Scientific texts written with ChatGPT can be accurately identified

Researchers have developed a tool to accurately determine, above 99% academic scientific texts which are written with Artificial intelligence.

The study was published in the journalCell Reports Physical Science“.

Even though many AI text scanners are available online and work quite well, no tool has been created specifically for this. academic writing. To fill the void, this particular research group sought to create a more efficient tool for exactly this purpose.

The study focused on a type of article called prospects, which provide an overview of specific research topics written by scholars. The team chose 64 such articles and created 128 ChatGPT articles on the same topics.

When they compared the articles, the researchers found a marker of articles written by AI: predictability.

Unlike artificial intelligence, humans have a more complex paragraph structure, differing in the number of sentences and the total number of words in the paragraph, as well as the length of the sentence.

Preferences in punctuation and vocabulary are also indicative. The team put together 20 features that the tool should pay attention to.

When testing the tool successfully 100% accuracy distinguish AI-generated articles from human-written articles. To identify individual paragraphs in an article, the accuracy of the model was 92%.

In the next phase, the research team wants to test the tool on larger datasets and different types of academic scientific writing.

However, the tool was not designed to collect student reports generated by artificial intelligence, but as the lead author and professor at the University of Kansas points out, Heather Desirepeople can easily copy these methods to create models for their own research purposes.