TikTok gave us slang like rizz, whereas X popularized ratio and doomscroll. However in response to new research from Florida State College, the most recent pressure shaping language isn’t an individual or platform: It’s artificial intelligence.
In a peer-reviewed examine revealed within the Cornell College archive arXiv, FSU researchers discovered that AI is influencing not simply how we write, however how we converse. After analyzing greater than 22 million phrases from unscripted podcasts, the group noticed a surge in phrases favored by massive language fashions (LLMs) like OpenAI’s ChatGPT (delve, boast, meticulous, and garner to call just a few), whereas use of their synonyms remained comparatively flat.
The researchers name this the “seep-in impact” or “lexical seepage.” In contrast to slang unfold by subcultures or mass media, this shift originates with an algorithm. In cognitive psychology, this is named implicit learning, the place recurring phrasing and phrase selections are unconsciously saved in reminiscence. Likewise, language analysis additionally highlights a phenomenon often known as priming, the place publicity to particular phrases or syntax results in an elevated chance of utilizing them later. In just some years, the chatbot’s most popular vocabulary has moved off-screen and into each day dialog.
“AI could actually be placing phrases into our mouths, as repeated publicity leads individuals to internalize and reuse buzzwords they may not have chosen naturally,” says Tom Juzek, a computational linguistics professor at FSU and lead writer of the examine. “The deeper concern is that the exact same mechanism might form not simply vocabulary but in addition beliefs and values.”
The examine group—Juzek, Bryce Anderson, and Riley Galpin—analyzed 1,326 episodes of tech and science podcasts, cut up evenly between a pre-ChatGPT interval (2019 to 2021) and a post-ChatGPT interval (2023 to 2025). They drew on transcripts the place doable, or generated them with OpenAI’s Whisper mannequin, leading to a dataset of about 22 million phrases. They then in contrast per-million utilization charges of AI-associated buzzwords towards shut synonyms to check whether or not shifts mirrored atypical drift or a definite AI-style affect.
“It was essential that this was unscripted language, so we targeted on conversational exhibits—Lex Fridman, Radiolab, Ologies—to seize one thing near spontaneous speech,” Juzek says. “We explicitly excluded sources reminiscent of convention talks or lectures, which are sometimes scripted and will even be AI-assisted.”
He explains that LLMs don’t inherently overuse buzzwords throughout pretraining on huge datasets. The impact arises later, throughout human choice studying. “From what we all know, raters are usually younger, so concepts about what counts as formal writing could fluctuate,” says Juzek. “AI mannequin fine-tuning entails difficult trade-offs to realize usefulness, truthfulness/grounding, and getting high-quality choice information is pricey and laborious to acquire. People typically reward fashion over substance, so fashions could decide up ‘polished’ buzzwords within the course of.”
A similar study in Germany discovered near-identical patterns on YouTube, suggesting the phenomenon extends past American podcasts to different languages and contexts.
Is AI Standardizing Human Speech?
The implications attain past phrase selection. If OpenAI, Anthropic, or Google fine-tune their fashions in another way, populations might undertake subtly distinct speech patterns. Consultants warn this might flatten dialects, erase regional slang, and dampen creativity.
“Whereas AI does replicate patterns already current, by amplifying and projecting the ‘highest-value’ model of these patterns realized from hundreds of thousands of interactions, it dramatically shifts the stability of which language types dominate,” says Moti Moravia, cofounder and CTO of Leo AI. “Although you may set parameters for range, the primary purpose of AI fashions is to maximise perceived high quality.”
Whereas speech patterns have all the time developed, at the moment the shift is going on with unprecedented pace. AI Fashions like ChatGPT, Bard, and Claude are skilled on billions of phrases by internet scraping and are utilized by hundreds of thousands of individuals virtually day by day. If algorithms quietly prune our synonym selections, they is also narrowing how we body concepts.
AI programs are inclined to enlarge dominant language patterns, which accelerates their adoption in broader tradition. With out continued human enter, they might stagnate, replaying the previous as an alternative of adapting to the current. The end result may be a inventive panorama that feels out of sync with actuality—except new frameworks are developed to prioritize originality. “This can be a terrifying future, however we nonetheless have time to alter this and construct in frameworks in order that unique human creativity remains to be rewarded,” says Journey Adler, cofounder and CEO of Created by People.
Likewise, Moravia argues that firms like OpenAI, Anthropic, and Google will proceed to chase greater benchmarks by coaching on the very best information obtainable, optimizing for the metrics they know tips on how to measure. The safeguard, he suggests, is to ascertain a brand new benchmark—one which explicitly values range in language and past. Firms needs to be incentivized to optimize for extra diversified outputs in the way in which fashions converse. “This could be a delicate but highly effective technique to encourage AI programs to protect linguistic range quite than unintentionally slim it.”
Holding on to the Human Tone
Juzek cautions that the rise in sure phrases doesn’t show AI is the only real driver. Many had been already trending earlier than 2022, and AI could merely be accelerating an present shift. “It took years earlier than we understood the complete combine of advantages and dangers that got here with social media, and I believe it will likely be related with AI fashions,” he says. “Conversations with colleagues inform me that this ‘small tweaks snowball’ impact could also be inherent to gradient descent, the optimization process on the core of how the fashions study. Understanding that correctly would require extra foundational analysis.”
Trying forward, he expects language change to speed up. Some AI-favored phrases could fade, very like generational slang, however the bigger threat is delicate homogenization.
“Culturally, this issues for belief and creativity,” Juzek says. “In the end, that very same uncertainty will attain spoken interactions, for instance, cellphone calls. Arguably, face-to-face conversations stay protected for the foreseeable future.”
