Researchers at the University of Kansas published a study last Wednesday (7) in which represents an algorithm capable of identifying an academic text created ChatGPT with more than 99% accuracy.

With the development of AI-powered chatbots, one of the biggest challenges is reliably distinguishing words generated by robots from those spoken by real people. While some attempts have already been made to develop ChatGPT detectors, so far none of the options have proven effective, including those created by herself. Open AI, the company responsible for ChatGPT. Existing tools are so inefficient that they are practically useless.

The algorithm described in the ChatGPT study has specific applications, but its promising results are encouraging. During the development process, the researchers identified the distinctive features of the AI-generated writing.

A study that has been peer-reviewed and published in a journal Cell Reports Physical Science, describes a method capable of identifying academic papers produced by AI. To do this, the researchers selected a set of 64 scientific articles written by people in different fields, such as biology and physics. These articles were used as the basis for ChatGPT training, resulting in a dataset of 128 AI-generated articles containing a total of 1276 paragraphs of nonsense content. From these artificial paragraphs, the scientists developed the ChatGPT detection algorithm. They then assembled a new dataset to test the algorithm, containing 30 real articles and 60 articles written by ChatGPT, for a total of 1210 paragraphs.

According to the researchers, the algorithm was able to correctly identify articles created by ChatGPT 100% of the time. At the paragraph level, the accuracy was slightly lower, but still impressive, with a 92% detection rate of AI-generated paragraphs being achieved.

Research can be a starting point for tool development

The researchers hope that others will use their work to develop discovery software tailored to different contexts and purposes. “We aim to create an accessible method so that even high school students can develop an AI detector for different types of handwriting.“, said Heather DeSare, one of the authors of the study and professor of chemistry at the University of Kansas, in an interview with EurekAlert.The problem of AI-generated writing needs to be addressed, and people don’t need to have a background in computer science to contribute to the field.“

The study highlights some of the hallmarks of writing ChatGPT. For example, human writers tend to create longer paragraphs, use a wider vocabulary, include more punctuation, and qualify their statements with words like “however”, “but” and “although”. In addition, ChatGPT has less specificity when quoting numbers and mentioning other scientists.

It is important to note that the model developed by Dezer and his team is not ready for use by teachers who want to detect high school plagiarism. The algorithm was developed specifically for academic writing, especially for scientific journals. This may be frustrating for instructors and administrators who have been concerned about plagiarism by ChatGPT for the past six months. However, Deser mentions that it is theoretically possible to apply the same technique to develop a model capable of detecting other types of writing.

It is important to consider that a writer can easily make small changes to the text generated by the chatbot, making it difficult to detect. Despite this, the researchers describe their work as a “proof of concept” and argue that it is possible to develop a more robust and accurate tool with a larger data set.

While the results are encouraging, tech companies and AI advocates say tools like ChatGPT are still in their infancy. It is impossible to say whether such detection methods will remain effective as AI continues to evolve rapidly. The more language models approach human writing, the more difficult it will be to determine the characteristics of writing created by robots.

Mundo Conectado Deal Center: Selection of Discounts and Lowest Prices
Best deals on electronics, cell phones, TVs, soundbars, drones and more

Via: Gizmodo Source: EurekAlert, Cell Reports Physical Science

…..