– Within weeks of ChatGPT's launch, concerns arose about students using the chatbot to produce essays quickly.
– Startups responded to these concerns by developing products that claim to identify whether text was written by a human or an AI.
– New research, yet to be peer-reviewed, reveals that these tools can be easily tricked and fail to detect AI-generated text.
– Debora Weber-Wulff, a professor at HTW Berlin, collaborated with researchers from various universities to evaluate 14 detection tools, including Turnitin, GPT Zero, and Compilatio.
– The majority of these tools rely on identifying AI-generated text characteristics, such as repetition, to determine the likelihood of text being AI-generated.
– However, the study found that the tested tools struggled to detect ChatGPT-generated text that had been slightly rearranged by humans and obfuscated using a paraphrasing tool.
– This suggests that students can bypass the detectors by making slight modifications to the AI-generated essays.