AI UGC Ads on YouTube Shorts and Demand Gen: Specs and Disclosure Rules

Quick answer: a YouTube Shorts ad is not a file you attach to a campaign. Google states that "the videos you want to use as the YouTube Shorts ads must be uploaded to YouTube", public or unlisted. That one sentence is why AI UGC on Shorts sits under twodisclosure regimes at once: YouTube's altered-or-synthetic content rule, which governs the upload, and the Google Ads AI label that started rolling out on 9 July 2026, which governs the ad. On specs, build 9:16 at 1080x1920, keep it under 60 seconds, and know that Demand Gen's floor is 5 seconds while anything under 10 seconds is ineligible to serve on YouTube In-stream. That last number rules out an 8-second single clip from some models, including Veo 3.1.

Two rulebooks, one file

Most coverage of AI and YouTube treats the creator disclosure and the advertiser label as the same thing. They are not. They come from different teams, they trigger on different tests, and they land in different places on the page.

The Google Ads page is honest about its own limits in a way vendor summaries usually drop. It says outright that "use of the AI label setting in Google's advertising products doesn't guarantee compliance with specific regulations" and tells you to seek legal guidance. Treat the toggle as plumbing, not as a defence. What the setting does and does not do across the five products it ships in is worth reading once: the July 2026 Google Ads AI labeling policy is a permission to label, not a duty to.

What YouTube actually asks you to disclose

YouTube's disclosure pagegives three triggers. Creators must disclose content that "makes a real person appear to say or do something they didn't do", "alters footage of a real event or place", or "generates a realistic scene that didn't actually occur".

For AI UGC, the third one is the one that catches you. An invented presenter in a kitchen holding your product is not a real person saying something they did not say, and it is not altered footage of a real place. It is a realistic scene that never happened. Disclose it.

The carve-outs are real too. YouTube says creators do not need to disclose "non-realistic content that's made with AI, or edits to realistic content that are minor", and its examples include beauty filters, colour adjustments, cloning your own voice, and "production assistance, like using generative AI tools to create or improve a video outline, script, thumbnail, title, or infographic". So an AI-written script read by a real creator is outside the rule. A fully generated spokesperson reading the same script is inside it. If you want the line drawn for the other two big platforms, our TikTok AI content disclosure rules and Meta AI-generated creative ad policy posts do the same job for TikTok and Meta, and the three are not aligned.

Where the label shows up depends on the content. YouTube says labels appear "in the video player" for photorealistic content, and "may appear in the expanded description" for animated or non-photorealistic AI. Photorealistic UGC gets the more visible treatment. That is the version worth planning your creative around.

The same page says YouTube may apply that label automatically when the file carries C2PA metadata. Do not budget for it. We injected a C2PA manifest into a test clip and ran it through our own export path, and the manifest did not survive even a lossless remux, which is the usual outcome for anything that rewrites the container, and nearly every toolchain does. Plan on declaring the content yourself and treat an automatic label as a bonus.

The spec sheet, from Google's own pages

Demand Gen is where most ecom buyers reach Shorts, because the same campaign serves "Discover Feed, Gmail, YouTube videos, YouTube Search Results, YouTube Watch Next, YouTube Home Feed, YouTube Shorts, Google Display Network websites, and apps". Here is what the asset pages actually require:

Two of those rows quietly conflict with how AI video is generated. The 10-second in-stream floor and the "more than 15 seconds" recommendation both sit above what several models will produce in a single clip.

Where the models land against those numbers

This next table is ours, not Google's. It is read straight off our production pricing table and model registry, which is where our app decides what a render costs and which parameters it may send. Nobody else publishes this because nobody else is dispatching to a dozen models from one codebase.

Read the top row again, because it is the practical trap. Veo 3.1 is the model most people reach for first, and its longest single clip in our catalogue is 8 seconds. That clears Demand Gen's 5-second minimum, so the asset uploads and serves on Shorts and in-feed placements, and then quietly never appears in-stream. If in-stream is part of the plan, either stitch multiple generations into one asset or pick a model that reaches 10 seconds and beyond in a single pass. Our Veo 3.1 model page lists the priced durations, and the cost benchmarks post compares what each of them costs to finish.

The bottom three rows need one clarification, because the number people repeat is a price-table artefact rather than a capability. Our published tiers for OmniHuman and Fabric stop at 30 seconds, but those models have no duration input, so the app does not charge from that table at all. It prices them per spoken second. Where each one stops is a different number again: our registry caps OmniHuman 1.5 at 30 seconds because its schema requires an audio track under 35 seconds, VEED Fabric 1.0 at 120, and Pruna Avatar bills to 60, where that model truncates speech. For a Shorts ad none of that binds: the feed only plays the first 60 seconds anyway.

One deliberate omission from that table: Sora 2 and Sora 2 Pro sell 4s, 8s and 12s clips, and OpenAI is retiring both on 24 September 2026. Do not build a Shorts workflow on a model with a published shutdown date.

