Plenty of apps aimed at families use the word "safe" as if it were a product feature. It's not — at best it's a marketing claim; at worst it's a promise no software can actually make. What matters is the specific design decisions made at specific points in the system, and how you as a parent can see them for yourself.

This post is an honest description of how Nooodle's content filtering actually works. Not marketing language. The real decisions. Nothing here is a guarantee — content filtering is probabilistic and parental supervision is always required.

The problem we were designing around

AI generation models are trained on the internet. The internet contains everything. Even with extensive filtering at the model level, unexpected outputs happen — from innocent prompts, without warning, in ways that are hard to predict in advance.

A young creator asking about "fire" might get something fine. Or they might get something jarring. We can't control what the model does. What we can try to control is what reaches the space.

The core design principle: don't rely on the AI provider alone. Run our own checks, at every step, independently of whatever the model does.

Two checkpoints, not one

Every creation in Nooodle goes through two independent content checks:

Checkpoint 1 — Before generation (prompt screening)
When a prompt is submitted, it doesn't go directly to the AI model. It passes through Nooodle's own content classifier first. We check for categories that have no place in a family creative app: violence, sexual content, hate speech, targeting of real individuals, dangerous activities. Prompts that fail this check are blocked. The creator sees a friendly message explaining why, with alternative ideas to try.

Checkpoint 2 — After generation (output screening)
Even when a prompt passes, we run the generated image or video through a second classifier before displaying it. This catches edge cases — unexpected model outputs, composite effects, results that don't match what the prompt implied. If something fails this check, we suppress the output.

Neither checkpoint is perfect. Filters miss things. When that happens, we want to hear about it.

Why there's no chat, no social feed, no follows

This wasn't a feature we cut for time. It was a first-principles design choice.

Most online incidents involving young users happen in social features: direct messages, comment sections, follower systems, recommendation feeds. Those are the vectors — not the content creation tools.

Nooodle is a creative space, not a social network. There is no direct messaging, no follower feed, no profile search. The only public-facing surface is the Gallery, where approved works appear alongside a chosen nickname — never a real name, age, or location. Every Gallery submission goes through additional human-assisted moderation before it appears publicly.

What the Parent Dashboard shows

The Parent Dashboard surfaces account activity:

  • Every image and video generated — with the exact prompt used
  • Filter events — any prompt that was blocked by the content filter
  • Published works — what's in the Gallery
  • Credit usage — how many tokens have been used and when

What we don't do

No advertising. No behavioural advertising profiles. No third-party ad SDKs. No data sold to anyone. Our revenue comes from parent subscriptions, which keeps incentives aligned with families rather than with ad networks.

We also don't train AI models on creations made in Nooodle accounts. The images and videos generated inside an account are stored to display them back in the space — that's it.

What we're still improving

Content filtering is a practice, not a launch. We review filter events, update classifier categories as edge cases emerge, and tune thresholds when we see false positives blocking innocent prompts.

If the filter blocks something that shouldn't have been blocked — or misses something that should have been — please tell us at hello@nooodle.ai. We read every message.

We're not going to claim the system is perfect. We're describing exactly what we built and how it works, and we'll keep improving it from there.