The detector is in trouble. More than half of the English articles written by non-native English speakers are thought to have been written by AI.

CTOnews.com, July 13 (Xinhua) researchers at Stanford University have found that if non-native English speakers do not have enough English vocabulary, it is easy to be mistaken for AI-generated content by GPT detectors when writing English articles.

The study, led by James Zou, an assistant professor of biomedical data science at Stanford University, published in the journal Patterns, said the current GPT detector was actually inadequate to assess whether the article was written by AI.

The researchers of ▲ source journal papers used seven popular GPT detectors to examine a total of 179 articles written by real people and found that these detectors treated more than half of articles written by non-native English speakers as AI.

▲ source journal papers CTOnews.com query papers found that the researchers did not publish the name of the GPT detector used for the experiment, but the research results show that when the current GPT detector "identifies whether the TOEFL article is written by AI", the error rate is as high as 61.3%, and there is even a GPT detector with an error rate as high as 97.8%, and all detectors agree that among these articles written by real people, 19.8% of the content is "definitely generated by AI".

Through further analysis, the researchers of ▲ Picture Source periodical papers found that the main reason for the identification errors is that non-native English speakers are limited in the diversity of languages and the choice of words, and the English vocabulary they master is relatively more "formal" and is not often used in some life scenes, thus resulting in the misjudgment of AI.

▲ diagram source Pexels