AI generated b-roll for a brand video: the cutaway is where the rules actually bite

The standard a generated cutaway has to meet was set by people who were not thinking about generation. Yale Film Studies defines continuity editing as a system of cutting that relies on matching screen direction, position, and temporal relations from shot to shot. Every term in that is a relationship between two shots rather than a property of one.

Yale treats the cutaway and the cut-in as a single operation and ties both to the scene: an instantaneous shift from a distant framing to a closer view of some portion of the same space, and back. On that definition a cutaway is a closer view of a portion of the space the audience is already standing in, rather than supplementary footage arriving from elsewhere.

The failure has a published name. Yale defines the cheat cut as a cut which purports to show continuous time and space from shot to shot but which actually mismatches the position of figures or objects in the scene.

B-roll itself is used here operationally, from what Yale requires of the cut. The Columbia Film Language Glossary's own alphabetical index, read on 4 September 2026, carries no entry for cutaway, cut-in, insert shot or b-roll.

YouTube requires creators to disclose when they use AI to meaningfully alter or generate photorealistic content. Its published triggers are content that makes a real person appear to say or do something they did not, content that alters footage of a real event or place, and content generating a realistic scene that did not occur.

The sentence that reaches a single cutaway sits underneath. The obligation, YouTube writes, could include content that is fully or partially altered or created using AI tools. Three seconds inside a sixty-second film is inside the policy.

Then the platform gives two worked examples. Must disclose: AI generated extra footage of a real place, like a video of a surfer in Maui for a promotional travel video. Need not disclose: AI generating or extending a backdrop to simulate a moving car. Its test for the second list is that a minor edit is primarily aesthetic, and don't alter the content in a way that could mislead the viewer about what actually happened.

Read the two against each other and the line falls somewhere else. The large visual element, a generated backdrop behind a car, needs no disclosure. The three seconds nobody watches on purpose, an establishing shot of a real place, needs one. The variable is whether the shot makes a claim about a real place, rather than how big the shot is or whether it carries the film.

What each library warrants about the contents of the frame turns on a release. Unless specifically warranted above, Getty writes, it does not grant any right or make any warranty with regard to names, people, trademarks, trade dress, logos, copyrighted audio, designs, works of art or architecture depicted or contained in the content. What sits above is itself conditioned: its royalty-free content will not infringe a trademark or other intellectual property right where a property release has been obtained. Pond5 excludes the same list except for locations for which the item page states it has a property release.

Getty carries a separate clause for the case that matters here. In some jurisdictions, its cultural assets note says, using content depicting artwork, architecture or landscapes of cultural importance may require additional authorisation and fees from the appropriate authorities, and Getty disclaims all liability for failing to obtain them.

Now put a Dubai establishing shot against that list. A skyline is architecture. The three seconds a brand film here opens on is, on both companies' published terms, the category where cover depends on whether a release was obtained.

Google reaches the same problem from the generative side. Its indemnity does not apply where the allegation is based on a trademark-related right as a result of Customer's use of such Generated Output in trade or commerce. A brand film is use in trade or commerce by definition.

Audio inside a stock clip is not licensed with it either. Pond5 publishes that music, dialogue or ambient audio in a clip is incidental only, and clearing it is the buyer's job.

Start with the gap, because it is a finding rather than a hedge. No published evaluation was located that measures whether a generated clip cuts against filmed material in the same sequence. MSVBench, 27 February 2026, says as much of its own field: existing benchmarks remain anchored to single-shot paradigms, lacking the cross-shot metrics required to assess long-form coherence.

Across 20 methods MSVBench reports that current models primarily behave as visual interpolators rather than true world models, and that character consistency degrades across multi-shot narratives. Identity goes at the cut rather than inside the shot, and a second benchmark across 17 models splits the same way.

