AI is quietly reshaping the way we talk

TikTok gave us slang like rizz, while X popularized ratio and doomscroll. But according to new research from Florida State University, the newest force shaping language isn’t a person or platform: It’s artificial intelligence.

In a peer-reviewed study published in the Cornell University archive arXiv, FSU researchers found that AI is influencing not just how we write, but how we speak. After analyzing more than 22 million words from unscripted podcasts, the team observed a surge in terms favored by large language models (LLMs) like OpenAI’s ChatGPT (delve, boast, meticulous, and garner to name a few), while use of their synonyms remained relatively flat.

The researchers call this the “seep-in effect” or “lexical seepage.” Unlike slang spread by subcultures or mass media, this shift originates with an algorithm. In cognitive psychology, this is known as implicit learning, where recurring phrasing and word choices are unconsciously stored in memory. Likewise, language research also highlights a phenomenon known as priming, where exposure to specific words or syntax leads to an increased likelihood of using them later. In just a few years, the chatbot’s preferred vocabulary has moved off-screen and into daily conversation.

“AI may literally be putting words into our mouths, as repeated exposure leads people to internalize and reuse buzzwords they might not have chosen naturally,” says Tom Juzek, a computational linguistics professor at FSU and lead author of the study. “The deeper concern is that the very same mechanism could shape not just vocabulary but also beliefs and values.”

The study team—Juzek, Bryce Anderson, and Riley Galpin—analyzed 1,326 episodes of tech and science podcasts, split evenly between a pre-ChatGPT period (2019 to 2021) and a post-ChatGPT period (2023 to 2025). They drew on transcripts where possible, or generated them with OpenAI’s Whisper model, resulting in a dataset of about 22 million words. They then compared per-million usage rates of AI-associated buzzwords against close synonyms to test whether shifts reflected ordinary drift or a distinct AI-style influence.

“It was important that this was unscripted language, so we focused on conversational shows—Lex Fridman, Radiolab, Ologies—to capture something close to spontaneous speech,” Juzek says. “We explicitly excluded sources such as conference talks or lectures, which are often scripted and may even be AI-assisted.”

He explains that LLMs don’t inherently overuse buzzwords during pretraining on massive datasets. The effect arises later, during human preference learning. “From what we know, raters tend to be young, so ideas about what counts as formal writing may vary,” says Juzek. “AI model fine-tuning involves tricky trade-offs to achieve usefulness, truthfulness/grounding, and getting high-quality preference data is expensive and hard to obtain. Humans often reward style over substance, so models may pick up ‘polished’ buzzwords in the process.”

A similar study in Germany found near-identical patterns on YouTube, suggesting the phenomenon extends beyond American podcasts to other languages and contexts.

Is AI Standardizing Human Speech?

The implications reach beyond word choice. If OpenAI, Anthropic, or Google fine-tune their models differently, populations could adopt subtly distinct speech patterns. Experts warn this could flatten dialects, erase regional slang, and dampen creativity.

“While AI does reflect patterns already present, by amplifying and projecting the ‘highest-value’ version of those patterns learned from millions of interactions, it dramatically shifts the balance of which language forms dominate,” says Moti Moravia, cofounder and CTO of Leo AI. “Even though you can set parameters for diversity, the main goal of AI models is to maximize perceived quality.”

While speech patterns have always evolved, today the shift is happening with unprecedented speed. AI Models like ChatGPT, Bard, and Claude are trained on billions of words through web scraping and are used by millions of people almost every day. If algorithms quietly prune our synonym choices, they could also be narrowing how we frame ideas.

AI systems tend to magnify dominant language patterns, which speeds up their adoption in broader culture. Without continued human input, they could stagnate, replaying the past instead of adapting to the present. The result might be a creative landscape that feels out of sync with reality—unless new frameworks are developed to prioritize originality. “This is a terrifying future, but we still have time to change this and build in frameworks so that original human creativity is still rewarded,” says Trip Adler, cofounder and CEO of Created by Humans.

Likewise, Moravia argues that companies like OpenAI, Anthropic, and Google will continue to chase higher benchmarks by training on the best data available, optimizing for the metrics they know how to measure. The safeguard, he suggests, is to establish a new benchmark—one that explicitly values diversity in language and beyond. Companies should be incentivized to optimize for more varied outputs in the way models speak. “This would be a subtle yet powerful way to encourage AI systems to preserve linguistic diversity rather than unintentionally narrow it.”

Holding on to the Human Tone

Juzek cautions that the rise in certain words doesn’t prove AI is the sole driver. Many were already trending before 2022, and AI may simply be accelerating an existing shift. “It took years before we understood the full mix of benefits and risks that came with social media, and I suspect it will be similar with AI models,” he says. “Conversations with colleagues tell me that this ‘small tweaks snowball’ effect may be inherent to gradient descent, the optimization procedure at the core of how the models learn. Understanding that properly will require more foundational research.”

Looking ahead, he expects language change to accelerate. Some AI-favored words may fade, much like generational slang, but the larger risk is subtle homogenization.

“Culturally, this matters for trust and creativity,” Juzek says. “Sooner or later, that same uncertainty will reach spoken interactions, for example, phone calls. Arguably, face-to-face conversations remain safe for the foreseeable future.”

https://www.fastcompany.com/91398460/ai-is-quietly-reshaping-the-way-we-talk?partner=rss&utm_source=rss&utm_medium=feed&utm_campaign=rss+fastcompany&utm_content=rss