In English, AI text generators like ChatGPT are often indistinguishable from human text - but that's not the case in all languages. — Photo: Hannes P Albert/dpa
TOKYO: Recent advances with and popularisation of so-called "generative" artificial intelligence has stoked concerns about fake homework, fake academic papers, fake journalism and fake identities.
But one place where there might not be as much to worry about is the world’s third biggest economy and one of its most advanced.
A team of Japanese researchers have shown that "ChatGPT-authored Japanese writing" is stylistically distinguishable "with up to 100% accuracy from human-authored text." The academics, from Mejiro University in Tokyo and Kyoto University of Advanced Science, took 72 academic papers and compared them to ChatGPT-derived equivalents using an AI system for detecting AI texts.
The outcome, published August 9, was in stark contrast to AI’s ability to detect fakery in English, which going by some measurements and tests stands at only 26%. In other words, as the team reported when writing up their findings in the journal PLOS One, "74% of AI-generated texts are incorrectly classified as human-written." The team said similar tests could be done using other platforms, such as Google Bard and Microsoft’s Bing AI.
Other recently-published research found the AI detector bots to be "biased" against non-native speakers of English, whose writings were more often incorrectly labelled as AI-generated compared with those of native speakers. Another paper showed that speakers of English and Mandarin Chinese were equally likely to fall for AI-generated deepfake voices. – dpa
