At the start of the year, OpenAI, a renowned artificial intelligence organization, introduced a revolutionary tool with the potential to make a significant impact. This tool aimed to not only benefit the world but also alleviate the challenges faced by professors and teachers, potentially preserving their mental well-being. This ingenious creation was designed to determine whether a piece of content was created using generative AI tools, similar to its well-known counterpart, ChatGPT.
Six months later, the once-promising tool has unfortunately met an untimely demise. It was forced to shut down because it was unable to fulfill its intended purpose.
OpenAI’s AI Classifier, previously known as such, was recently and discreetly deactivated by the company due to its acknowledged “low rate of accuracy.” While the admission was not easily noticeable, the explanation was discreetly added to the original blog post that announced the release of the tool. Unfortunately, the link to the classifier is no longer functional.
However, the company is not easily discouraged by setbacks. OpenAI has openly acknowledged its mistakes and expressed a commitment to learn from them. They have pledged to actively incorporate valuable feedback and are embarking on research to explore more effective provenance techniques for textual content. Furthermore, they have taken the initiative to develop and implement mechanisms that enable users to distinguish between AI-generated audio or visual content and content created by humans.
The current era is characterized by the continuous advancement of highly advanced AI tools, leading to the emergence of a rapidly growing industry focused on AI detectors.
The AI Classifier developed by OpenAI was launched with great excitement, highlighting its capability to distinguish between text written by humans and text generated by an AI. From the very beginning, the company cautiously acknowledged the inherent unreliability of the classifier. The evaluation of a “challenge set” of English texts showed that the identification of AI-written text as “likely AI-written” was only accurate in 26% of cases. However, it was concerning that human-written text was mistakenly labeled as AI-crafted in 9% of cases.
The limitations of the AI Classifier were a significant obstacle. The model faced difficulties when dealing with shorter texts that were less than 1,000 characters long. Additionally, it made mistakes by incorrectly identifying human-written texts as being generated by AI. Furthermore, classifiers that were constructed using neural networks exhibited subpar performance when applied to data that was outside of their designated training set.
Education is a specific domain where accurate AI detection is extremely important. Ever since ChatGPT was introduced in November, educators have expressed concerns about students potentially misusing the chatbot to generate essays and other academic content.
OpenAI has recognized the importance of identifying text generated by AI, particularly in the field of education. The company acknowledged the influence and constraints of AI-generated text classifiers in educational settings and expressed its dedication to increasing efforts to reach out and share knowledge.
Currently, Decrypt’s request for OpenAI’s response has not been answered, which leaves us feeling curious and eager for a reply. The pursuit of perfecting AI detection tools is an ongoing journey, and OpenAI remains committed to relentless innovation and progress.