How well do you really
know your partner?
Write a love letter below. We’ll teach an AI to forge one back in your voice. Then your partner has to guess which words were actually yours.
With love,
Write a love letter below. We’ll teach an AI to forge one back in your voice. Then your partner has to guess which words were actually yours.
That AI forgery we just wrote in your voice? It’s a one-off run of the exact technique Pangram uses to train its AI-detection model. Here’s the real thing.
For each real, human-written text, Pangram prompts a frontier model to write a brand-new one that matches its topic, length, tone, and style — close, but never a copy. Every human example gets an AI shadow.
Without matched pairs, a detector cheats — flagging text as AI just because it’s a different topic or length than the humans it trained on. Mirroring erases those surface tells, so the model has to learn the genuine fingerprints of machine writing.
Train the classifier on ~1M documents
Mine hard negatives — human text wrongly flagged as AI
Synthetically mirror them into AI twins
Retrain on the harder set
So when your partner squints at two letters trying to spot the fake, that’s the very problem Pangram’s model has already solved — millions of times over.