A Nonprofit Dedicated To
Ending Online Child Abuse Through
Polycreek develops advanced AI tools that enable earlier detection, faster reporting, and more effective intervention to protect children online.
Our Mission
Polycreek works to protect children from online exploitation and abuse, including grooming and the spread of harmful content. We develop advanced tools that enable earlier detection, faster reporting, and more effective intervention, while equipping parents, organizations, and digital platforms to create safer online environments. Our mission is to reduce harm and strengthen protection for children in an increasingly connected world.
Our Need to Act Now
21.3M
CyberTipline reports in 2025
Including 61.8 million images, videos, and other files.
NCMEC, 2025 ↗
1 in 4
children face online sexual solicitation
During their childhood. 9.1% face online sexual extortion at least once.
Childlight / Univ. of Edinburgh, 2026 ↗
1.4M
online enticement reports in 2025
Up 158% from 2024, including more than 80,000 sextortion reports.
NCMEC, 2025 ↗
400K+
reports with a generative AI link
Over 182,000 involved offenders possessing or trying to generate AI CSAM.
NCMEC, 2025 ↗
137/day
financial sextortion reports
Up 37% from 2024. At least three dozen U.S. teen boys have taken their lives.
NCMEC, 2025 ↗
1 in 6
kids 9–12 had a sexual interaction online
Across all minors aged 9 to 17, it is 1 in 4.
Thorn, 2025 ↗
19sec
to a high-risk grooming conversation
The fastest case seen in social gaming. The average is 45 minutes.
WeProtect / Crisp, 2023 ↗
4.78M
suspected CSAM records exchanged
Across INHOPE hotlines in 2025. 67% were confirmed illegal.
INHOPE, 2025 ↗
91%
of CSAM victims are aged 3 to 13
Pre-pubescent children remain the vast majority of victims.
INHOPE, 2025 ↗
1 in 5
children never tell anyone
Of minors who had an online sexual interaction, 1 in 5 told no one.
Thorn, 2025 ↗
4 in 10
teens online “almost constantly”
97% of U.S. teens go online every day.
Pew Research, 2025 ↗
141M+
takedown notices issued
176B+ images processed by Project Arachnid as of July 2026.
C3P / Project Arachnid, 2026 ↗
311K
CSAM reports confirmed in 2025
Up 7% on 2024. 317,101 new images and videos were added to IWF’s hash list.
IWF, 2025 ↗
22%
of minors targeted by adults online
Had a sexual interaction with someone they believed to be an adult.
Thorn, 2025 ↗
$5B+
spent yearly jailing offenders
More than 3,000 times the U.S. budget for child abuse prevention research.
WeProtect, 2025 ↗
Behind every one of these numbers is a child. Polycreek exists to protect that child.
Because every child deserves to just be a kid.
Kids used to spend their afternoons outside with bikes and toys. Now they're in Discord servers, Roblox lobbies, and Snapchat group chats.
It's also where the predators are hiding. That's what we're working to end.
Meet Aletheia
Most filters just tell you a conversation looks bad based on keywords or phrases. Aletheia, trained on a 2.53-million-conversation corpus, tells you which user is the predator, the risk score of the conversation, and the confidence value all in about 38 milliseconds or less.
Measured on a held-out benchmark at 20% positive prevalence. Recall, false-positive rate, and AUC don't depend on base rate; real-world precision varies with how rare grooming is in your traffic, which is why Aletheia is built to rank a human review queue rather than auto-action. Full methodology →
Each conversation is providing behavioral pattern context
Aletheia scores your DMs, comments, community chats or any place where children and adults might interact. Each message context comes back with a risk score, so your trust and safety team works the highest priority first, instead of having to manually review and tag.
Drop in and go
One quick API call. You send a conversation, you get a risk score, a rating, and which speaker set it off. Batch up to 100 threads per request. Teams are usually running on live traffic in under a week.
The worst cases surface first
Every conversation comes back scored 0 to 1 and rated Safe through Critical. The Critical ones go to the front of your queue. Everything else is ranked underneath, so your moderators start at the top and work down instead of guessing what to open first.
Audit-first, compliant by design
We don't store conversations. If one gets flagged, we keep a hash of it and the score, and that's it! No message text ever touches our storage. The hash is how you match our verdict back to the thread in your own system. And because every score is timestamped, you have something to show a regulator asking about COPPA, GDPR, DSA, or KOSA.
hey saw your post earlier, you seem cool
haha thanks ur post was funny
ur really mature for ur age tho lol
Flattery + age referencewe should hop in a vc 1on1 sometime, just us
Isolation tacticidk maybe
don't tell ur parents tho they wouldnt get it
Secrecy / parent isolationIllustrative results
Built on real, multi-source data
Aletheia is trained on a 2.53-million-conversation corpus that combines public and private research datasets, real U.S. federal court records, and synthetic grooming scenarios. Pulling from many sources keeps it from leaning on any single platform's language. We validate real-world detection on a held-out set of real court-record conversations.
Two ways to run Aletheia
Call our hosted API, or run the entire model inside your own infrastructure.
Hosted API access
Point your traffic at our endpoint and you're scoring the same day. We handle the servers, the scaling, and the model updates.
- Fastest to integrate
- ~38ms median response
- Always running the latest model version
Recommended for standard platform use
Single-tenant container
The full model runs in your environment. Nothing leaves your network.
- Deploys in your own cloud, on-prem, or air-gapped
- Full data residency and network isolation
- You hold the keys and the logs
Recommended for large-scale commercial use
I founded Polycreek because I believe we can use AI for good, to protect the most vulnerable among us. Every child deserves to grow up safe online, and we won't stop building until that's a reality.
Zeran Johannsen
Founder, Polycreek
“In 2025, 21.3 million reports of suspected child sexual exploitation were made to NCMEC.”
The scale of online child exploitation is growing faster than human review can keep up. That's why we build technology to help close the gap between harm and action through faster detection, faster reporting, and faster intervention.
Notes from Polycreek
A monthly note on what we're building, the stories behind the work, and where we could use help.