Researchers Chengzhi Mao, Carl Vondrick, Hao Wang and Junfeng Yang found that large language models (LLMs) used by artificial intelligence applications are more likely to modify human-written text than human-generated text. AI when they are responsible for revision or rewriting.
Difficult to hide your talent
AIs scrupulously follow the rules of the language; they are much less likely to correct a text written by a Logical AI. This is precisely the key point.
The Raidar app developed by the researchers measures the “distance” between the original text and the text rewritten by the AI. They submitted academic or commercial texts, literary texts, student work, computer code, reviews on Yelp (a business and services evaluation service), arXiv articles (scientific articles before revision). In all cases, the application did better than existing tools in detecting articles written by AI, up to a gain of 29 points.
Even more interesting, even if AIs are asked to write “in human style” or to “avoid being detected by detection applications”, they will still write following rules of what they determine to be altered handwriting, which will appear strange to say the least and produce a specific characteristic writing imprint.
It will always be possible with a lot of work and effort to pass off an AI-generated text as an original text, but in the end, it will be faster for a student to write an original text, and above all better for their student. self-esteem.
The robust results obtained by the researchers illustrate the unique imprint of machine-generated texts through the alterations produced by the machines themselves.
Already teachers recognize relatively easily when a student strays from their own style, soon they will be able to demonstrate even more easily what has been written by an AI Too standard, it’s suspicious!
Image : Pexels – Pixabay
For the original text: Raidar : geneRative AI Detection viA Rewriting – Chengzhi Mao & Carl Vondrick & Hao Wang & Junfeng Yang – Columbia University and Rutgers University
https://arxiv.org/pdf/2401.12970
See more articles by this author
