Japan startup develops system to tell human writing from AI-generated text

Photo shows the screen of Puddin AI, a system that identifies whether a piece of writing was written by a human or generated by AI, with sentences highlighted in purple indicating portions of the thesis that were copied and pasted.Photo courtesy of Valar Intelligence, Kyodo News An artificial intelligence startup recently developed a system that analyzes academic papers to determine whether they were human-written or AI-generated, hoping for broader adoption by academic institutions. Developed by Valar Intelligence, an Osaka-based startup, the identification system "Puddin AI" requires users to write and submit their papers on its platform.

Kyushu University has been experimentally using it in classes while several dozen universities are considering its adoption, according to the company. The system focuses on the process of writing, recording when a user begins writing, their writing speed and the chronological order of their revisions. When the identification button is pressed, it assesses the "humanness" of a piece of writing using around 200 indicators, such as common human spelling mistakes, pauses taken in between and the estimated amount of time typically required to produce the piece.

Copy and pasting an AI-generated composition drastically shortens the writing process, allowing the system to determine that a particular piece was not written by a human. The results are divided into three categories: AI, AI-supported and Human. The system handles Japanese, English and four other languages, and it could analyze university assignments, academic papers and corporate documents, according to the developer.

Andrew John Chapman, an associate professor of energy economics at Kyushu University, said, "I want students to take their time and write their work on their own. This system can verify originality, making fair evaluation possible." IT