이미지: Ai2 (Allen AI) Blog
Summary
- Researchers used Ai2's infini-gram engine to analyze how much AI-generated text overlaps with rare phrases from existing publications
- Among the top 200 Amazon bestsellers, books with higher AI-text ratios showed 4.4 percentage points more overlap in rare phrases than those with lower ratios
- When compared against award-winning or shortlisted literary works, the gap widened to 22.5 percentage points
- 연구 주체
- 스토니브룩대 Tuhin Chakrabarty 교수 연구팀
- 사용 도구
- Ai2의 infini-gram 엔진, Google Books 데이터
- 비교 대상
- 아마존 자가출판 고수익 상위 200권씩 AI텍스트 유무 그룹
- 격차 1
- 일반 비교군 대비 희귀 표현 겹침 4.4%p 차이
- 격차 2
- 수상·후보 문학 대비 격차 22.5%p로 확대
- 발단 사건
- Commonwealth Foundation 수상작이 AI 작성 의혹을 받은 'GrantaGate'
When the Award-Winning Story Turned Out to Be AI-Written
The so-called "GrantaGate" incident broke out when a short story awarded by the Commonwealth Foundation was suspected by readers of being machine-written. AI detectors could produce a score indicating the likelihood that a text was machine-generated, but they couldn't reveal where the sentences actually came from. To fill this gap, a research team led by Tuhin Chakrabarty, assistant professor of computer science at Stony Brook University, turned to Ai2's infini-gram engine. Using Google Books data alongside infini-gram, the team split the top 200 highest-earning self-published Amazon bestsellers into two groups based on the proportion of detected AI text, then compared them. The results showed that books with a higher share of AI-generated text overlapped with rare phrases from existing publications 4.4 percentage points more than books with a lower share. When compared against award-winning or shortlisted works, this gap widened to 22.5 percentage points. Indeed, when the researchers ran the original short story at the center of the GrantaGate controversy through infini-gram, the phrase "sour tang of fermenting" was found to nearly match a sentence from a fan fiction site.
infini-gram, a Tool for Finding the Fingerprints of Sentences
infini-gram is a search engine that indexes massive text datasets, allowing users to count how many times a phrase of any length appears across an entire corpus. Ai2 (Allen Institute for AI) has already applied this engine to two projects built on its fully open language model, Olmo: OlmoTrace, which directly compares a model's output sentences against the original text in its training data, and the Creativity Index, which scores a text's originality relative to a reference corpus. Professor Chakrabarty said he learned about infini-gram through these two projects. He noted that "provenance attribution is an extremely difficult problem in machine learning, particularly with LLMs." The starting point for this research was the recognition that an AI detection score alone can only indicate the probability that a text was machine-written, and nothing more. According to the research team, once something is known to be AI-generated, tools like infini-gram can then be used to actually dissect the text sentence by sentence. The study was posted as a paper on arXiv in late July, and that same month, both The Atlantic and The New York Times reportedly covered the research while examining the proliferation of AI-generated books and related copyright issues.
So What Changes as a Result
What this analysis suggests is that AI-written text tends not to generate entirely new language, but rather to recombine rare expressions drawn from existing texts. While a detection score alone makes it difficult to determine plagiarism or potential copyright infringement, combining it with phrase-matching tools like infini-gram opens a path to tracing specifically where a given sentence originated. This could serve as empirical grounding that publishers or literary award organizers might use to screen out AI-generated manuscripts, while also offering a clue for the broader copyright debate over how much and in what ways AI models reproduce specific expressions from their training data.
