Image credit – artificialcorner.com
One of the most widely used online applications in history, ChatGPT has gained popularity among professionals and students for completing homework assignments, university essays, and other jobs.
A Twitter user submitted a screenshot of one famous instance when well-known AI plagiarism checkers indicated that the US Constitution was AI-generated text. 92.15% of the United States Constitution, according to the AI content detector “ZeroGPT,” was purportedly written by AI. A different AI content-detecting tool was reported by another user. It projected that 59% of the US Constitution was produced by AI.
If you submit the United States Constitution to a tool designed to detect text written by AI models, such as ChatGPT. It concludes that the document was almost certainly written by AI.
AI writing detectors like GPTZero, ZeroGPT, and OpenAI’s text classifier can’t be relied on to correctly find text written by large language models (LLMs) like ChatGPT because of the false positives.
The same problem comes up with Bible texts, which also look like they were made by AI. Before we can figure out why these tools make such clear mistakes, we need to know how they work.
What Are the Concepts Of AI Detection?
The different AI writing detectors use somewhat distinct detection methods. But they all start with an AI model that was trained on a large corpus of text (millions of writing examples) and a set of believed rules that decide whether the writing is human or AI.
Next, the system classifies the text using “perplexity” and “burstiness” features. Perplexity measures a text’s variance from an AI model’s training. Perplexity measures how shocking the syntax is based on its observations. It is measured because AI models like ChatGPT use their training data to write texts. It decreases as output matches training data. It will be low if a text’s language matches the model’s training, making it more likely for the AI detector to categorize it as AI-generated. The interesting U.S. Constitution case follows. These models misclassify Constitutional language as AI-generated because it’s so ingrained.
The trouble is that it is perfectly possible for humans to write content with low perplexity, which makes AI writing detectors much less reliable. AI falls short once again in this field.