2023-06-09 11:15:36
The debut of the artificial intelligence system called ChatGPT, and especially that of its latest version, has not ceased to amaze the world with its remarkable ability to produce coherent and informative texts on the topics that are requested. The ability of this artificial intelligence to generate texts on any subject on which there is information on the Internet can serve as a shortcut to obtain summaries on issues that do not have a specific word in an encyclopedia. But what if a student, instead of writing the essay, project or thesis that his teacher has commissioned, asks artificial intelligence to write a text of the required topic and length and then presents it to the teacher as his own? This is much more sophisticated than the trick of copying content from the internet and signing it as your own.
Recent research has explored this question.
Heather Desaire’s team, from the University of Kansas in the United States, asked ChatGPT to write 128 articles on as many topics. In each article, the most important aspects of the subject matter should be exposed, as a general vision of it.
Comparing those 128 texts with others with the same themes and focus but written by humans, Desaire and his colleagues detected telltale features of the non-human authorship of the texts. Although they turned to another artificial intelligence to help them with the very subtle features, many other features are obvious to the human mind with a simple read.
Contrary to what happens with artificial intelligence, humans have more complex paragraph structures, with large variations in the number of sentences and total words per paragraph, as well as notable fluctuations in sentence length. A text written by an artificial intelligence is usually quite schematic and orderly, and usually has rather short sentences and a relatively simple and predictable architecture. Preferences in terms of punctuation and vocabulary also reveal whether the text has been written by a person or by artificial intelligence.
The team identified a total of twenty telltale features of this non-human authorship. By feeding this information to another artificial intelligence, Desaire and his colleagues tested it on its ability to discern when a text had been written by a human and when by ChatGPT.
In cases where the inspection system was allowed to read the entire document, the hit rate in detecting that the authorship was not human was 100 percent. When he was only allowed to read one paragraph of each document, his hit rate dropped to 92 percent.
Casting as your own a text that has actually been written by an artificial intelligence is not as easy as it may seem. There are ways to detect if a human has written it or if, on the contrary, it has been done by an artificial intelligence like ChatGPT. (Image: Heather Snub/Romana Jarosova/University of Kansas. CC BY-SA)
The inspection system prepared by the research team far outperformed a commercially available system that does the same type of work.
El estudio se titula “Distinguishing academic science writing from humans or ChatGPT with over 99% accuracy using off-the-shelf machine learning tools“. Y se ha publicado en la revista académica Cell Reports Physical Science. (Fuente: NCYT de Amazings)
#Artificial #intelligence #school #work #ultimate #trick #cheating #students