Brand safety for AI-generated ads is won in the brief and only verified in review. Every rule you enforce as a review note gets broken again next batch, while every rule written into a standing constraint block is applied to all output from then on, which is the only version that survives 200 variations a month. On top of that you need one fixed review checklist and a short list of claim categories that never ship without a named human signing off.
Why constraints belong in the brief
A constraint in the brief is enforced on every variation. The same constraint delivered in review is enforced on the one variation somebody noticed, and it reappears next week.
The arithmetic makes this unavoidable. A careful review pass over 20 variations takes 30 to 45 minutes, and each output was decided separately rather than stamped from one template, so you cannot approve a sample and infer the rest. If your process depends on catching problems downstream, your safe output volume is capped at whatever one person can inspect.
The block we would keep in a document and paste into every brief:
- Claims you may make, verbatim. Not the topic, the sentence. "Machine washable at 40 degrees" rather than "mention durability".
- Claims that are blocked, with the same specificity, including anything legal has already refused.
- Required disclaimers, the exact wording, and where they must appear.
- Prohibited footage. Discontinued packaging, superseded logos, former staff, identifiable third parties, competitor products in frame, anything shot under a licence that has expired.
- Prohibited settings. Categories vary, and this is where category rules bite: no product near a road, no alcohol with a car, no children with the product, no clinical or medical setting unless cleared.
- Brand mechanics. Which logo file, which colour values, the correct pronunciation of the brand name in the voiceover, and whether the name may be abbreviated.
- Tone prohibitions. Harder to write, worth the effort. "Never sarcastic, never uses the word premium, no fear-based framing, no urgency language beyond the stated deadline."
- Format limits. Maximum duration, ratios required, no text inside platform safe zones.
- Market-specific rules, per language, since a claim cleared in one market may be blocked in another.
One warning. Constraints subtract, and they compound. Thirty prohibitions over a thin footage library leaves nowhere to go, and you get six near-identical ads and a false conclusion that the tool has no range. If your batch lacks variety, audit the constraint list before blaming the model. Our campaign setup documentation covers where these fields live in practice.
A review checklist that fits in ten minutes
Order matters. Check the things that cost real money first, and automate everything mechanical so human attention is not spent on spelling.
Automate these. Spellcheck, duration limits, aspect ratio and resolution, loudness, safe-zone text collision, presence of required disclaimers, and whether the file name matches your convention. None of these need judgement and all of them are cheap to catch by rule.
A human checks these, in this order.
- Claims. Does any variation say something you cannot substantiate, or something blocked. Read the voiceover transcript, not just watch the video, because a claim in audio is as binding as a claim on screen.
- Numbers. Price, discount, dates, percentages, quantities, delivery times. Any of these being wrong is a consumer-protection matter, not a typo.
- Product accuracy. Is the product shown the product you currently ship, in current packaging, in a colour you actually stock.
- Implied claims. The hardest category. Footage of someone running a marathon next to a supplement implies a performance claim nobody wrote down.
- People and permissions. Everyone on screen has a current release, and nobody has left under circumstances that make their appearance awkward.
- Third-party material. No visible competitor branding, no music outside your licensed set, no logos in the background of a street shot.
- Concept distinctness. Six ideas or one idea six times. This is a quality check, but it belongs here because clustering usually means over-constraining.
Log the rejections with a one-line reason. After two batches the pattern is obvious and it converts directly into new constraints, which is how review load falls over time instead of rising with volume.
Claim categories that always need a human
No batch in these categories should reach a platform without a named person approving it. The consequence of being wrong here is regulatory rather than commercial.
