There is a shift in spoken discourse, a flattening of the texture of human speech. It began in the quiet corners of the academy, but it has since spilled out into the noisy, unscripted sprawl of the internet. You hear it in the way a podcast host might, with suspicious fluidity, promise to “underscore a meticulous execution” or that they “delved into the opportunity to pinpoint rigor” or to “boast of a meticulous integration plan that will bolster adoption”. These are artifacts, little flags planted in our consciousness by an LLM.
A group of researchers from the Max Planck Institute for Human Development set out to find them. The influence of LLMs such as ChatGPT is already taken for granted in written discourse. Instead, the study takes a phonocentric approach, analyzing 740,249 hours of human speech. They analyzed 360,445 academic talks on YouTube and 771,591 conversational podcast episodes, searching for the moment when English broke. Following the release of ChatGPT in November 2022, they found a “measurable and abrupt increase” in specific words falling on a not-so-gentle upward curve.
