For better or worse, the rise of ChatGPT as a writing tool, search engine, or conversational buddy has considerably changed how we communicate with each other and with technology. At the same time, ChatGPT’s widespread use has also sparked numerous online debates about whether it’s possible to spot AI-created content by looking at certain cues, like the em dash.

But new research suggests that such AI cues might become increasingly harder to pick out—because we’re starting to speak more like ChatGPT, and not the other way around. 

Researchers at the Max Planck Institute for Human Development, Germany, found that in the 18 months since ChatGPT’s release, so-called “GPT words” noticeably increased in frequency among human users. Previous research had found that ChatGPT influenced written communication for humans, but the researchers were curious as to whether the proliferation of AI impacted how we spoke. 

For the study, the researchers uploaded millions of pages of e-mails, essays, academic papers, and news stories to ChatGPT, then prompted the AI to “polish” the text. Then they identified several words that ChatGPT seemed to favor, such as “delve,” “realm,” or “meticulous”—dubbing them “GPT words.” Finally, they tracked the frequency of GPT words in over 360,000 YouTube videos and 771,000 podcast episodes from before and after ChatGPT’s release. The paper, posted to the preprint server arXiv, has not yet been peer reviewed.

Even with controls to account for synonyms or scripted content, the researchers found that indeed, GPT words have risen to prominence in spoken English. It appears that a cultural feedback loop of sorts has emerged between English-speaking humans and AI. 

“The patterns that are stored in AI technology seem to be transmitting back to the human mind,” study co-author Levin Brinkmann told Scientific American. “It’s natural for humans to imitate one another, but we don’t imitate everyone around us equally,” he added. “We’re more likely to copy what someone else is doing if we perceive them as being knowledgeable or important.”

An increasing number of people are looking to AI as a cultural authority, wherein “machines, originally trained on human data and subsequently exhibiting their own cultural traits, can, in turn, measurably reshape human culture,” the authors wrote in the study.

“‘Delve’ is only the tip of the iceberg,” Brinkmann noted to the Verge. Other frequently used GPT words included “underscore,” “comprehend,” “bolster,” “boast,” “swift,” “inquiry,” “meticulous,” and “groundbreak.”

The study offers some provocative food for thought, but there are some caveats worth noting. First, the researchers analyzed data from a specific set of GPT models: GPT-4, GPT-3.5-turbo, GPT-4-turbo, and GPT-4o. This anchors the study to these specific versions of ChatGPT. OpenAI will undoubtedly introduce new models over the coming months and years, and those upcoming versions are likely to exhibit new forms of language use and word preference. As a result, this study could become dated rather quickly. 

It’s also not clear if ChatGPT truly has a significant influence on more casual forms of verbal language, especially given that the researchers pulled a considerable amount of data from academic sources. What’s more, language and word use evolve over time owing to a wide variety of factors; while ChatGPT may be contributing in some small way to changes in the words we use, it’s important to point out the many other sources in society and culture that contribute to language shift.

AI is entering our subconscious, informing the linguistic patterns that allow us to communicate with one another. What that means for us humans, we’ll have to wait to see. But in the meantime, experts caution that it’d be smart for us to keep a close eye on AI’s influence on culture, communication, and beyond.