How it works
Real-time AI safety, running entirely on your phone.
No cloud, no accounts, no waiting on a server. Here's the path a single frame takes, from the moment it appears on your screen to the moment SafeScreen decides whether you should see it.
Capture the screen
A background service mirrors what is on your screen, frame by frame, in any app. It uses the same secure mechanism your phone already has for screen recording, kept entirely on-device.
A fast first pass, every frame
A compact image classifier (~10MB) checks every frame for explicit content, paired with a skin-tone backstop so two weak signals make one robust decision. Cheap enough to run continuously.
A deeper check, only when needed
When something looks like an AI-generated face, a second model takes a closer look. Only on candidate frames, never every frame, so the expensive check never slows things down.
Decide, then intervene
A policy engine combines the signals, smooths them over time, and decides: clear, warn, or blur. If it blurs, you see a label and tap to reveal. You stay in control.
Two jobs, two models
What SafeScreen looks for.
RUNS EVERY FRAME
Explicit imagery
A purpose-built classifier flags nudity and explicit content, reinforced by a skin-tone signal so it stays robust on real-world content. This is the workhorse. It runs continuously and drives the blur.
RUNS WHEN TRIGGERED
AI-generated faces
A second model estimates whether a face was synthetically generated. This is genuinely hard, so SafeScreen surfaces it as a confidence badge that says "possibly AI-generated", never a verdict.
Where the limits are
Said plainly.
- AI-image detection doesn't generalize perfectly. No on-device detector does today. We report it as a confidence signal and keep improving it, never certainty.
- It blurs the whole frame, not a region. Region-precise blur is on the roadmap.
- Today it focuses on images and video. Text moderation is a future addition, not part of v1.