Equity · Wednesday, September 2, 2026
Pangram trains its AI detection models using a dataset composed of 50% human-written text and 50% AI-generated text. This approach allows the model to learn stylistic differences and estimate the degree of AI influence in written content.
“So basically the total training set is 50% human text and 50% AI text.”
“And then we're able to teach our model to estimate the degree of change by just looking at the final product.”