What clears the floor on a free trial

Two limits stack here, and vendors usually mention only the first. Six models are blocked outright during our 3-day trial: Kling 3.0, Seedance 2.0, Sora 2, Sora 2 Pro, OmniHuman 1.5 and VEED Fabric 1.0. Separately, any single trial render is capped at 250 credits, and a render over that is refused rather than trimmed. Put the 10-second in-stream floor on top and only three models satisfy all three at once:

• Grok Video, 10 seconds at 720p, 140 credits, or at 480p for 100.
• Seedance 1.5 Pro, 12 seconds at 1080p, 245 credits.
• Pruna Avatar, script-length, 18.4 credits per spoken second, so a 10-second read is 184 credits, 13 seconds is 239, and 14 seconds crosses the cap at 258.

Kling 2.6 and Happy Horse 1.1 are both open during the trial and both reach 10 seconds on a paid plan, but their 10-second tiers are 260 and 370 credits, so on a trial you only get their 5-second tiers, which do not clear the in-stream floor. Veo 3.1 Fast fits the cap comfortably at 245 for 8 seconds and still cannot reach 10. Worth knowing before you spend a trial render finding out.

The 4:5 gap nobody mentions

Demand Gen accepts 4:5 vertical at 1080x1350, and that orientation earns real impressions in the YouTube home feed and on Discover. But a comment in our own model registry describes 4:5 as a value "which no model here accepts". Every video model in our catalogue exposes some subset of 16:9, 9:16 and 1:1, and the three audio-driven talking models simply inherit the aspect of the still image you feed them. There is no 4:5 generation path.

So 4:5 is a post-processing step in every AI stack, whether or not the tool tells you. Our render endpoint does it in ffmpeg against a fixed 1080x1350 canvas, with two behaviours: the default scales to cover and centre-crops the overflow, and a pad mode letterboxes instead. Crop is usually right for UGC, because Google notes that a rotated device shows a vertical ad "with black bars to either side of the video" and you do not want to add your own bars on top of that. But crop eats headroom, so if the presenter's face is near the top of a 9:16 frame, generate at 1:1 and crop down from there instead.

What to actually ship

• Generate 9:16 at 1080x1920 as the master. It is the format Google says is best suited to Shorts, and it re-crops cleanly to 1:1 and 4:5.
• Target 15 to 45 seconds. Above 10 seconds keeps in-stream eligibility, above 15 matches Google's own reach recommendation, and under 60 keeps the whole ad inside the Shorts feed.
• If your best-performing model caps below 10 seconds, stitch. A three-shot 8-second cutdown solves the eligibility problem and gives you a hook, demo and call to action anyway. Our 30-second UGC script structure is built around exactly that shot count.
• Upload the asset to your own YouTube channel as unlisted, and set the altered-or-synthetic disclosure on that upload before you attach it to a campaign.
• Burn a visible AI disclosure into the creative if you serve the EU, India or New York. Google states that labels you add directly to your creatives will not violate its text-overlay and watermark policies.
• Write the 90-character description before you build the ad, not after. It is short enough to need a real edit.
• Upload two to five variants per ad, in more than one orientation, so one campaign can fill Shorts, in-feed and Discover placements.

None of this is exotic. It is the same discipline as any other channel: read the spec sheet, respect the two rulebooks that apply, and pick a model whose single-clip ceiling matches the placement you are actually buying. The part that trips people is that on YouTube the ad is also an upload, so the creator rules follow it in.

Sources

• YouTube Help, "Disclosing use of altered or synthetic content"
• Google Ads Help, "YouTube Shorts ads: Asset specs and best practices"
• Google Ads Help, "About video assets specifications and ad format guidelines for Demand Gen campaigns"
• Google Ads Help, "Demand Gen campaign creative: Asset specifications and guidelines"
• Google Ads Help, "Use square and vertical video to engage mobile customers"
• Google Ads Policy Help, "Updates to AI labeling requirements (July 2026)", posted 9 July 2026

All six were read directly in August 2026. Model durations and aspect-ratio support are read from our own production pricing table and model registry on the same date. Policies and model catalogues both move; those dates are your check for whether this page has gone stale.

UGC Vids AI generates the vertical creative itself, and re-aspects and stitches it in the same pass, so a Shorts master and a 4:5 cutdown come out of one render. Plans start at $49/month for 5,000 credits, and there is a free 3-day trial. Six models are blocked during the trial (Kling 3.0, Seedance 2.0, Sora 2, Sora 2 Pro, OmniHuman 1.5 and VEED Fabric 1.0), and any single trial render is capped at 250 credits, so if you want a trial clip that clears the 10-second in-stream floor, use Grok Video at 10 seconds and 720p, Seedance 1.5 Pro at 12 seconds, or a short Pruna Avatar script. If you are choosing between them, the best AI video model for product demos compares them on output, not on price.