Academics from the University of Kansas have devised an algorithm that can tell the difference between written content produced by an AI and that created by a human writer, with 99% accuracy. While previous attempts to build algorithms had failed, the Kansas team claimed success by focusing on certain habits unique to human writers. These include their use of long scientific paragraphs and their preference for equivocal language such as although and however. The researchers also stated that their algorithm could be developed to detect other types of AI-written content.
Scientists have increasingly turned to AI as a means to produce written content, sparking debate about the technology’s potential risks and exciting possibilities. OpenAI, the company behind ChatGPT – an AI program designed to produce written text – was among several firms that had attempted to build technology to differentiate between text written by humans and that written by AI, but had struggled to do so effectively until now.
While the algorithm developed by the Kansas team was specifically designed to detect AI in scientific writing, its authors said it was designed to be accessible, allowing even high school students to build an AI detector for different types of writing.
However, the scientists also acknowledged the limitations of their algorithm. It could still be subverted by a human writer who edited an AI-generated text. As a result, the authors said more research was needed to determine the extent of its applicability.
