The advert got cheap. The disclosure did not

Somewhere below the level at which most boards look, a line item is being repriced. Media spend remains stubbornly expensive and reasonably well understood. Production does not. The cost of making thirty seconds of moving image is falling quickly.

For as long as anyone in a European marketing department has been budgeting, that cost has behaved like a physical cost, because it was one. A crew, a location, a day, a grade, a delivery. It could be reduced by using cheaper people or fewer days, but it could not be reduced by an order of magnitude, and it certainly could not be reduced to the price of a modest lunch.

The board-level question is no longer whether the footage looks convincing. It is what the company must disclose when AI-generated advertising becomes cheap enough for routine use.

What is actually on offer

The system drawing much of the current interest is Seedance 2.5, which ByteDance unveiled on 23 June 2026 at its Volcano Engine FORCE conference. It remains in enterprise beta. A browser platform branded Seedance 2.5 currently provides access to the available 2.0 family and lists the newer model as coming soon.

There is no public Seedance 2.5 specification sheet or API rate card. The launch presentation centred on a continuous thirty-second shot and a larger reference budget, reported as up to fifty inputs across images, video, audio, 3D models and style material. Those are beta claims, not procurement terms. Native 4K and jointly generated audio were already part of the 2.0 generation.

ByteDance reports roughly twenty per cent better adherence to written instructions than its predecessor. That figure originates with the vendor and has not been independently benchmarked, which is the appropriate weight to give it.

The browser service’s commercial model is consumption based. These prices apply to the models it offers today, not to the unreleased 2.5 beta. One-off packs begin at $12.99 for 160 credits and remain valid for forty-five days. A monthly plan starts at $29.99 for 550 credits. Subscription credits expire thirty days after each grant; sign-up and check-in credits expire in seven. Refunded credits from a failed render do not expire.

Two things follow from that structure. Firstly, longer and higher-resolution output consumes materially more, so nobody is producing a finished thirty-second piece at 4K on free sign-up credits, whatever the marketing implies. Secondly, this is consumption spend rather than a retainer, which means it lands in a different budget line, with a different approval path and a different failure mode. Unused credits with a validity window are simply money that evaporates when a campaign slips a month.

The capability curve is steeper than the planning cycle

The reason this arrives as a surprise is the slope. Seedance 2.0 already produces clips of up to fifteen seconds with synchronised audio and a substantial reference set. Its official limit is nine images, three video clips and three audio clips, plus the written instruction. That is enough for campaign work, even though the shorter window still favours cutaways and compact product shots.

If the 2.5 beta’s thirty-second window and larger reference budget survive public release, the change will have happened within one budget cycle. A media plan written last year may therefore contain a production-cost assumption that is already stale.

Where the liability actually sits

Here is the part that European boards are getting wrong, and it is a matter of who the regulation names.

Article 50(2) of the EU AI Act places an obligation on providers of systems generating synthetic audio, image, video or text content to ensure that outputs are, in the Act’s words, “marked in a machine-readable format and detectable as artificially generated or manipulated”. That duty sits with the model provider, subject to feasibility and state-of-the-art qualifiers.

Article 50(4) is the one that matters to the company buying the credits. “Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.” The deployer is the European firm running the campaign. The obligation does not travel with the vendor, and it is not discharged by the vendor complying with its own.

The scope is broader than the phrase suggests, because the Act defines its terms. Article 3(60) defines a deep fake as “AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful”. That is not a definition about politicians and pornography. Objects, places and events are in it. A generated establishing shot of a plausible city street, or of a product on a plausible shelf, sits closer to that definition than most marketing directors assume.

There is relief for creative work. Where the content forms part of an evidently artistic, creative, satirical or fictional work, the transparency obligation is limited to disclosure in a manner that does not hamper the display or enjoyment of the work. An obviously stylised brand film is in easier territory than a clip engineered to look like documentary footage of a real place. Which of those a given campaign has commissioned is a question of fact, decided after the fact, by somebody else.

The Act’s timetable is staggered and parts of it have been politically contested, so no responsible summary should assert a specific compliance date. Firms should take their own advice on when, not whether.

The questions procurement should be asking

The technical specification may be the easiest page in the file. Procurement still needs answers to less glamorous questions.

Do the service terms permit the intended commercial use, in writing, and has a copy been retained with the campaign records? A product page may promise watermark-free commercial output, but the terms of service are the document to keep.

Does the absence of a visible watermark say anything about machine-readable marking? No. A clean-looking file tells a deployer nothing about whether provenance metadata is present. Get the answer in writing.

What happens to reference material that gets uploaded? A run accepting fifty inputs invites teams to upload unreleased product photography, customer footage and brand assets. Whether those inputs are retained, and on what terms, is a data question with a contractual answer, not an assumption.

Is there a register of which delivered assets were generated? Without one, a clip drifts into a folder, and eighteen months later somebody presents it as evidence of something. That is the failure mode, and it is administrative rather than creative.

The dependency nobody wants to discuss

The strategic point is uncomfortable and familiar. The frontier of this capability is not European. The model in question belongs to a Chinese company, its nearest competitors belong to American ones, and the European position in this layer is that of a customer.

This is the cloud argument and the semiconductor argument again, one level up the stack, and with a twist: the artefact being imported is not compute or silicon but the raw material of a company’s public voice. A European brand’s video output can now be produced end to end by a system it does not control, cannot inspect, and did not train.

Abstention hands the cost advantage to competitors. The more useful response is ordinary supplier discipline: read the contract, record provenance, assign an owner and keep an exit route. A multi-model AI API aggregator can reduce the engineering cost of changing video providers, but it does not transfer the disclosure duty or the need to review each model’s terms.

Production prices can fall by an order of magnitude without making the campaign low risk. A cheap render still carries the same brand name when it is published.