AI text detectors easily fooled

 

A new study shows that tools purporting to detect AI-generated text can be easily fooled.  The study evaluated 14 detection tools, including Turnitin, GPT Zero, and Compilatio, to identify words generated by ChatGPT. To bypass these detectors, students could simply tweak their essays and coursework, the results suggest. The researchers found that the tools, often marketed as anti-plagiarism platforms for teachers, had trouble detecting text that had been slightly rearranged by humans and altered using paraphrasing. While the tools accurately detected human-written text an average of 96% of the time, the detection accuracy rate of ChatGPT text was only 74% and noticeably lower — 42% — for slightly tweaked ChatGPT content. According to lead author Debora Weber-Wulff, the tools "don't work" and "don't do what they say they do." The study was published to the preprint server arXiv, so it has yet to be peer-reviewed.