Since ChatGPT was released last November, many have taken refuge in OpenAI's chatbot to cheat on schoolwork (and write it without lifting a finger) and also to create large-scale disinformation campaigns.
At OpenAI they seem to be aware of the dire consequences of their creation and have set to work to create a sort of antidote. The startup specialising in artificial intelligence has launched a programme that prides itself on distinguishing between texts written by real human beings and those generated by machines.
The detection system lurking in the bowels of OpenAI's new program still norway number data works rather poorly, as the company confessed in a post published on its corporate blog on Tuesday. In tests carried out by OpenAI, the software correctly identifies texts generated by machines in only 26% of cases. However, only 9% of texts originally written by real people are wrongly assigned to artificial intelligence .
In this sense, and while Open AI perfects its new tool, it is advisable not to trust the detector 100% when evaluating the authorship (human or robotic) of the texts.
OpenAI’s new detector is accessible for free from the website platform.openai.com/ai-text-classifier , where the user can paste the text they want to analyze and press the “submit” button to be confronted with results , which are classified according to the probability that a text was or was not illuminated by an AI. OpenAI’s program considers four possible options when classifying texts: very unlikely, unlikely, unclear or possible.
It is also worth noting that the program needs a minimum of 1,000 characters to properly carry out its work and is not yet trained to detect texts written jointly by humans and machines.
To train its detector, OpenAI has created a language model using data sets with texts generated by both artificial intelligence and humans. The "dataset" used by OpenAI comes specifically from three sources: Wikipedia, WebText and InstructGPT.