| Claim category | Why it needs a human every time | Practical rule |
|---|---|---|
| Health, medical, supplements | Implied treatment claims are regulated in most markets, and the implication can come from footage rather than words | Pre-cleared wording only, plus a footage prohibition list |
| Financial products, earnings, returns | Requires prescribed risk wording and often a licensed reviewer | Never let a model draft the disclaimer |
| Weight loss and body transformation | Before-and-after imagery is restricted or banned on major platforms | Ban the imagery in the brief, not in review |
| Comparative and superlative claims | "Best", "fastest", "number one" need substantiation you must be able to produce on request | Require the source alongside the claim, or block the words |
| Guarantees, warranties, returns | These are contract terms. Wrong duration is a legal exposure | Exact wording as a constraint, checked against your policy page |
| Environmental and sustainability claims | Greenwashing rules have tightened across markets and vague wording is now the risk | Only specific, verifiable statements. No "eco-friendly" |
| Testimonials and reviews | Must be genuine, current, and typical. Synthesised or paraphrased quotes are not | Quote verbatim from a real review, with the rating and sample size |
| Pricing and promotional terms | Reference pricing and countdown claims are enforced tightly in several markets | Keep price off screen unless the offer is locked |
| Safety, certification, compliance marks | A certification shown that has lapsed is a serious problem | Verify currency of every mark, every quarter |
| Restricted categories: alcohol, gambling, credit, employment, housing | Platform policy plus statutory rules, often with targeting restrictions attached | Category specialist reviews the batch, not a generalist |
| Anything involving children | Consent, depiction rules and category-specific restrictions | Human review of every frame |
If you operate in more than one market, treat each language as a separate review. We support scripts written natively in English, German, French, Spanish, Italian and Hindi, and native drafting is better than translation for readability, but it does not know that a claim cleared for one market is blocked in another. A native speaker who knows the local rules is not optional.
Why generated footage of your product is a compliance risk
This is the part specific to AI, and it is more serious than the tone-of-voice worries that usually dominate the conversation.
A generative model has never seen your SKU. It produces a category-average object, so label typography drifts, proportions shift, seams sit in the wrong place, materials read as the wrong material. The output looks like your product and is not your product. Show it to a customer as the thing they are buying and you have misrepresented what you ship, which sits squarely inside consumer-protection law and platform misleading-content policy, regardless of intent.
Two further mechanisms make it worse. Generated on-screen text is produced as visual texture with no character-level representation, so a price or a claim rendered inside a generated frame can come out subtly wrong while looking authoritative. And generated people can imply endorsement, testimonial or demographic claims that nobody wrote in a brief and nobody cleared.
Our rule, and we apply it to our own product: generate atmosphere, never evidence. Exteriors, weather, water, smoke, abstract texture, background plates behind captions. Short, peripheral, under about two seconds, with no product, no hands and no readable text. Anything a viewer is being asked to believe should be real footage. The mechanism behind these failures is set out in our glossary entry on how generative video models work.
Add one governance item that gets forgotten: keep a record of which variations used generated footage and which shots they were. When somebody asks in six months whether an ad showed the real product, you want a file rather than a memory.
Where we sit, and what we do not do
Disclosure: Genyad is our product. It assembles variations from a library of footage you already own, which is the brand-safe default, since every frame is real material you shot and already own the rights to. Old approved ads make particularly good input because they have already passed brand and legal review.
We have no AI avatars or synthetic presenters, no static banner formats, no product-URL import, no product-feed or CSV template rendering, no predicted performance scores, and no direct publishing to Meta or TikTok. That last one is a safety feature as much as a roadmap gap: you export and upload, so the approval gate stays with a person. We do not review your claims for you, and no tool should tell you it does. On the adjacent question of who can see your footage, our data and privacy documentation has the specifics.
Frequently asked questions
How do I stop an AI tool from producing off-brand ads?
Write the brand rules as constraints in the brief rather than correcting them in review, including tone prohibitions and a list of prohibited footage. A rule in the brief applies to every variation, while a review note applies to one. Anything you have corrected twice belongs in the standing constraint block.
Is AI-generated video safe to use in regulated categories?
Generated footage of the product itself is not, because the model renders a category-average object rather than what you actually ship, which is a misrepresentation risk. Generated atmosphere with no product, no people in close-up and no readable text is usually fine. In health, financial and similar categories, keep every frame that carries a claim as real footage.
Who should sign off on AI-generated ad creative?
A named individual, with a checklist, before anything carries spend. In regulated categories that person needs category knowledge rather than general marketing knowledge, and in multi-market campaigns you need one reviewer per language. Automating approval removes the only step in the process whose purpose is accountability.
How long should reviewing a batch of AI ads take?
Around 30 to 45 minutes for 20 variations if the mechanical checks are automated and the human checks follow a fixed order. If it takes materially longer, the usual cause is missing constraints, so the same corrections keep reappearing. Review time falling batch over batch is the sign your brief is improving.