TikTok doubled the length. Europe started counting the labels
Two things happened in the same week, and only a fox with its nose down would have caught both. TikTok doubled how much video its generator will make for you, and Europe's marking and disclosure duties began to apply. Only one of them arrived with a press release.
By Katie Delaney · 2026-08-11 · 11 min read
Thirty seconds changes what AI generated ads can be

On 3 August 2026 TikTok published an upgrade to Symphony, its creative suite, moving it onto Dreamina Seedance 2.5. Two numbers moved. In TikTok's own words, Symphony "now supports up to 30 seconds of video generation, up from a 15 second maximum before", and advertisers "can now upload up to 50 multi-modal references (image, video, and audio), up from 9 previously".
Generated video ceiling
Up from 15 seconds
Multi-modal references per generation
Up from 9
Reference uplift
9 to 50 inputs
Take the length first. Fifteen seconds is a variant. It is the format you generate forty of, run against each other, and discard. Thirty seconds is a spot. It carries a beginning, a turn and an end, and crucially it carries a disclosure without eating the idea. That is the first quietly important thing about ai generated ads this month: the format grew large enough to hold its own compliance. PPC Land covered the upgrade the same week, and the length is the part everyone quoted.
The reference jump is the bigger shift, and it is the scent worth following. Nine inputs is a mood board. Fifty is a brand book: product shots, a voice sample, three prior spots, the packaging, the mandated end frame. An ai video ads generator fed fifty controlled references is doing something different from one fed a text prompt, and the output stops being a lucky draw. It becomes an interpolation of assets you own, which is a far easier thing to defend to a legal team.
Europe started counting labels on 2 August
The same week, with no announcement aimed at advertisers at all, the EU's transparency obligations for generated content began to apply. There was the Commission's own announcement on 2 August, framed for citizens rather than for anyone buying media, which is why it left no tracks in the trade press. The European Commission's own guidance states that "Article 50 of the AI Act applies as from 2 August 2026".
The structure matters more than the date, because it splits the duty in two and puts one half on you. Miss that and the whole trail goes cold. Article 50(2) sits on providers: systems generating synthetic audio, image, video or text "shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated". Article 50(4) sits on deployers: anyone deploying a system that generates or manipulates content constituting a deep fake "shall disclose that the content has been artificially generated or manipulated".
| Duty | Who carries it | What it requires | Where it lands in practice |
|---|---|---|---|
| Machine-readable marking | Provider of the generating system | Outputs marked and detectable as artificially generated | TikTok and its model partner |
| Human-visible disclosure | Deployer of the system | Disclose that content was artificially generated or manipulated | The advertiser and its agency |
| Timing of disclosure | Deployer | Disclosed to a person on first exposure at the latest | The first frame, not the caption |
The Commission's transparency FAQ is specific about timing: deployers must disclose deepfake content to a natural person "upon first exposure at the latest", in "a clear and distinguishable manner". Read against a 30-second generated spot, that is a first-frame requirement, not a caption requirement. Read against the Article 50 text itself, the two duties are separate and neither discharges the other.
Honest caveat, because it matters: the Commission's FAQ carries no advertising-specific guidance, and the Article 50 page notes that amendments are pending and not reflected in the displayed text. This is a live area, not a settled one. What is settled is that a duty now applies to deployers, and most advertisers running ai generated ads have not worked out who in their organisation the deployer actually is.
TikTok's own labelling rule is three years older
None of this is new to TikTok, and ai generated ads have been inside its rules for years. Its policy predates the European duty by nearly three years. In a newsroom post of September 2023: "The policy requires people to label AI-generated content that contains realistic images, audio or video, in order to help viewers contextualize the video." It applies to content completely generated or significantly edited by AI, and permits a label, a sticker or a caption.
So a brand running ai video ads on TikTok has been inside a labelling regime since 2023, whether or not anyone in the den noticed. What has changed is the enforcer and the standard. A platform policy is enforced by moderation and the sanction is your content. A regulation is enforced by an authority and the sanction is your company. Those are not the same risk and they should not share a checklist.
Machine-readable is a provenance problem
The marking half of ai generated ads is where advertising quietly meets infrastructure. Content credentials, the C2PA standard, are the practical route to machine-readable marking, and they are already being read in the wild. Cloudflare's WebMCP preview ships a tool that sweeps a page's images for content credentials and returns a count, in its own worked example finding credentials on 8 of 12 images, and reporting them without cryptographic verification.
That last clause is the one to sit with. Provenance is being surfaced before it is being verified. For anyone shipping ai advertising creative that means the mark on your asset will be read, summarised and possibly shown to a customer by software you do not control, some time before anyone agrees what an unverified credential means. Keeping your own record of what was generated, from which references, and when, is cheap now and expensive to reconstruct later.
The uncomfortable price of the label
Now the part nobody selling an ai video ads generator puts on the landing page, and the reason ai generated ads are an editorial question rather than a procurement one. Labelling appears to cost you performance.
An experiment published in Equilibrium used a two-by-two design, crossing the actual source of an advertisement (AI or human) with how it was labelled (as AI or as human), across a low-involvement product and a high-involvement one. Its finding, verbatim: "Advertisements labelled as human-generated received higher consumer trust and purchase intent ratings, even if they were AI-generated. Conversely, transparently labelled AI-generated ads caused lower trust and intent, particularly for high-involvement products."
Read that carefully, because it is a trap dressed as a finding, a snare set in soft undergrowth. It does not say lie. It says the label, not the machine, moves the needle, and it moves it downward. That is precisely the situation in which a regulation is doing its job and a growth target is doing its own, and someone senior has to decide which one wins. In regulated categories the answer is already written down for you, which is a mercy.
