What actually went to air
On 9 August 2026, TV Asahi broadcast a 30-second commercial titled ‘Fun Summer with GREEN DA・KA・RA’ during its live coverage of the Japan–Mongolia men's basketball international and in other terrestrial slots. The advertiser was Suntory Beverage & Food. Planning and production sat with TV Asahi's Content Programming Division AI Creative Studio, a unit established in April, while AI model Inc. was named as the production company.
TV Asahi described it as the broadcaster's first television commercial made with ‘full generative AI’ for the video. The release then supplied the more important qualification: AI was used for the picture, but not for the captions, sound or narration. TV Asahi announcer Rumi Watanabe delivered the voice-over. The resulting object was not one undifferentiated AI file. It was a controlled assembly of layers with different authorship, risk and approval routes.
That distinction is why this is a useful industry case. The commercial was not presented only as a lab test, festival short or social experiment. It entered a conventional national advertising slot for an established consumer brand. Once generated footage reaches that environment, the production question changes from ‘can the model make 30 seconds?’ to ‘can a team identify, approve and warrant every component of the broadcast master?’
The boundary did more work than the label
‘Full-AI commercial’ is effective publicity language, but it is too broad for a production record. TV Asahi's own explanation is more operational because it describes a perimeter. The generated picture layer carried the summer scenarios, the child, the liquid heart and the visual transitions. Text, audio and narration remained outside it. Each of those choices reduces a different category of uncertainty.
Keeping on-screen copy separate allows the legal line, product qualification, brand spelling and timing to be checked as exact text instead of trusted to generated pixels. Keeping the voice separate gives the production a named performer and a conventional recording asset. Keeping sound outside the model allows music, effects, mix levels and broadcast loudness to travel through familiar clearance and finishing processes. None of this makes the picture layer automatically safe. It makes the boundary reviewable.
The practical document is a layer map attached to the brief and final master. It should name picture, product pack, logos, supers, legal copy, voice, music and sound effects; mark each as generated, photographed, designed, performed or licensed; identify its owner; and link the approval evidence. ‘AI used’ then becomes a set of traceable decisions instead of a single ambiguous flag.
A broadcaster became part of the production company
The structural change is as significant as the imagery. TV Asahi did not simply sell airtime to a finished campaign. Its in-house AI Creative Studio planned and produced the commercial with an external specialist, while the network supplied the transmission environment and the narrator. The broadcaster therefore occupied several positions in one delivery chain: creative producer, production partner, media owner and public narrator of the case.
That arrangement follows TV Asahi Holdings' 2026–2029 management plan. AI is one of five stated strategic pillars, with an AI Creative Studio intended to create anime, drama and commercials and to develop new revenue from content and business development. This spot is evidence of a broadcaster attempting to turn production capability into a commercial product, not merely adding an AI effect to an existing programme.
For agencies and production companies, the competitive signal is not that broadcasters will replace every supplier. It is that the boundary between production, media and case-study distribution is compressing. The organisation controlling the slot can also build the asset, explain the process on its own channels and package the experiment as programming. A supplier now has to define the value it adds after access to generation has become widely available: direction, brand control, rights evidence, finishing, versioning and reliable delivery.
The making-of became a second deliverable
TV Asahi announced that the making-of would appear three days later inside its regular late-night programme AI Grand Strategy. That turned the commercial's production process into editorial content. One commission therefore produced at least two public-facing assets: the brand message itself and a programme segment about how the broadcaster made it.
This changes what must be cleared. A production diary, prompt screen, reference board, model interface or rejected take can expose sources, personal data, confidential brand material or unapproved likenesses even when none appears in the master. If the making-of is planned from the start, it needs its own shot list, screen-clearance pass, contributor permissions and explanation of what each visible asset represents.
It also changes the commercial measurement. The airtime delivers reach; the making-of delivers novelty, trade coverage and a narrative about innovation. Those outcomes should be separated. A campaign may attract attention because it is AI-made without proving that the generated approach improved recall, product understanding or sales. The production report should distinguish media performance, brand performance and earned coverage from the method used to make the pixels.
Rights evidence has to match the broadcaster's own position
The public release names the organisations involved but does not identify the generation models, input assets, training-data position, number of artists, production schedule, budget, consent route for depicted people or provenance standard used for delivery. Their absence from a press release is not proof that the work was unrecorded. It does define what an outside reader cannot verify from the public case.
