ChatGPT: how the new system works to detect texts written with Artificial Intelligence
Now users will be able to quickly and easily verify if a text was written by a person or if it was created using Artificial Intelligence thanks to a new tool developed by the creators of ChatGPT
OpenAI, the company responsible for the creation of ChatGPT, presented a new system powered by Artificial Intelligence (AI) capable of detecting those texts that were written by a machine.
The new system comes to respond to a problem that became widespread after the arrival of ChatGPT on the market, such as the indiscriminate use of the tool by students to cheat by submitting work that was not written by them, but by the machine.
“We have trained a classifier to distinguish between human-written text and AI-written text from a variety of providers. Although it is impossible to reliably detect all text written by AI, we believe that this type of system can help prevent fake news.
They clarify that although this kind of tool is not infallible, it can help detect possible cases in which an AI was used to write a text.
In this sense, they indicated that their system has a false positive rate of 9%. This means that on average 1 in 10 texts reviewed by the OpenAI AI will return a message that it was written by a machine, when in fact it was written by a human being.
“The reliability of our classifier generally improves as the length of the input text increases. Compared to our previously released classifier, this new classifier is significantly more reliable on the text of newer AI systems.”
OpenAI highlighted that the tool will be freely available worldwide to make it easy for anyone who might need it to have access to it. They added that they hope to continue to develop this system in such a way that it improves its effectiveness and continues to be useful.
The text detection system has a series of limitations which were exposed by its developer. This includes aspects such as the fact that its reliability decreases significantly when analyzing short texts with an extension of less than 1,000 words.
Similarly, they recommend that the tool be used only to analyze texts written in English, since doing so in other languages can generate erroneous results.
Finally, they add that in those cases in which the text written by an AI was modified by a person, the system will not be able to classify them as written by a machine.