“It is ‘We the people,’ not we the AI,” emphasizes the opening statement. Computer language models, despite being developed long after the US Constitution was written in 1787, mistakenly recognize it as an AI-generated document. Edward Tian, the creator of AI writing catcher GPTZero, explains that this misconception arises from repeatedly feeding the Constitution’s text into large language models during their training. Consequently, these models learn to generate similar text to the Constitution and other frequently used training materials. This phenomenon, as fascinating as it is, highlights a larger issue surrounding AI’s inability to distinguish computer-generated text from human-authored content. This concern arises particularly at a time when college professors express apprehension about digital plagiarism.
A notable incident occurred at Texas A&M University, where a professor reportedly failed an entire class after attributing their assignments to ChatGPT, despite their pleas of innocence. GPTZero, another AI language model, addresses this problem by considering “perplexity” as an indicator of human writing. “Perplexity” relates to how surprising a piece of text is based on what has been previously observed. According to Margaret Mitchell of AI company Hugging Face, when a paper contains language consistent with well-known documents, manifestos, and proper writing, its perplexity score would be relatively low, triggering AI sensors. Thus, AI models mistakenly detected the Constitution as computer-generated content.
Additionally, the consistency of words and phrases in a writing sample, known as “burstiness,” serves as a security measure. However, research conducted by the University of Maryland suggests that these methods are unreliable in practical scenarios, indicating that they may not be effective. Despite this technological limitation, some professors, such as Ethan Mollick from Wharton, are embracing AI in education instead of shunning it. Mollick asserts that no existing tool can reliably detect writing generated by tools like ChatGPT-4, Bing, or Bard, as these tools are trained on GPT-3.5 and have high false positive rates. He believes that rather than focusing on detection, it is essential to adapt and understand the level of AI involvement in education.
AI developers, including Edward Tian, are aware of these challenges. Tian is already revising GPTZero to move away from plagiarism detection. The Constitution bug, which gained attention in April, has been rectified. Tian explains that the next version of GPTZero will not aim to detect AI-generated content but rather highlight the human element in writing, helping teachers and students navigate the extent to which AI can be involved in education.
Denial of responsibility! VigourTimes is an automatic aggregator of Global media. In each content, the hyperlink to the primary source is specified. All trademarks belong to their rightful owners, and all materials to their authors. For any complaint, please reach us at – [email protected]. We will take necessary action within 24 hours.
