a first frame using an image model, then hand that frame to a video model and instruct it to generate motion from there. Veo 3.1 accepts up to four reference images to hold a person, an object, or a style across a generation. Kling's Elements feature lets a producer register a character from multiple angle shots, or in the Omni model, from a few seconds of reference video, and then call that character by name in later prompts. These are the tools of the trade now, built specifically around conditioning on reference material.

The legal exposure enters right at the upload step, before any generation happens. Whatever image or clip gets uploaded carries its own rights status into the workflow. The AI process doesn't wash that away. A copyrighted still image uploaded as a style reference is still a copyrighted still image once it's inside the system, no matter how many generations separate it from the final output.

The closest appellate guidance on this sits in a ruling that didn't even involve video. In September 2026, the Third Circuit decided Thomson Reuters v. Ross Intelligence, the first federal appellate decision addressing whether using copyrighted material to build an AI system counts as fair use. The case involved legal headnotes, not footage, but the reasoning is the most direct precedent available for anyone trying to figure out where reference-conditioning stands. The lesson travels: using someone else's creative work to build or guide a new tool, even indirectly, doesn't get a free pass just because the end product looks different.

The drift problem with likeness and style references

Almost nobody sets out to infringe a copyright in these workflows. The actual failure mode is slower and harder to notice: iterative drift. A reference goes into the process early, the producer runs generation after generation, and by the final cut, traces of that original reference are still sitting in the output, even though the producer has long since stopped thinking of the project as derived from anything.

Likeness drift appears most often with faces. By the end, the output still carries enough of that person's actual facial structure to be recognizable. That's a right of publicity problem, not just a copyright one: the right to control commercial use of someone's name, image, likeness, and voice is governed by state law, and a growing number of states have extended that protection to cover AI-generated depictions specifically.

Style drift works the same way but through a different channel. Courts look at whether those stylistic fingerprints can be traced back to an identifiable source, and if they can, frame-level originality stops being the question.

The clean fix for likeness drift is simple: don't use real-person references for character work. That removes the judgment call entirely, and removing the judgment call is the entire point. If a project genuinely requires a real likeness, written consent is the floor, not a nice-to-have. Voice carries its own version of the same risk. General-purpose synthetic voices, offered under published commercial terms, sit on much safer ground.

Adapted source material raises a distinct version of drift: chain-of-title failure. Adaptation rights, distribution rights, and audiovisual rights frequently sit with different parties entirely, and a signed contract with the wrong party is an expensive piece of paper.

Platform indemnification claims for reference inputs

Getting a green light from an AI platform to generate something is not the same as being protected once that content goes out into the world. IP indemnification offers from these platforms are written to exclude exactly the thing most producers assume they cover: infringement that comes from reference material the user supplied.

A platform can permit commercial use of its outputs and still leave a brand wide open to a copyright claim, a right-of-publicity claim, a disclosure violation, or content that quietly erodes trust with an audience. The license to generate is a starting point. It is not a guarantee of anything downstream.

Concrete examples make this easier to see. The MPA and ByteDance signed a Memorandum of Understanding on August 17, 2026, setting up IP safeguard frameworks for the Seedance and Seedream models across TikTok, the TikTok USDS Joint Venture, CapCut, and Dreamina. The protection moved upstream, into the model and platform layer. It did not move downstream into indemnity for the user's own inputs.

Litigation settlements carry the same narrowness. The final approval of the Bartz v. Anthropic settlement in July 2026 does not authorize any future training use and does not resolve claims tied to model outputs or other copyrighted works beyond what that specific case covered. Each settlement closes one dispute. None of them function as a blanket shield for everyone using the underlying technology.

California AB 2013, which took effect in January 2026, requires AI developers to publish information about the datasets used to train their generative AI systems, including whether copyrighted material is in there and whether it was purchased or licensed. That gives commercial users, for the first time, a concrete basis for doing due diligence on model provenance before building a campaign on top of a given tool.

Disclosure obligations that layer on top of copyright

Disclosure law and copyright law are two separate systems, running two separate tests, and clearing one has no bearing on the other. A fully disclosed AI-generated video can still infringe someone's copyright. A video built from entirely cleared, properly licensed footage can still violate a disclosure law if it fails to label what it is. Neither one covers the other.

New York's Synthetic Performer Disclosure Law takes effect June 9, 2026, and requires advertisers to conspicuously label AI-generated synthetic performers, meaning non-identifiable fabricated human likenesses appearing in ads. A campaign running nationally almost certainly touches New York audiences through normal ad delivery, which pulls it into scope.

The EU AI Act adds its own transparency requirements on top of that, creating disclosure obligations for AI-generated or manipulated content distributed to audiences inside the EU. A U.S. production with any EU distribution or EU audience exposure needs to check this independently of whatever copyright clearance work has already been done.

The practical consequence for reference footage work is straightforward: a producer who does everything right on the copyright side, clears every reference image, documents every license, can still release a synthetic human likeness without the required label and land in violation of a completely different law. Clearing copyright does nothing to satisfy disclosure requirements. Both need separate attention, every time.

A repeatable clearance process for reference footage in commercial AI video projects

What actually keeps a commercial AI video project safe is a review process run on every project, not a one-time check of a platform's terms of service or a general sense that the tool is "cleared."

The first track is the underlying source material. A signed contract with the wrong party doesn't protect anyone.

The second track is likeness and voice. The default should be no real-person references for character work. Where a real likeness is genuinely required, written consent is the floor, not a formality. For any synthetic voice used commercially, save the document showing the commercial terms at the exact time the voice was used.

The third track covers the assets sitting inside the workflow itself: music, stock footage, fonts, and any reference clips uploaded for conditioning. Music, stock footage, fonts, and uploaded reference material all need separate review from a producer and from legal counsel, asset by asset.

The fourth track is record-keeping, and it fails most often in practice because nobody wrote any of it down. A single folder per project holding the license text, the date it was obtained, and the source URL closes this gap permanently. It costs almost nothing to maintain and solves the problem that causes the most actual disputes.

Model provenance belongs in this process too. California AB 2013's dataset disclosure requirement gives producers a legitimate basis for asking a model provider directly whether its training data included licensed copyrighted material. Providers who answer that question clearly are worth building a workflow around.

Alongside all four tracks, keep records of human authorship on every AI-assisted project: prompt development, selection criteria used to pick among generations, editorial decisions, and any post-production changes. That documentation supports a copyright claim on the finished work and demonstrates the human authorship courts require for protection to attach.

Disclosure compliance runs in parallel with all of this, not after it. Platform AI labeling requirements, New York's Synthetic Performer Disclosure Law, and the EU's labeling obligations all need to be checked against the finished asset that's actually going out the door, not against the list of references that were cleared earlier in the process.

An integrated production environment, one bringing models like Veo, Kling, Sora, and Flux together with a full audio stack such as ElevenLabs' in a single interface, speeds up the generation and asset-management side of this work considerably. It does not replace the clearance discipline itself. The tracks above still need to run on every project, by a person, every time. What changes is that a producer stops having to chase licenses and reference files across five disconnected tools and can run the same checks from one place.

Keep the folder. Date the license. Save the URL. That single habit, more than any platform feature or legal argument, is what closes the gap that actually gets commercial AI video projects into trouble.