NewSpark

Features / Authenticity Detection

Know which submissions are actually real.

AI-generated photos and video are showing up in contest and UGC submissions at a scale no moderation queue was built to catch by eye alone. NewSpark trains a content authenticity model on your own moderators' real/fake decisions — not a generic, one-size-fits-all classifier that's never seen your campaign — so flagging gets sharper the more your team uses it.

Trained on your own moderation decisions Per-project model, not a generic classifier Accuracy reported on every training run

A model that learns your submissions, not the whole internet

Off-the-shelf AI-image detectors are trained once, on generic web data, and never adapt to what's actually coming into your campaign. NewSpark instead builds a classifier from your own review history — every approve/reject decision your moderators make becomes a labeled training example specific to your content, your audience, and the generative tools currently being used against it.

🧬 Trained on real moderation decisions

No separate labeling project required — the training data is simply the real/fake calls your team already makes reviewing submissions.

📊 Transparent accuracy reporting

Every training run shows sample counts, the real/fake balance of your data, and measured accuracy — no black box, no vendor-supplied benchmark from someone else's dataset.

🔁 Retrain as new patterns emerge

Generative tools change constantly. Kick off a new training run whenever you want the model caught up on the latest submissions your moderators have reviewed.

🎯 Scoped per project

A model trained on one campaign's submissions doesn't have to guess at a different campaign's audience — each project can train its own.

Built into moderation, not bolted on after

Authenticity scoring shows up right where your team is already working — the same moderation queue used for rights tracking and approval workflows across NewSpark's UGC & gamification tools. There's no separate dashboard to check, and no submission is auto-rejected by the model alone — a human always makes the final call.

Questions we hear a lot

Does this replace human moderators?

No — it's a decision-support tool for your existing moderation queue. Every submission still goes through review; the model surfaces a likelihood score so your team can prioritize what to look at closely instead of treating every submission as equally trustworthy.

How is this different from a generic AI-image detector?

Generic detectors are trained once on the open web and never see your actual submissions. NewSpark trains a model per project, using your own moderators' real/fake decisions as labeled training data — so it learns the specific patterns of content coming into your specific campaign.

Can I see how accurate the model is?

Yes. Every training run reports accuracy against a held-out sample, and you can see the real/fake balance of your training data at a glance before deciding whether to retrain.

Do I need a data science team to use it?

No. Training happens from inside the same moderation workflow your team already uses — approve or reject a submission as usual, and that decision becomes a labeled example. Starting a new training run is a single click.

Worried about AI-generated submissions in your next campaign?

We'll map your moderation workflow in 30 minutes.