Tracing distinctive language in AI-written text

Phrase matching can add inspectable evidence to a claim that text was AI-written. Ai2's infini-gram searches large text collections for exact expressions, showing where wording also occurs. Researchers used it alongside a book index to distinguish rare book-associated phrases from language common on the web. Ai2 describes the method and findings.

In the reported comparison of 100 high-revenue self-published genre books per group, rare expressions covered 43.2% of text in books with substantial detector-flagged AI writing, versus 37.6% in those without detected AI text. This is an association between detector classifications and phrase overlap. A match alone does not establish authorship, the phrase's origin or which source a model trained on.