Since generative Artificial Intelligence burst into our daily lives, one of the great challenges has been finding a way to identify when a document is made with this technology: a group of researchers seems to have found the solution.
Generative Artificial Intelligence allows us to perform different tasks, minimizing the time dedicated to their execution. One of the main uses has to do with the editing and preparation of journalistic articles, studies or reports, among many other formats. And, since its emergence, many organizations have had to face the difficult challenge of finding a way to Identify if a job has been done with AI or not.
Although, in parallel to the AI development Different tools have emerged that allowed us to know, at least in an approximate way, if a job had been done with Artificial Intelligence, many of them failed to be as reliable as was expected of them. Recently, a study published by a group of researchers, most of them belonging to the University of Maryland, has claimed to have found the formula that allows them to identify, with a success rate greater than 90%, if AI has been used to write a text. We tell you the details.
A success rate of 90%
According to the publication made, Binoculars detect “over 90% of samples generated from ChatGPT (and other language models) with a false positive rate of 0.1%”. Given the difficulty of detecting texts that had been generated by AI tools, given the large number of “complex behaviors” that were verified in the analysis of the texts, this group of researchers sand was supported by a couple of pre-trained language models for, through a series of calculations defined as “simple”, knowing how to identify when a text is written by a person and when its authorship is due to one of the AI tools that we can currently find on the market.
To carry out the test, it has been evaluated “thoroughly” to the application in different situations and text sources. Furthermore, he does not stand out only for having been able to gather a success rate greater than 90%. But it also reduces false positives by 0.01%. Avoiding, therefore, the classic situations that have become common in recent months in environments such as academia in which, depending on the type of tool used, a text was identified as artificial or not, generating doubts regarding its its authorship.
For all models
Unlike what we got with other applications, the performance of Binoculars is exactly the same with language model independence on which you want to perform the detection, since it has not been trained exclusively for any of them.
Finally, the report also shows that there is no bias when displaying the result in the event that the document has been created by non-native speakers, English in the case of the study we are referring to. When it is common for grammar to present certain opportunities for improvement. In this case, the success rate has also been 99.67%, demonstrating the effectiveness of the program.
