Simply put, an AI (artificial intelligence) classifier sorts data into one or more labeled "classes" (or categories). These classifiers can be a useful tool when trying to determine if content is written by a human, or produced by AI (like ChatGPT).
How accurate are classifiers at detecting AI-produced content?
While classifiers are a useful tool for detecting AI-produced content, none are foolproof. All classifiers produce false-positives (AI-generated) and false-negatives (human created). They are not intended, at this time, to provide definitive proof that content is AI-generated. Rather, they should be used as a tool to guide assessment.
For more information, OpenAI has created extensive documentation which can be viewed here.
CopyLeaks detection report highlights the specific elements of the text that have been written by a human and those written by AI, along with a confidence level, offering a new level of insight and transparency.
GPTZero uses "burstiness" and "perplexity" as its two main indicators when determining if text is AI-generated or human created. Perplexity looks at the complexity of the sentences, while burstiness compares sentence variation/structure.
*Signing up with email is required.
*AI text classifiers previously created by Open AI and its predecessor (Open AI text classifier and GLTR) are no longer supported due to low performance rates*