Early Research! This system is a very early experimental prototype.

Do NOT trust the model's predictions for real-world decisions, we've trained it for very few steps. And keep in mind that we iterate frequently, so any outputs you see here may change significantly as we keep developing it. Our goal here is to showcase how a real system could work, for example when it comes to showing the annotated reasons for the predictions, but you should not expect it to perform well, at least for now.

SYSTEM DEMONSTRATION

TELL: Traced Evidence for Learned Labels

This is a demo of a GRPO-trained model that detects AI-generated text by marking the specific "tells" that indicate AI or human authorship, where we focus on interpretability and evidence rather than just a classifier score (read more in How this works).

OUR APPROACH

Imagine you see this text:

Dublin is the capital city of Ireland and one of the most exciting places I have ever visited. The city is full of history, culture, and friendly people.

and you go to check it on an AI detector:

Other detectors

87%

AI

↑ this is telling

TELL

Dublin is the capital city of IrelandModerate AI | score 0.55: stating the obvious to complete the geographic introduction; models are trained to be didactical and tend to add well-known facts even when they add little new information and one of the most exciting places I have ever visitedWeak Human | score -0.30: personal POV; this kind of grounded first-person phrasing can be a signal that there is a real human perspective behind the text. The city is full of history, culture, and friendly peopleModerate AI | score 0.65: triad; good writing often uses them, but AI text overuses this pattern because neat, balanced lists of positives are statistically overrepresented in training data.

↑ this is showing

Current AI detectors only give you a score. But is that enough?

We think that's not how we can build trust. For example, if I want to "accuse" someone of writing with AI, I need evidence, I can't just say "this website told me it's 87% AI".

So if we want proper AI text detection, we need to know why something is AI-generated. That's what TELL does.

TEST IT YOURSELF

Inference engine: loading

0%