AI that protects children without watching them.
Our models run on a child's own phone. They catch unsafe content and the way predators talk, flag it to a parent, and keep nothing.
Children deserve protection that does not spy on them. We think that is possible, and we are building it.
Most online-safety tools watch everything a child does and send it off to a server. We don't want to build that. Our question is harder. Can a model catch real danger while seeing as little as possible, and keeping none of it?
Four years of work say yes. What we have runs on the phone, raises a short flag a parent can act on, and forgets the rest.
Two systems, one job.
The research turns into apps families can actually install. The camera comes first, the full shield follows.
PH Camera
A camera that will not take or keep an unsafe photo. Every frame is checked on the phone and thrown away. The app has no internet permission, so nothing it sees can leave the device.
PH Bulwark
The shield for the whole device. It filters unsafe content in place across apps and the web, warns a parent when something is wrong, and keeps the rest of the page working.
Block in place, not the whole web.
Most filters ban a whole site the moment it might show something unsafe, so a child loses the search, social and learning sites they actually need. We do two things instead. When we can, we pull out only the unsafe parts and leave the page working. When the content is serious, we block it outright.
Filtered in place. The page keeps working.
Blocked outright when it must be.
Small models, each with one safety job.
Most of these already run in our alpha. None of them need the cloud.
Grooming-pattern recognition
A small language model that learns how predators talk. Secrecy pressure, pushing a child onto another app, fishing for their age or address. When it sees the pattern it tells a parent, with none of the message contents attached.
On-device content filtering
It checks images and text as a page loads and blocks the unsafe ones in place. Only the harmful part is removed, so the rest of the page still works. Illegal child-abuse material is blocked on sight and reported as the law requires, and it is never stored or shown.
Reading the screen in encrypted apps
Plain OCR, running on the phone, reads the text already on screen. That lets it catch grooming even inside end-to-end-encrypted chats. It never logs keystrokes or passwords, and nothing it reads leaves the device.
Offender-record matching
Early research into linking convictions that are already on the public court record, to support our journalism. A person reviews every match, and only after a case has been to court.
Our child-safety benchmark
We asked the big AI models to help protect children. Most said no.
We built a benchmark of 36 real child-safety tasks and ran six frontier models through it. GPT-5 refused all five categories. Gemini refused most. Only xAI's Grok took the work on, and Grok-4.1 scored 79.9% on average and 100% on real grooming cases. That gap is a big part of why we build our own models.
Four lines we will not cross.
On device by default
The models run on the child's own phone. An optional filter can route traffic through our servers or one you host yourself, never a third party.
We remember nothing
No raw messages or images are ever kept. A parent gets a short, redacted alert and nothing more.
A person always decides
The models raise a concern. A person decides what happens next, whether that is a parent or one of our editors. Nothing is ever an automated accusation.
Open methods, closed data
We are happy to explain how the models work. We will never hand over a child's data, a grooming dataset, or live model weights.
- 2017
- On the front line since
- 76K
- People in our community
- 0
- Raw messages or images kept
- 100%
- Runs on the phone
A small team, four years in.
Predator Hunters Research is the AI side of a child-protection group that has run since 2017. We are small and self-funded, and we care as much about a child's privacy as we do about their safety.
Builds the models and the systems they run on, and decides what we won't build.
Backers · partners · researchers
Help us build the AI that keeps children safer.
We are a small, self-funded team, four years into this. If you fund safety research, want to build with us, or want to put these models to work protecting children, get in touch.