ChatGPT could also be altering the way you’re writing

From “rizz” to “meat proxy,” phrases and phrases have fallen out and in of favor for hundreds of years. Culture, politics and technological growth have lengthy influenced these tendencies, and now, new research shows that ChatGPT is starting to affect how people talk.

The findings from a staff on the Pew Research Center present how, in a short while since OpenAI launched ChatGPT in November 2022 and different chatbots adopted, the usage of sure phrases and phrases has spiked on the internet. The researchers, led by knowledge scientist Samuel Bestvater, studied almost half one million English language webpage texts in snapshots over the previous 5 years. Not solely has AI-written textual content develop into extra frequent, however it is also influencing how folks write, reminiscent of attempting to keep away from utilizing the “tells” which have develop into related to AI-generated textual content.

Bestvater and his staff detected double the variety of em dashes, double the frequency of specific phrases — reminiscent of “delve,” “testomony” and “interaction” — and almost triple the usage of a phrasing often called adverse parallelisms (“it’s not X, it’s Y”), in addition to a 63% enhance in Oxford commas. But whereas AI labs regularly attempt to tweak their fashions to generate extra humanlike textual content, the folks they’re mimicking aren’t staying the identical. 

“Human writers may also be altering the way in which that they write in response to AI. I’ve seen some anecdotal proof of individuals saying, ‘Now I keep away from em dashes,’” Bestvater mentioned. “If human writers are adjusting their decisions of phrases and phrases and punctuation to attempt to differentiate their writing from AI-generated textual content, then you definately would see a sample that appears type of like that within the chart.”

The researchers collected month-to-month knowledge from the nonprofit Common Crawl, spanning January 2021 by July 2026. Bestvater described the net archive “as near a full scrape of the observable web as they will get.” 

Some tendencies, like em dashes, rose steeply beginning in 2024, whereas others, like phrases apparently favored by chatbots, grew to become extra frequent on the web beginning in 2023, the evaluation exhibits. Other phrases generally showing in AI writing embrace “align,” “showcase,” “precious,” “essential,” “underscore,” “intricate,” “meticulous,” “tapestry,” “panorama” and “pivotal.” 

AI detectors don’t simply search for a guidelines of tells; as an alternative, Bestvater mentioned, they assess statistics involving phrase alternative, sentence rhythm, punctuation distribution and the like, after which they output a chance that it’s AI-generated. He emphasised, although, that for one webpage or textual content, it may be difficult to determine if it was written with AI, and there’s a danger of false positives, which is when somebody mistakenly flags human-written textual content as AI. 

“I get nervous once we’re attempting to make use of a single rating to make a consequential resolution, like in a hiring or educational integrity case,” he mentioned. 

The researchers used the open-source model of Pangram Labs’ AI detector, checking a few of their evaluation with the industrial Pangram 3.3 detector and discovering related outcomes. Pangram detects refined variations between fashions, reminiscent of ChatGPT, Claude and Gemini, Pangram’s founding AI engineer Elyas Masrour mentioned to SFGATE, in a press release. For instance, whereas ChatGPT makes use of em dashes greater than people do, Gemini makes use of it much less, he mentioned.

The research additionally comes as individuals are studying to make the excellence themselves. As AI-infused writing proliferates on the web and past, 76% of Americans really feel you will need to be capable to distinguish between AI- and human-written content material, Bestvater mentioned, citing a Pew survey in September 2025. But solely 12% of individuals really feel assured making such distinctions themselves, the survey discovered.

About 10% of webpages now present indicators of AI authorship, whereas greater than a 3rd of webpages revealed after the discharge of ChatGPT have them, Bestvater and his colleagues discovered. Signs of AI are the most typical in .com web sites, they usually’re the least frequent in .gov and .edu web sites. Pangram’s own research additionally confirmed in July that 41% of long-form content material (greater than 250 phrases) on LinkedIn is AI-written. Pangram typically wants 50 words or extra to make a significant AI detection. 

Some researchers have sought to grasp not simply how chatbots write textual content particularly nonhuman methods, but in addition why. Tom Juzek, a computational linguistics researcher at Florida State University, revealed a research in 2025 exploring why phrases like “delve” gave the impression to be so frequent. 

Juzek mentioned that as they’re skilled and refined, ChatGPT and different giant language fashions, or LLMs, apparently latch onto specific phrases that folks approve of as they supply suggestions on a mannequin’s capability to carry out advanced duties. That is, some human evaluators should themselves just like the phrase “delve,” subtly encouraging the fashions to make use of it extra.

In addition to phrase decisions, LLMs have been susceptible to deploying sure sorts of phrases and sentence constructions greater than folks do, Juzek mentioned. For instance, recent research confirmed that LLMs use the next about twice as a lot as people do: clause fragments like “But the remark fits him,” marker clauses like “…, and that’s a very good factor,” and amount noun-phrases like “many in Europe.” LLMs generally even have a penchant for lists, daring textual content and emojis, in response to another analysis.

Some researchers imagine that the affect could also be working in each instructions, with folks altering how they write and chatbots adjusting their outputs. 

“You can see adjustments in folks’s speech. People are tending to extra generally use these low-frequency phrases of their speech as a result of they’re spending lots of time listening to or studying AI-generated content material,” mentioned Roger Kreuz, a cognitive psychologist on the University of Memphis who recently wrote about how tough it’s for folks to determine bot-written textual content. 

Kreuz urged warning relating to the conclusions of the Pew research, because it’s doable that the tendencies Bestvater noticed and the frequent use of sure phrases and phrases won’t essentially be influenced by AI-written web sites. Instead, some might be written by individuals who steadily use chatbots like ChatGPT and who’re, consciously or not, writing just like the AI textual content they’ve been uncovered to. 

“It might be the case that the chatbots are shaping us as effectively,” he mentioned.

Kreuz cited a recent research displaying that, even after transient interactions with chatbots, folks start to undertake AI-like wording as their very own. Another study got here to related conclusions, noting that individuals who closely relied on a chatbot to assist write an essay felt that the consequence was considerably much less artistic and fewer in their very own voice.

In any case, Juzek emphasised that buzzwords come and go. Masrour mentioned Pangram is discovering that, for brand new variations of chatbots, “delve” is now on its method out. “Early fashions actually favored and overused the phrase delve. But as human preferences shifted, newer fashions began to make use of it much less,” he mentioned in an e-mail to SFGATE. 

Similarly, OpenAI could also be getting the em sprint drawback “underneath management,” Juzek mentioned, and the present flagship GPT mannequin seems to make use of them much less usually. That can be good for human writers, who haven’t completely given up on these em dashes. 

“It can be an actual disgrace if we let the fashions dictate our language and affect our voice.”