Machine learning - or machine teaching? Our everyday speech is increasingly being influenced by AI chatbots, which have a preference for certain words and phrases that are now rubbing off on us, researchers say. — Photo: Klaus-Dietmar Gabbert/dpa
BERLIN: Artificial intelligence chatbots have largely been 'trained' by being fed reams of information from the internet, some of it the outcome of years of hard work by some of the world's leading doers and thinkers.
But now it seems that it is people – including university lecturers and others described as intellectuals – who are being trained by AI, even if unwittingly.
A team of researchers based at Germany's Max Planck Institute for Human Development have analysed over a million recent academic talks and podcast episodes, finding what they described as a "measurable" and "abrupt" increase in the use of words that are "preferentially generated" by ChatGPT.
The team claimed their work provides "the first large-scale empirical evidence that AI-driven language shifts are propagating beyond written text into spontaneous spoken communication."
After sifting through 360,000 YouTube broadcasts and twice as many podcasts, the researchers found that since the launch of ChatGPT in 2022, speakers have become increasingly inclined to pepper their broadcasts with words that the chatbot uses regularly, such as delve, comprehend, boast, swift and meticulous.
The team’s research suggests that AI's "linguistic influence" is spreading beyond academia, science and technology, where early use of large language models was more common, to education and business.
Not only is the shift detectable in the "scripted or formal speech" heard in lectures posted on YouTube, but it can also be found in more "conversational" or off-the-cuff podcasting, according to the team, which warned that the machines' growing influence could erode "linguistic and cultural diversity."
In similar findings released in Science Advances, an "extensive word analysis" of medical research papers published between 2010 and 2024 showed "an abrupt increase in the frequency of certain style words" after AI tools were made widely available.
Last year, according to the research led by Germany's University of Tübingen, "at least 13.5%" of biomedical papers bore the hallmarks of being "processed by LLMs." – dpa
