Tracing distinctive language in AI-written text

Allen Institute (AI2) · 4 weeks ago · 39

Researchers at Stony Brook University used Ai2's infini-gram search engine to trace distinctive phrases in AI-generated text back to their sources in training corpora, finding that top-selling self-published books on Amazon with substantial detected AI content contained 4.4 percentage points more rare expressions from existing books compared to non-AI books. The analysis examined 200 highest-revenue self-published books in each category and found the gap widened to 22.5 percentage points when compared with award-winning literature. This method provides evidence beyond simple AI detection scores, enabling researchers to identify when AI models reproduce distinctive language patterns from published works and investigate potential copyright violations in AI training data.