Advertisements labelled as human-generated received higher consumer trust and purchase intent ratings, even if they were AI-generated.
The wider disclosure literature offers no comfort either. A scoping systematic review in PLOS ONE, covering 32 studies from 1997 to 2021, concluded that "the evidence is mixed as to whether features of interactive advertising increase consumer engagement, recall, awareness, or comprehension of product claims and risk disclosures". For ai generated ads, presence is measurable. Comprehension is not, and assuming the first delivers the second is how disclosure programmes quietly fail.
The prominence question is settled ground in the UK at least. CAP's advice on qualifications holds that "the more important the qualifier, the greater prominence is likely to be necessary", and that material information tucked into terms or a scroll away is unlikely to be sufficient. Apply that logic to a label in a caption under a 30-second generated spot and you can see where this ends.
Five rules for shipping generated creative
Written for a marketing team, not a legal one. None of them slows a campaign by more than an hour, and together they keep ai generated ads out of trouble.
Decide, in writing, whether the advertiser or the agency is the deployer for the disclosure duty. It is one line in a contract and it is the question an authority will ask first.
The Commission expects disclosure on first exposure at the latest. In a 30-second spot that means the opening frame or a persistent on-screen mark, not a caption a scrolling viewer never reads.
For each asset, record the tool, the date, the references supplied and who approved it. Fifty references make this more important, not less, because provenance now has fifty roots.
The research suggests a visible label costs trust, especially on considered purchases. That is a reason to test creative that earns trust back, not a reason to hide the label.
TikTok's labelling policy and Article 50 are different obligations with different sanctions. Meeting one does not evidence the other, and your records should show both.
Generate, ship, label if reminded
The tool produces variants, the best performer runs, and disclosure is handled by whichever platform toggle someone remembered to switch on during upload.
Generate, record, disclose on first exposure
Provenance captured at creation, a named deployer, a disclosure inside the creative rather than beside it, and a file that can answer a question a year later.
Where this bites hardest
Regulated categories, predictably, are where ai generated ads bite hardest. A gambling operator, a lender or a clinic already carries mandated wording, and now carries a second mandated mark on top of it. Thirty seconds is genuinely helpful there: it is the first generated format with room for both. Our paid social work and our content marketing practice spend most of their time on exactly this collision, and our brand strategy work is where the question of whether to generate at all gets answered.
Three folkfox pieces sit next to this one: AI creative in paid social, AI generated advertising and the EU AI Act, and the Garante's deepfake labelling ruling, which is what enforcement of this actually looks like when it arrives.
The pattern is the ordinary one. A tool gets better, a duty starts applying, and the gap between them is where the work lives. That gap is a hedgerow, and the quarry is always in it. Nobody needs to abandon ai generated ads, and nobody should pretend the label is optional. The teams that will regret this quarter are the ones that scaled generation without once writing down what they generated, and then had to reconstruct it from memory, tracking their own paw prints backwards, while somebody official waited.
Frequently asked questions
What changed in TikTok Symphony in August 2026?
TikTok moved Symphony onto Dreamina Seedance 2.5. Generated video length doubled from a 15-second maximum to 30 seconds, and the number of multi-modal references an advertiser can upload rose from 9 to 50. TikTok describes it as rolling out to selected paid advertisers in selected markets across Creative Studio, Automations and the Symphony API.
Do I have to label AI generated ads in the EU?
If you deploy a system that generates or manipulates content constituting a deep fake, Article 50(4) requires you to disclose that the content was artificially generated or manipulated. The European Commission states Article 50 applies from 2 August 2026 and that disclosure should reach a person on first exposure at the latest, in a clear and distinguishable manner.
Cooley's client alert on Article 50 taking effect 2 August 2026 strengthens the EU labelling timeline claim, as detailed in the alert.
Is the machine-readable marking my job or the platform's?
The marking duty sits on the provider of the generating system, not on the advertiser. The separate human-visible disclosure duty sits on the deployer, which in most campaigns is the advertiser or its agency. Meeting one does not discharge the other, and the two are enforced against different parties.
Does labelling hurt performance?
The available experimental evidence suggests it can. A two-by-two study crossing ad source with ad labelling found that transparently labelled AI-generated ads produced lower consumer trust and purchase intent, particularly for high-involvement products. Treat that as a reason to invest in the creative, not as a reason to omit a required disclosure.
What is an ai video ads generator actually good for now?
At 30 seconds with 50 reference inputs it moves from producing disposable variants to producing a usable spot built from assets you control. That makes it far more defensible than a text-prompt generation, because every element traces back to something you own and approved.
Has TikTok always required AI labels?
Since September 2023. Its policy requires people to label AI-generated content containing realistic images, audio or video, covering content completely generated or significantly edited by AI, and it accepts a label, a sticker or a caption. The European duty is newer, stricter about timing, and enforced by authorities rather than by moderation.
AI creative in paid social
What generated creative does to testing volume, and where the gains actually come from.
Read the pieceAI generated advertising and the EU AI Act
The wider obligations landing on generated marketing, beyond the labelling duty.
Read the pieceThe Garante's deepfake labelling ruling
What enforcement of an AI disclosure duty looks like once a regulator actually moves.
Read the pieceA single reading is a rumour, not a measurement
The measurement discipline that applies to everything else AI is now doing to your marketing.
Read the piece
Generating creative in a category that gets read closely?
folkfox builds paid social for regulated brands, where a new creative tool arrives with a platform policy attached and a regulation already running underneath it.

