AI B-roll Licensing: Why Generated Cutaways Replace Stock Footage but Not the Contract

What a Stock Licence Actually Bought You

AI B-roll licensing is not one question, it is three. A stock clip arrived with a single licence that answered whether you could use it, whether you could sell it, and who was liable. A generated cutaway decouples those answers into tool terms, copyright law and a disclosure test.

Buying a stock clip was never really a purchase of footage. It was a purchase of a document. That document named the territory, the media, the term and the indemnity, and it attached to a specific asset ID that anyone downstream could look up. When a client asked whether the shot was cleared, the answer was a licence number.

It also bought a chain of custody that nobody had to write down. The library had already cleared the location, the model release and the trademarks visible in frame. A producer never had to ask whether the office in the background was a real leased office, because the library carried that risk. That is the part that quietly disappears when the clip comes out of a generator instead.

What it did not buy was specificity. Runway's own guide puts the trade plainly: a stock library is limited by coverage, while a generated clip is limited only by description. Search for a person working and you get hundreds of clips; search for the shot your script actually describes and nothing comes close. That coverage gap is why generated B-roll is replacing stock on commercial timelines at all, and it is also the point where the paperwork disappears.

Why AI B-roll Licensing Is Three Answers, Not One

Runway states the licensing risk in one line: the risk sits in the gap between what a tool lets you make and what it lets you sell. Some generators grant commercial use on every tier and others restrict it to paid plans, so the terms have to be confirmed for the account tier before client work, and a record kept of which tool produced which clip.

The tool tier decides the first answer, and it is a vendor question rather than a legal one: {{link}} — which generator you picked now decides what you are allowed to sell.

The second answer is copyright, and it is the one teams assume away. The US Copyright Office has been publishing its report on copyright and artificial intelligence in parts; Part 2, released on 29 January 2025, addresses the copyrightability of outputs created using generative AI, and the Office's registration guidance treats human authorship as the prerequisite for protection. A purely generated cutaway therefore arrives with nothing to license onward and no registration to point at.

The third answer is disclosure, and it cannot be bought from anyone. It lives in platform rules and in statute, it is triggered by what the clip shows rather than by how it was made, and no vendor tier can waive it.

The tool tier decides the first answer, and it is a vendor question rather than a legal one: licensable AI video platforms — which generator you picked now decides what you are allowed to sell.

The Disclosure Trigger Is What the Clip Depicts

Runway's guidance is unusually concrete here. On YouTube, disclosure applies to realistic content that could mislead viewers about real events, people or places, while clearly unrealistic footage and minor production edits such as colour correction and upscaling are exempt. The operative example is the useful one: a generated texture cutaway over a voiceover generally sits outside the requirement, while a generated shot presented as documentary footage of a real place does not.

Read that as a test and it becomes an edit-bay routine. For every generated insert, ask what the clip claims to depict. A slow orbit around a matte black speaker on a concrete plinth is decoration and asserts nothing about the world. A handheld tracking shot of a delivery cyclist on a wet city street sits closer to the line, and the same shot presented as coverage of a real delivery operation is over it. The generation method is identical in all three cases; the answer is not.

The depict test belongs in the review, not in the export: {{link}} is where a clip that claims to show a real place gets caught before it ships.

This is also why saying a video used some generated B-roll is not an answerable disclosure statement. Nobody downstream can act on it, because it does not say which clips assert what. A per-clip note recording that a shot is generated, non-depicting and decorative takes seconds at the point of edit and turns a vague assurance into something a reviewer can check.

The depict test belongs in the review, not in the export: a trust-QC gate is where a clip that claims to show a real place gets caught before it ships.

Generated B-Roll Has to Be Matched, Not Just Made

The cost that replaces the stock fee is a matching cost. Runway lists four things to carry from the A-roll into every prompt: colour temperature, described in plain language such as warm tungsten or cool overcast daylight; light direction and hardness, so a hard key from the left does not become soft top light; lens character, meaning shallow or deep focus and wide or long; and camera energy, since a locked-off interview does not cut well against a handheld sprint.

Clip length sits inside the same discipline: one to three seconds for a fast social cutaway, three to six seconds for a mood or explanation shot. Lock the aspect ratio to the timeline before generating rather than cropping afterwards, because a 16:9 clip cropped to 9:16 loses the framing that was prompted for.

Two failure modes do most of the damage. Text and interface elements are the first: signage, phone screens, dashboards and packaging copy come back as convincing-looking gibberish, which is worse than obvious nonsense because it survives a quick review. The second is floaty, weightless motion in clips with a moving camera, which reads as generated even when every object in the frame is correct.

Continuity drift is not a generation bug you can prompt away, and it lands in the same bucket as {{link}} — three clips from one prompt return three different kitchens.

Reference images reduce that drift without eliminating it, so separate consecutive generated shots with A-roll wherever the edit allows.

Continuity drift is not a generation bug you can prompt away, and it lands in the same bucket as post-production repair for AI video — three clips from one prompt return three different kitchens.

The Per-Clip Origin Record Replaces the Licence

If the licence is gone, the thing that replaces it is a record. C2PA describes Content Credentials as an open technical standard for publishers, creators and consumers to establish the origin and edits of digital content — a nutrition label for digital content, in the coalition's own framing, that gives a peek at the content's history and can be accessed by anyone at any time.

The value for B-roll is specific. A stock invoice proved that a named clip was cleared for a named use. A provenance record proves that a named clip came from a named tool on a named date and carries the edit history with it, which is the question a client asks six months later, when the master has to be re-cut or re-disclosed.

Statute points the same way. Under the EU AI Act, providers of generative AI have to ensure that AI-generated content is identifiable, and certain generated content — deep fakes, and text published to inform the public on matters of public interest — should be clearly and visibly labelled. The transparency rules come into effect in August 2026, which places the obligation on the current production calendar rather than a distant one.

Regulated buyers already ask for this in writing: {{link}} turns the origin record into something a legal reviewer can sign off.

Regulated buyers already ask for this in writing: regulated-industry compliance turns the origin record into something a legal reviewer can sign off.

What to Keep on the Stock Subscription

The substitution is not total, and pretending otherwise is where budgets go wrong. Runway's own comparison sets out three routes rather than two. Generating gives high specificity at moderate continuity risk, with licensing answered by the platform's usage terms, and suits shots that do not exist in stock. AI-selected stock returns clips in seconds at low specificity and still earns its place on high-volume short-form. Shooting it yourself remains the only answer for anything that must be verifiably real, such as the property you are selling or the machine you own.

Sort the B-roll list into those three piles before a single prompt is written and the licensing question mostly answers itself. What is left is the middle: clips specific enough to generate and realistic enough to depict something. Those are the ones that need the note, and they are usually a minority of a timeline.

The economics only improve where the clip lands first time, which is why {{link}} is the number to watch rather than the price per generation.

Keep the subscription for high-volume generic coverage. Generate the shots that do not exist. Shoot the ones that have to be real. Write down which is which, per clip, at the point of edit, and keep the tool, the tier and the date beside it. That record is the licence.

The economics only improve where the clip lands first time, which is why cost per usable clip is the number to watch rather than the price per generation.