From “rizz” to “meat proxy,” words and phrases have fallen in and out of favor for centuries. Culture, politics and technological development have long influenced these trends, and now, new research shows that ChatGPT is beginning to influence how humans communicate.
The findings from a team at the Pew Research Center show how, in a short time since OpenAI released ChatGPT in November 2022 and other chatbots followed, the use of certain words and phrases has spiked on the web. The researchers, led by data scientist Samuel Bestvater, studied nearly half a million English language webpage texts in snapshots over the past five years. Not only has AI-written text become more common, but it could also be influencing how people write, such as trying to avoid using the “tells” that have become associated with AI-generated text.
Bestvater and his team detected double the number of em dashes, double the frequency of particular words — such as “delve,” “testament” and “interplay” — and nearly triple the use of a phrasing known as negative parallelisms (“it’s not X, it’s Y”), as well as a 63% increase in Oxford commas. But while AI labs continually try to tweak their models to generate more humanlike text, the people they’re mimicking aren’t staying the same.
Article continues below this ad
“Human writers might also be changing the way that they write in response to AI. I’ve seen some anecdotal evidence of people saying, ‘Now I avoid em dashes,’” Bestvater said. “If human writers are adjusting their choices of words and phrases and punctuation to try to differentiate their writing from AI-generated text, then you would see a pattern that looks sort of like that in the chart.”
The researchers collected monthly data from the nonprofit Common Crawl, spanning January 2021 through July 2026. Bestvater described the web archive “as close to a full scrape of the observable internet as they can get.”
Some trends, like em dashes, rose steeply starting in 2024, while others, like words apparently favored by chatbots, became more frequent on the internet starting in 2023, the analysis shows. Other words commonly appearing in AI writing include “align,” “showcase,” “valuable,” “crucial,” “underscore,” “intricate,” “meticulous,” “tapestry,” “landscape” and “pivotal.”
AI detectors don’t just look for a checklist of tells; instead, Bestvater said, they assess statistics involving word choice, sentence rhythm, punctuation distribution and the like, and then they output a probability that it’s AI-generated. He emphasized, though, that for one webpage or text, it can be challenging to figure out if it was written with AI, and there’s a risk of false positives, which is when someone mistakenly flags human-written text as AI.
“I get nervous when we’re trying to use a single score to make a consequential decision, like in a hiring or academic integrity case,” he said.
The researchers used the open- analysis with the commercial Pangram 3.3 detector and finding similar results. Pangram detects subtle differences between models, such as ChatGPT, Claude and Gemini, Pangram’s founding AI engineer Elyas Masrour said to SFGATE, in a statement. For example, while ChatGPT uses em dashes more than humans do, Gemini uses it less, he said
The study also comes as people are learning to make the distinction themselves. As AI-infused writing proliferates on the internet and beyond, 76% of Americans feel it is important to be able to distinguish between AI- and human-written content, Bestvater said, citing a Pew survey in September 2025. But only 12% of people feel confident making such distinctions themselves, the survey found.
About 10% of webpages now show signs of AI authorship, while more than a third of webpages published after the release of ChatGPT have them, Bestvater and his colleagues found. Signs of AI are the most common in .com websites, and they’re the least common in .gov and .edu websites. Pangram’s own research also showed in July that 41% of long-form content (more than 250 words) on LinkedIn is AI-written. Pangram generally needs 50 words or more to make a meaningful AI detection.
AI impacts
Some researchers have sought to understand not just how chatbots write text in particular nonhuman ways, but also why. Tom Juzek, a computational linguistics researcher at Florida State University, published a study in 2025 exploring why words like “delve” seemed to be so common.
Juzek said that as they’re trained and refined, ChatGPT and other large language models, or LLMs, apparently latch onto particular words that people approve of as they provide feedback on a model’s ability to perform complex tasks. That is, some human evaluators must themselves like the word “delve,” subtly encouraging the models to use it more.
In addition to word choices, LLMs have been prone to deploying certain kinds of phrases and sentence constructions more than people do, Juzek said. For example, recent research showed that LLMs use the following about twice as much as humans do: clause fragments like “But the observation suits him,” marker clauses like “…, and that’s a good thing,” and quantity noun-phrases like “many in Europe.” LLMs sometimes also have a penchant for lists, bold text and emojis, according to another analysis.
A 2-way street
Some researchers believe that the influence may be working in both directions, with people changing how they write and chatbots adjusting their outputs.
“You can see changes in people’s speech. People are tending to more commonly use these low-frequency words in their speech because they’re spending a lot of time listening to or reading AI-generated content,” said Roger Kreuz, a cognitive psychologist at the University of Memphis who recently wrote about how difficult it is for people to identify bot-written text.
Kreuz urged caution regarding the conclusions of the Pew study, since it’s possible that the trends Bestvater observed and the frequent use of certain words and phrases might not necessarily be influenced by AI-written websites. Instead, some could be written by people who frequently use chatbots like ChatGPT and who are, consciously or not, writing like the AI text they’ve been exposed to.
“It could be the case that the chatbots are shaping us as well,” he said.
Kreuz cited a recent study showing that, even after brief interactions with chatbots, people begin to adopt AI-like wording as their own. Another study came to similar conclusions, noting that people who heavily relied on a chatbot to help write an essay felt that the result was significantly less creative and less in their own voice.
Access our new app with ad-free reading, curated maps and more.
GET IT NOW
In any case, Juzek emphasized that buzzwords come and go. Masrour said Pangram is finding that, for new versions of chatbots, “delve” is now on its way out. “Early models really liked and overused the word delve. But as human preferences shifted, newer models started to use it less,” he said in an email to SFGATE.
Similarly, OpenAI may be getting the em dash problem “under control,” Juzek said, and the current flagship GPT model appears to use them less often. That would be good for human writers, who haven’t totally given up on those em dashes.
“It would be a real shame if we let the models dictate our language and influence our voice.”
— Husband of Central Coast teacher who died sues pool contractor
— Ex-child actor claims she was repeatedly raped during Disney filming
— Tonight’s lunar eclipse will be the best in California until 2029
— Ohio police criticize Jed York after prostitution sting arrest