That gap matters because TV Asahi is also a rights holder. Its copyright page says its programmes, images, music and other content contain multiple underlying rights, and it prohibits unlicensed use of its content for generative-AI training where an enjoyment purpose exists. The Japan Commercial Broadcasters Association, chaired by TV Asahi chairman Hiroshi Hayakawa, has separately called for member content not to be trained on without permission and for systems to prevent identical or similar outputs.
A broadcaster commissioning generated video therefore needs the same standard of respect for inputs that it asks model developers to show its outputs. The delivery pack should identify the model and account, applicable commercial terms, supplied references, brand-owned inputs, restrictions on training or retention, likeness and performance consents, similarity review and the person who accepted the residual risk. Provenance is not an ideological add-on here. It is how the buyer keeps its production practice consistent with its rights position.
Brand truth still needs a protected path
The commercial uses fantastical imagery to communicate hydration: the brand's heart motif fills with liquid while summer scenarios show the product in use. Generative video is well suited to transformations that would be expensive or cumbersome to photograph. The risk is that the same system can alter pack geometry, label text, liquid behaviour, a person's body or the physical meaning of the shot between frames.
Separate creative latitude from factual control. The director can allow broad interpretation in environment, transition and texture while locking the approved pack shot, logo proportions, product colour, campaign copy and required qualifications. Every generated person should be reviewed for identity similarity and demographic intent. Every product interaction should be checked at speed and frame by frame. A plausible still is not enough if the bottle, hand or mouth changes during motion.
The brand-safety pass should include product accuracy, prohibited claims, legal text duration, legibility, continuity, anatomically coherent action, unexpected marks or text, competitor similarity, cultural review and broadcast technical compliance. Record the accepted frame range and final overlays rather than approving only a storyboard. Generation moves more of the risk into temporal details that a static key visual will not reveal.
Build the delivery pack before the first generation
Start with a one-page boundary schedule. For every layer, state the intended method, source owner, permitted references, fixed brand elements, human contributor, review owner and final evidence. Add a model register covering provider, version where available, workspace, commercial terms, data-retention setting and whether client material can be used for provider training. Agree this before a vendor uploads a pack shot or treatment deck.
During production, retain the approved inputs, selected outputs, meaningful human edits and rejection reasons. Give each shot a stable ID. Log which output became which version of the master. Run similarity and likeness review before picture lock, then perform the usual online, copy, mix and broadcast checks on the assembled film. Preserve the final master hash, captions, audio stems, clearance summary and the exact public description of AI use.
At hand-off, the client should receive more than a video file. Deliver a layer manifest, asset and rights register, contributor list, approval log, model-use summary, disclosure decision, technical report and exceptions note. If a making-of or award entry will follow, specify which production material is cleared for publication. The goal is not to archive every failed prompt. It is to make every material decision behind the broadcast asset retrievable.
The next benchmark is repeatable delivery
TV Asahi's spot crossed a meaningful threshold because an established broadcaster and national advertiser put generated pictures into ordinary terrestrial distribution. Yet the most reusable choice was conservative: generated imagery was bounded by human-controlled copy, sound and narration. The project treated ‘full AI’ as one layer of a finished commercial, not a reason to abandon the rest of production.
Future cases should be judged on more than spectacle or apparent production savings. Ask how many generated shots passed, how many iterations each approval required, which fixes moved into conventional post, whether product and people remained coherent, how fast rights evidence could be retrieved, what the making-of added, and whether the campaign achieved its commercial objective. Cost per generated second is less informative than cost per approved, cleared and delivered second.
The mature lesson is not that television advertising no longer needs a shoot. It is that a commercial can now arrive through a different image-making route while still needing the same precision about authorship, claims, performance, timing and responsibility. AI changes the picture pipeline. Broadcast delivery still demands somebody who can say exactly what is in the master and why it is safe to air.
Build
Turn the AI boundary into a delivery record
I help production companies, agencies and brands connect generated assets to rights evidence, human approvals, version control and client-ready delivery records.
Launching a business of your own? Founder Launch OS connects the brand, offer, website and visual campaign in one guided Codex or Claude Code workspace.
Keep reading
