This is an observation that is as amusing as it is worrying, and which could well be the sign of an impoverishment of language over the years. While artificial intelligence is supposed to be inspired by a body of human-produced content, isn't the trend reversing?
If you feel like your favorite content creators or office colleagues are increasingly speaking like artificial intelligence, you're probably not alone in thinking this...
280,000 academic videos analyzed
In 2024, two years after the arrival of ChatGPT, researchers at the Max Planck Institute conducted an initial study on nearly 10,000 texts written by humans and compared them to artificial intelligence. The first observation was already very alarming, and already demonstrated that the words used by LLMs were increasingly present in university content.
In order not to rely solely on this first study, although it was already comprehensive, the researchers at the Max Planck Institute wanted to go further. Indeed, they did not only analyze written content, but also oral content.
To do this, they then carried out a study on the words used in nearly 280,000 academic videos published on YouTube over the last 18 months. After comparing these videos to other content published three years earlier, the research team found that some of the words were used more often, up to 51% for some of them.
Towards a universal and standardized language?
While it is regrettable that the corpus analyzed by the Max Planck Institute does not focus on French-language content, this trend seems to be confirmed through other studies.
In another study conducted by the University of California, Berkeley, researchers had proven that artificial intelligence produces more approximate answers, probably to encourage users not to jump to conclusions.
For the researchers who focused on English-language content, this is even more true when users use English that deviates from the United States through an accent or terms that would not be used by the majority of Americans.
Even more worrying, in addition to providing less precise answers due to fear of not understanding the exact requests of certain users, the study shows that AI would encourage the latter to express themselves differently.
While this standardization of language has not yet been proven on other languages such as French, regular users of ChatGPT are already used to reading expressions less used before such as "il est important" or "en plus", considered as markers of content produced by AI.
0 Comments