How Does Pangram Work?

> That’s how we train Pangram: for every element of our dataset of commercially licensed human-written texts, we generate an AI double that is as close as possible to its source. The doubles are the same length, same tone, same topic, but AI-generated. We then train the Pangram model on those human-AI text pairs.

Is this training occurring with commercial AI models or open-weight models?

It seems like if this were all being done in an above-board way with commercial models, OpenAI and Anthropic would be able to determine which texts were being used in the training process.

> our dataset of known human text is drawn from 2021 and earlier

That’s great for now, but that corpus is already missing five years of slang and won’t get any younger.

A recent post where the author (unsuccessfully) tries to fool Pangram:

https://thegustafson.com/blog/evading-ai-detection