The Brand Safety Checklist for AI-Generated Content
A brand safety checklist for AI-generated content: likeness, IP, claims, artifacts, audio, and platform policy checks to run before anything publishes.
The asset that causes a brand safety incident is almost never the one anyone worried about. It's the fourteenth variant in a batch, approved at 5:40 on a Thursday, with a background sign in a language nobody on the team reads, or a stock-looking presenter who bears an uncomfortable resemblance to a competitor's founder. Nobody was careless. There just wasn't a check that would have caught it.
That's what a brand safety checklist is for. Not to slow production down — a good one adds ninety seconds per asset — but to convert "did anyone look at this properly?" into a sequence of specific questions with yes/no answers. The value is that it catches the thing you weren't thinking about, which is by definition the thing that gets you.
This is the checklist, organized by risk category, with the specific failure modes each check exists to catch.
The pre-publish checklist
Run this before an asset enters the publishing queue. Green-tier formats spot-check one in ten; amber and red run it every time.
| # | Check | Fails if |
|---|---|---|
| 1 | Likeness | Any face resembles a real identifiable person you haven't cleared |
| 2 | Third-party IP | Logos, packaging, characters, or architecture belonging to someone else |
| 3 | Text in frame | Any rendered text is misspelled, nonsensical, or in an unintended language |
| 4 | Claims | Any implied performance, health, financial, or comparative claim |
| 5 | Artifacts | Hands, limbs, object continuity, physics that breaks in playback |
| 6 | Audio | Voice matches an approved voice; music is cleared; no unintended dialogue |
| 7 | Context and culture | Gestures, symbols, dress, or setting that reads wrong in a target market |
| 8 | Platform policy | Synthetic-content labeling applied where required |
| 9 | Brand standards | Product colorway, typography, logo placement, safe areas |
| 10 | Adjacency | Where this will run and what it'll sit next to |
Below, the reasoning for the ones that catch the most.
Likeness is the highest-consequence check
Generated faces are synthetic, but "synthetic" doesn't mean "safe." Two distinct problems:
Unintended resemblance. Models occasionally produce faces close enough to a real public figure to be noticed — rare, and catastrophic in a commercial context. The check is simple: does anyone on the review call recognize this person? If two people hesitate, regenerate.
Deliberate likeness without consent. Digital twins and avatars are legitimate — a twin of your own executive, with their agreement, is a real workflow. What isn't legitimate is generating a recognizable person who hasn't agreed, in any context, including "it's obviously satirical." Make this a hard line rather than a judgment call, and make sure freelancers working in your account know it.
The safe route for anything featuring a specific person is an approved avatar or a lipsync route with a documented consent record, not a text prompt describing them.
Text in frame catches more errors than anything else
Rendered text is where generated content most visibly breaks. Modern image models handle typography far better than they used to — Seedream 5.0 Pro handles multilingual text credibly — but "far better" is not "reliably," and the failure mode is highly visible.
Specific things to look for:
- Misspellings in your own brand or product name. Common and embarrassing.
- Background signage in a language you didn't intend, sometimes saying something you'd rather it didn't.
- Numbers. A generated "40% off" that should read 30% is a claims problem, not a typo.
- Text legible at full resolution but illegible at the size it'll actually be viewed.
The reliable fix is structural: don't generate the text that matters. Generate the frame, add the text as an overlay with your caption and typography presets. This also makes it consistent, translatable, and editable later.
Claims: the one legal cares about
Generated video will happily imply things you never intended to claim. A product shot with a visible before-and-after implies efficacy. A financial explainer with a rising chart implies returns. A wellness clip with a confident presenter implies medical authority.
The check is to watch with the sound off and ask: what would a skeptical regulator say this video is claiming? Then confirm you can substantiate it. Same standard as any advertising creative — the difference is that generated content produces implications faster than a brief can anticipate, because the model fills in details nobody specified.