Physics-IQ, co-authored inside Google DeepMind, normalises its score so real video reaches 100, reports the best model scoring 29.5%, and states that visual realism is uncorrelated with physical understanding. VideoPhy-2 reaches the same place by another route, and names the laws that break: momentum and mass conservation are the most frequently violated at 40 per cent, reflection and buoyancy under 20. Momentum is traffic and crowds. Mass conservation is water, smoke and cloth.

Asked for a locked-off shot, a model cuts. Physics-IQ reports generated videos from Sora frequently including transition cuts despite being instructed not to change the camera perspective, and Physics-IQ Verified rewrote its prompts fifteen months later because many models are still subject to camera drift. The fix reads like a call sheet: a static locked-off single-shot with fixed frame throughout. Shot discipline moves into the prompt.

A generated clip is designed to arrive already saying that it was generated. The C2PA technical specification 2.4 gives, as a worked example, a video generated from a text prompt whose manifest records a c2pa.created action with a digital source type of trainedAlgorithmicMedia.

The part that reaches a cutaway is the ingredient mechanism. The specification requires that when assets are composed together, for example placing an audio clip into a video in a video editing tool, information about any claim from the placed asset is recorded into the new asset, as an ingredient that is componentOf the finished file. On the specification's own terms a generated cutaway is recorded as a component of the film rather than vanishing into a timeline.

One tool publishes that this reaches the export. Adobe says it applies C2PA metadata automatically when content or a project that included a Firefly generated asset is downloaded or exported, and forbids stripping it. YouTube publishes that it may label automatically off C2PA metadata, and that such a label can not be adjusted.

Neither publishes whether that label fires on an ingredient or only on the finished film's own created action, and joining the two would be inference. The answer is cheap: upload the finished cut as private, look for a label, and do it before the campaign date.

The federal position gets one sentence and a link, because what the UAE actually requires of AI in a brand film owns that layer. The National Media Authority's twenty media content standards, in English, name no AI at all, and a search of that page returns zero occurrences of artificial intelligence or AI-generated.

The new part is out-of-home, and it starts with what the regime is not. Dubai's Decree No. (6) of 2020 regulating advertisements defines an Advertising Medium so that it does not include newspapers, magazines, periodicals, websites, radio, television, and cinema theatres. Boards are expressly inside it.

What the decree regulates is what an advertisement says. A permit holder must observe accuracy, objectivity, and honesty in presenting the contents of the Advertisement, and Arabic must be the main language, occupying at least fifty percent (50%) of the Advertising Medium, free from grammatical and spelling mistakes.

It says nothing about how the picture was made. A case-insensitive search of its 22,018 characters of English text returns zero occurrences of artificial intelligence, AI-generated, digital, simulat or virtual. The Out-of-Home Advertising Manual, approved by Dubai Municipality under Administrative Resolution No. (180) of 2025, sets twenty-six advisory content guidelines in Section 2.6, Table 42, and none refers to the provenance of an image.

The place a brand film's cutaways do reach that regime is the billboard cutdown, and what a permit holder answers for there is the advertisement's own content. Arabic delivery is a subject in itself, covered in Arabic video production.

A cutaway is a closer view of a portion of the space the audience is already in, matched for screen direction, position and time, and the failure has had a name for decades. The platform's line runs between altering belief about a real place and dressing the space behind a subject, not between hero and cutaway. Three agreements read on 4 September 2026 condition their cover on the asset not being changed, two of them on the context of use as well, and matching a cutaway into a filmed timeline is that change. What either library warrants inside the frame turns on whether a release was obtained, and a skyline is architecture. No measurement of the join was located. And Dubai's out-of-home regime regulates what an advertisement says, not how its picture was made.

Then the dating, which is why the piece stays useful. OpenAI publishes that the Sora API will be discontinued on September 24, 2026, twenty days after this was researched, and one paper records Sora 2 performing notably worse in April 2026 than in October 2025. A capability claim here has a shelf life the vendors put in months, so the durable part is where to look and what to ask. The question a reader can use tomorrow is whether this shot makes a claim about a real place.