AI in the UGC loop, part 3 — moderation: the layer most teams skip

Real brand safety is the difference between catching the swear word and catching the competitor’s logo in the background of the unboxing video that just went live on the homepage. The first is table stakes. The second is what costs you a conversation with the CMO. This is the layer most teams skip — and the day job taught me that the unsexy layer is usually the one that matters.

What "brand-safe" actually means now

The list of things a moderation layer has to catch has grown long:

• Profanity, hate speech and slurs — the classic filter, still necessary, no longer sufficient.
• Faces of minors — almost every brand’s policy says no, almost no brand actively detects it.
• Competing brand logos and packaging in the background of a clip.
• Unsafe product claims — "cleared my acne in three days" on a beauty PDP is a regulatory issue.
• Copyrighted music, especially in clips repurposed across platforms.
• Sensitive context where the brand association is simply wrong.
• Quality signals — vertical-only, low-light, unstable handheld footage that looks bad on a PDP regardless of legality.

A profanity filter catches one of those. The rest need vision, language understanding and brand context. That is the AI part. But the AI alone is not enough — what makes a moderation programme work is the queue model around it.

The three-tier moderation queue

The single biggest mistake brands make is treating moderation as binary: approved or rejected. Real moderation is three-tier, because the cost of a wrong decision differs wildly across categories.

Tier 1 — hard auto-reject

The unambiguous stuff: clear profanity, faces of minors, high-confidence competitor logos, copyrighted music with a strike. The model rejects, the asset never reaches a human, and the creator gets a polite templated rejection with the reason. No human time spent.

Tier 2 — soft flag, human review

The grey area: "probably a minor, low confidence", "possible competitor product, partial occlusion", "claim language that might be regulated". This is where human judgement matters, and where the AI’s job is to surface the asset with the specific concern timestamped, so the reviewer is not re-watching the whole clip hunting for what is wrong. A good Tier 2 review is 30 to 60 seconds.

Tier 3 — auto-approve, sample audit

The clean stuff goes live. But — and this is the tier most brands forget — you still sample 5 to 10% of Tier 3 for a weekly human audit. Not to catch escapes, but to catch model drift before it becomes a problem.

The SLAs that make it work

A moderation queue without SLAs is a queue that grows. Set them, post them in the channel, report on them weekly.

The numbers to track

You need three, not one. Tier 2 P75 review latency in hours — the health-of-queue metric. Escape rate — of assets that went live, how many were later flagged as a miss. And reject-overturn rate — of the assets the model auto-rejected, how many a human would have approved. A high overturn rate means the model is too aggressive and you are throwing away good content.

Three things to do this quarter

1. 1Write a one-page moderation policy. Not a 40-page legal doc. One page listing what is auto-reject, what is soft-flag, what is auto-approve. If it does not fit on a page, your reviewers cannot apply it consistently.
2. 2Set the three SLAs above and post them where the moderation team can see them. Report weekly.
3. 3Run a Tier 2 review queue, even manually for the first month. The queue model itself is the unlock. AI just makes it scale.

Last in the series: part 4, personalisation — why "newest first" leaves conversion on the table, and the maturity ladder to 1:1 matching. The product view of this stage is the Creator Review page.