Regulated categories — health, finance, supplements, legal services, children's products — should sit in the top approval tier permanently. The governance structure for that is in AI content governance for brands.
Artifacts: watch it at speed, then at quarter speed
The visual failures that survive a casual look:
- Hands and fingers during fast manipulation. Still the most common giveaway.
- Object continuity across a cut — a cup changes handle position, a logo shifts, a character's shirt changes shade between scenes.
- Physics that reads wrong at normal speed but not in a still. Liquid that pours upward, fabric that doesn't settle, a foot that slides.
- Eye contact where you didn't want it. Models default to subjects facing camera, which reads as a spokesperson rather than a cutaway.
Review method that catches these efficiently: watch once at normal speed for overall feel, once at 0.25x for artifacts, once muted for what the visuals alone imply. Ninety seconds total for a short-form asset.
For multi-scene work, continuity between scenes matters more than perfection within one. Reference images and consistent character generation across scenes reduce this class of problem substantially — visual consistency with reference images covers the mechanics.
Audio checks people skip
- Voice identity. Is this one of your approved voices? An unfamiliar narrator across two of your videos is a consistency failure viewers register even if they can't name it.
- Unintended dialogue. Several video models generate native audio, and a clip prompted as silent b-roll occasionally comes back with someone speaking. Listen before publishing, always.
- Voice cloning consent. A cloned brand voice needs a documented agreement from the person it's cloned from, kept on file. This is the audio equivalent of the likeness check.
Platform policy and disclosure
Requirements differ by platform and change without much notice. The stable approach is a policy of your own that's at least as strict as the strictest platform you publish to.
A workable default: label any asset that depicts a realistic human who is not a real employee or customer, and any asset that depicts an event that did not occur. Apply the label in the platform's native disclosure field where one exists, and in the caption where one doesn't. Put it in the publishing checklist so it isn't a judgment call at 5:40 on a Thursday. Advertising has separate rules in several markets, particularly for political and regulated categories — confirm those before a campaign launches rather than during it.
Adjacency, and making the checklist stick
A perfectly safe asset can become a problem depending on placement — an upbeat product clip running against disaster coverage, a humorous format in a serious context. Two controls: keep a suppression list of formats that shouldn't run during sensitive news cycles, and give your escalation owner authority to pause a whole format rather than one asset. Pausing a single clip while the format keeps publishing variants is the most common version of this going wrong.
As for the checklist itself: put it in the publishing step rather than in a document, keep it to ten items so it doesn't get skimmed, and log which check caught what. After a quarter you'll find two or three checks catch almost everything — and you'll know exactly where your prompts and formats need fixing.
FAQ
What's the biggest brand safety risk with AI-generated content?
Unintended likeness to a real identifiable person, because the consequences are legal rather than reputational and there's no easy remediation after publication. The second is implied claims in regulated categories, which generate quickly and substantiate slowly.
Do we need to label AI-generated content?
Several platforms require labeling synthetic content depicting realistic people or events, and requirements vary by market. The practical approach is a single internal policy at least as strict as your strictest platform, applied at the publishing step rather than case by case.
How do we stop generated text from having spelling errors?
Don't generate the text that matters. Generate the frame and add copy as an overlay using your typography presets — that keeps it consistent, translatable, and editable, and removes the entire class of error.
Can we generate customer testimonials?
No. A fabricated testimonial presented as a real customer is deceptive regardless of how it's produced, and it's the clearest red line in this whole area. Generated presenters delivering clearly-labeled illustrative scenarios are a different thing, and should be labeled as such.
How long should a brand safety review take per asset?
About ninety seconds for short-form: once at normal speed, once at quarter speed for artifacts, once muted for implied claims. Anything longer won't survive contact with a busy week, which is why the list has to stay at ten items.
Build the checklist into the step where assets actually leave your hands — the review and scheduling stage in the AI video editor — so it runs on the way out rather than living in a document somebody bookmarks.