It's tempting to market a content filter as if it just "knows" what's safe. It doesn't — it's a model making a prediction, and being honest about how that prediction works is the only way to trust it. Here's what's actually happening when Shield looks at something.
The short version
Shield runs an on-device object-detection model that's been trained to recognize specific categories of exposed body parts in an image — not "is this photo bad" as one vague judgment, but a list of specific regions with specific labels and a confidence score for each. Only the categories you've turned on in settings get covered; everything else is left alone.
Photos vs. video
For a photo, Shield can check the actual full-resolution image before you ever see it. For video, checking every single frame in real time isn't realistic on a laptop CPU — so Shield samples a frame roughly once a second while you're watching and re-checks the flagged regions from there. That means video protection is closer to "checked every second" than "checked continuously," which matters if you're thinking about edge cases like something appearing for a fraction of a second between samples.
Where it's cautious on purpose
- Nothing shows until it's checked. If detection hasn't finished, is too slow, or fails outright, the media stays fully covered. It never defaults to visible.
- Ambiguous cases stay covered. A low-confidence detection is still treated as a detection — Shield isn't tuned to only flag the obvious ones.
- It doesn't try to understand context. It's not reading captions, judging intent, or distinguishing art from anything else. It's looking at pixels and matching patterns it was trained on.
Where it can be wrong
Like any detection model, it can miss things (a false negative) or flag something that didn't need it (a false positive). Unusual angles, heavy compression, low light, or content the model simply wasn't trained to recognize can all reduce accuracy. This is exactly why the "cover first" rule exists independent of how good the model is on any given day — the safety property doesn't depend on the model being perfect, it depends on the failure mode always defaulting to covered.
The model will get better over time. The rule that nothing shows before it's checked isn't a version 1 limitation — it's permanent, on every plan Shield ever has.
What you control
You choose which categories get flagged, whether covered regions can be revealed with a click or stay locked, and the visual style of the cover itself. What you can't do — on any plan — is turn off the "cover before confirming" behavior. That part isn't a setting.