AI Content Moderation for Brand Creative

Quick answer — Moderating brand creative means two separate checks: a specialist model screening for explicit imagery, then a language model scoring the asset against written cultural rules for each market you are shipping to. The output is one of three verdicts, not a yes or no.

Quality control in Vitra.ai Universe runs both checks over image, video, audio and text.

This is not platform moderation

Moderating user-generated content is a volume problem: millions of uploads, a small percentage abusive, and the goal is catching them cheaply.

Brand creative is the opposite. Low volume, high stakes, and the failures are rarely explicit. A campaign gets pulled because a gesture reads differently in São Paulo than in Milan, not because of pornography. That means the check has to know things a general safety classifier does not.

Two stages, and the order matters

Stage one is a specialist classifier. A purpose-built NSFW model returns a single probability per image. Not a taxonomy, not bounding boxes — one number against a threshold. It exists to catch the obvious case and to stop it before anything expensive runs.

Stage two is a language model scoring written rules. This is where the actual work happens, and it only runs if stage one passes. Each market has a rule pack — plain-English policy statements like "Hindu deities on footwear, undergarments, or commercial products" — and the model scores the asset against every rule in every selected market in a single call.

Putting the cheap classifier first is deliberate. A blocked image never reaches the paid model.

Three verdicts, not two

Review is the point of the design. A binary pass/fail either blocks work that was fine or ships work that wasn't, and the middle band is where most real creative lands. At the shipped default threshold the bar for Approved is a perfect score, so anything with a flagged concern routes to a person by design.

Scoring is weighted, not averaged

Rules carry a severity — high, medium or low — and the region score is a severity-weighted mean rather than a flat one. A high-severity rule counts four times a low one. That stops a single serious problem being averaged away by twenty clean checks, which is exactly how a naive scorer fails.

Across selected markets the worst verdict wins. Clean in eight markets and blocked in one is blocked.

What it does not do

It does not understand your brand guidelines, your legal position or your campaign strategy. It checks written rules and returns a score with a rationale. The decision is yours, which is why the audit trail matters more than the verdict.

Where to start

It belongs in the pipeline that produces the work, which is what agentic workflows is for.

Run last quarter's campaign through it for the markets you actually shipped to. Not to catch anything — to see what the rule pack flags and decide whether you agree. That disagreement is the useful output on day one.

FAQ

How is brand creative moderation different from platform moderation? Platform moderation is a volume problem hunting explicit content. Brand creative is low volume and high stakes, and the failures are usually cultural rather than explicit - a gesture that reads differently in one market.

What are the three content moderation verdicts? Approved, Review and Blocked. Review is the deliberate middle band for content that flagged something without clearly violating anything, which is where most real creative lands.

Why run a specialist classifier before a language model? Cost and speed. The classifier returns one probability per image and catches the obvious case, so a blocked asset never reaches the more expensive model that scores cultural rules.

How are content moderation scores calculated? As a severity-weighted mean rather than a flat average, so a high-severity rule counts several times a low one. Across markets the worst verdict wins, so clean in eight and blocked in one is blocked.