AI is quietly reshaping the way we talk
Florida State University researchers found AI buzzwords are seeping into spontaneous speech through unconscious learning, potentially creating algorithmic influence over how we frame ideas and express thoughts.
Artificial intelligence is changing how we speak. Florida State University researchers analyzed 22 million words from unscripted podcasts and found that ChatGPT's signature terms—"delve," "boast," "meticulous," and "garner"—are surging in everyday conversation while their synonyms stay flat.
The researchers call this "lexical seepage," and it differs from typical slang spread. Instead of bubbling up from subcultures or social media, these language shifts originate with algorithms. The mechanism appears rooted in cognitive psychology: repeated exposure unconsciously stores word choices in memory through implicit learning. AI may be putting words in our mouths.
Tom Juzek, the computational linguistics professor who led the study, analyzed 1,326 tech and science podcast episodes split between pre-ChatGPT (2019-2021) and post-ChatGPT (2023-2025) periods. He chose unscripted shows like Lex Fridman and Radiolab to capture spontaneous speech, excluding scripted content that might already be AI-assisted.
The effect doesn't come from initial training on massive datasets. It emerges during human preference learning, where young raters often reward polished-sounding buzzwords over substance. The language drift reflects not just what AI learned from the internet, but what humans told it to prioritize.
The implications extend beyond vocabulary. If different AI companies fine-tune their models differently, populations could adopt distinct speech patterns. Moti Moravia from Leo AI points out that while AI reflects existing patterns, it amplifies the "highest-value" versions learned from millions of interactions, shifting which language forms dominate.
This affects how we think, not just how we talk. If algorithms limit our synonym choices, they're also narrowing how we frame ideas. AI systems magnify dominant language patterns, accelerating their adoption across culture. Without continued human input, they risk replaying the past instead of adapting to present realities.
Juzek cautions this might be inherent to gradient descent—the optimization procedure at the core of how models learn. The "small tweaks snowball" effect he describes suggests something more fundamental than a temporary trend.
Face-to-face conversations remain largely unaffected for now, but phone calls and other mediated speech will probably see similar effects. The risk is homogenization—not dramatic enough to notice immediately, but sufficient to flatten regional dialects and dampen linguistic creativity over time.
https://www.fastcompany.com/91398460/ai-is-quietly-reshaping-the-way-we-talk
