When chatGPT analyzes itself | | | | Turtles AI

OpenAI has just publicly released a tool that can assess whether or not a text is written by artificial intelligence. To date, this tool has relative (and limited) effectiveness. In fact, as can be read on the OpenAI website, out of a sample of English-language texts, the classifier was able to correctly identify 26 per cent of texts written by an AI, while 9 per cent of the time it incorrectly classified texts written by humans (i.e., indicating them as written by an AI). For this explicit reason, OpenAI warns of the risks of using this system as a tool to make specific decisions (e.g., to exclude school work submitted by a student...or an employee's report). It should therefore be used only as an adjunct to careful, specific assessment, left to us humans, of course. Some other points of note are that the classifier is ineffective for short texts (under 1000 characters). It is also suggested that the classifier be used, for now, only on English texts, since for other languages the results are significantly worse. Furthermore (and it seems obvious to me) it is very difficult to use the classifier successfully on highly predictable texts (e.g., a sequence of numbers, mathematical formulas, or lines of computer code). Finally, this tool will clearly be the subject of future updates aimed at making the detection system increasingly accurate. You can use the classifier by clicking on this link.