AI-assisted media classification scores images and video for content signals so a platform can route items for review, apply labels, or enforce its own policy at scale. It is a triage tool that narrows a human reviewer’s attention; it is not a decision-maker, and it does not verify anything about the people depicted.
Signals
Classification models emit probability-like scores for a set of content signals. A practical set for adult platforms:
- NSFW — an aggregate score that the item contains explicit material.
- Hard nudity — explicit genital or sexual content.
- Soft nudity — partial or suggestive nudity without explicit content.
- Other labels — violence, drugs, or platform-specific categories you define.
Treat these as independent signals, not a single verdict. A high soft-nudity score with a low hard-nudity score is a different review outcome than both being high, and the customer’s policy decides which combinations are acceptable for which surface.
Configurable thresholds
A score becomes an action through a threshold, and thresholds belong to the customer. Typical bands:
| Score band | Suggested action |
|---|---|
| Below low threshold | Auto-accept under the customer’s policy |
| Between thresholds | Route to human review |
| Above high threshold | Auto-flag for priority review or block per policy |
Thresholds should differ by surface — a public catalogue, a logged-in feed, and a private creator page can each tolerate different content. Make thresholds configurable per tenant and per surface, and log which threshold produced an action so a disputed decision can be explained.
Frame sampling
Video classification samples frames rather than scoring every frame. Sampling design drives both cost and accuracy:
- Even sampling across the duration gives broad coverage cheaply.
- Scene-change sampling concentrates on transitions, catching content that appears only briefly.
- Dense sampling near flagged regions refines a borderline result.
Too few frames miss short explicit scenes; too many raise cost without improving the verdict. The right strategy depends on the content type — a long-form scene and a short clip need different sampling. Record which frames were scored so a reviewer can jump to the evidence.
Routing to review
Classification output is only useful if it lands somewhere actionable. Route scored items into a review queue with:
- The signal scores and the frames that produced them.
- The threshold that triggered review and the surface’s policy.
- Enough context for a reviewer to decide without re-watching the whole asset.
Prioritise the queue by score confidence and by exposure — an item about to be promoted or surfaced publicly is more urgent than a dormant archive item.
What this is not
Be explicit about the boundary. Classification does not and cannot verify:
- Age — it does not establish that anyone depicted is an adult.
- Identity — it does not confirm who is depicted.
- Consent — it does not establish that anyone consented.
- Rights — it does not confirm ownership or licensing.
These are verification and legal obligations that no content classifier satisfies. Presenting a NSFW score as if it were an age or consent check is both inaccurate and dangerous. Age assurance, consent records, and rights documentation are separate systems and separate responsibilities.
Human review remains necessary
Classification reduces the volume a human must inspect; it does not remove the need for review. Models drift, thresholds are imperfect, and adversarial uploads are designed to evade scoring. Keep human reviewers in the loop for anything routed above a threshold, retain an appeal path, and treat the model’s output as evidence rather than a verdict.
Where AdultInfra fits
AdultInfra runs AI-assisted classification with customer-configured thresholds and a clear boundary between content signals and verification. If you need to triage a growing upload queue without outsourcing your policy, start with adult AI inference and bring one content type and one surface.