Polycreek's solution to detect and prevent online grooming.
Organizations around the world have made remarkable progress in detecting known CSAM.
Trust & Safety teams have developed sophisticated tools for text and image moderation.
But no model sees the whole conversation that leads up to the exploitation.
Which is where grooming actually happens and where Polycreek has found a solution.
The scale of the problem
How grooming actually works
Two decades of academic research have converged on a recognizable lifecycle. Aletheia is designed to recognize eight in-conversation phases, synthesized from at least seven scholarly models including O'Connell (2003), Olson et al. (2007), Black et al. (2015), and Winters & Jeglic (2017), together with Polycreek's internal grooming guidelines. These phases are the conceptual basis for the signals the model captures, not labels it emits.
Behind the screen is a predator, typically catfishing as a minor. Predators will go out of their way to keep up with current video game trends, memes, and what kids are into in order to seem like one of them.
The predator looks for public profiles of children who may look vulnerable. They'll typically be the first to message the child with shared interests.
First contact is almost always casual and flattering with a compliment and a shared interest. The predator's goal isn't the conversation itself, but to appear familiar and safe, like a friend who just gets them.
Over days or weeks the predator becomes the one person who always listens and always understands. By manufacturing a real sense of closeness and secrecy, they make the child feel singled out and special, so the bond feels safer than the people actually around them.
Before escalating, the predator quietly measures the risk. Whether a parent ever checks the phone, who else is around, where the child is when they talk. A normal stranger doesn't ask these questions in this order, which makes this one of the clearest warning signs of grooming.
Isolation rarely looks like a demand. The predator slowly widens the gap between the child and everyone else, becoming the only person who truly understands them. Then the conversation moves somewhere private and secret, and step by step the child's real support network is replaced by the predator alone.
Once enough trust and secrecy are in place, the predator starts pushing for explicit photos or videos. It's framed as proof of trust, a private joke, or something that will always stay between them. To the child it can feel like a real relationship, just the kind of thing boyfriends and girlfriends do for each other, which is exactly what makes it so hard to recognize as abuse.
Once the first image is sent, the affection disappears and the threats begin. The predator uses what they already have to demand more, trapping the child through shame, fear, and the belief that they have nowhere to turn. Many children stay silent for exactly this reason, convinced that telling anyone will only make it worse.
How Aletheia works
Single-turn classifiers cannot capture the cross-turn patterns the literature identifies as discriminative. Aletheia is built to.
A grooming dialogue can run hundreds of turns over weeks. Standard transformer encoders have a fixed input window. The whole conversation does not fit.
Aletheia's solution: tokenize the conversation into overlapping 2,048-token segments, then handle the cross-segment structure as a separate problem. Past 64 segments, the model retains the first quarter and the last three-quarters, preserving both opening and recent behavior.
A shared pretrained transformer (ModernBERT-base, ~150M parameters) processes each segment independently. Pooling produces one hidden vector per segment.
Learned segment-position embeddings preserve the order. The aggregator that comes next will use that order to model phase transitions across the conversation.
A three-layer transformer encoder, sixteen heads, attends over the whole sequence of segment embeddings. Attention pooling collapses the sequence into a single conversation-level representation.
This is where the model learns that the discriminative signal lives in co-occurrence and sequencing across the dialogue, not in any single utterance.
A binary "is this harmful" output is not enough. Any flagged conversation needs to surface which participant is the predatory party so an analyst can review. Aletheia trains a second softmax head that emits user1 / user2 / neither, correct 98.3% of the time on the held-out set.
Joint training minimizes a weighted sum of the harmful (binary cross-entropy) and predator-attribution (cross-entropy) losses. The harmful score is bucketed into five operational risk bands (Safe, Low, Medium, High, Critical) tied to escalation policy.
What the numbers say
Every figure is measured through the exact inference code that serves production, on a held-out benchmark of real court-filing grooming and real human chat.
Held-out ROC AUC:
The training corpus
The academic literature has spent more than a decade overfitting to PAN12. Aletheia is trained against a corpus more than three orders of magnitude larger and substantially more diverse, drawing on public and private research datasets, real public court records, and controlled synthetic data.
Closing the gap
| The status quo | Aletheia | |
|---|---|---|
| Training data | ~1,200 conversations (PAN12) | 2.5M+ across 20+ sources |
| Languages | English only | Corpus spans 10+ languages; validated in English to date |
| Outputs | Binary harmful flag | Harmful score, five-band risk level, predator attribution, confidence |
| Conversation length | Truncated at single-encoder window | Hierarchical, arbitrary length |
| Vocabulary era | Pre-2012 | Contemporary sources, model dated July 2026 |
| Validation | F1 0.85–0.90 on PAN12 | ROC AUC 0.9991 on a real held-out set (court-filing grooming vs. real chat) |
| Delivery | Closed academic artifact | Nonprofit-priced API and licensed deployment |
The conversation is where most grooming actually happens. Aletheia is one piece of the work needed to make it visible to the systems that already protect children from everything else. The only thing worse than the gap that exists today is the assumption that someone else will close it.
A nonprofit mission
All dollars donated go directly toward that mission.