Empirical evidence of Large Language Model's influence on human spoken communication

Imagine the way human culture spreads is like a giant, endless game of "telephone," but instead of whispering in a line, we are all shouting into a massive, global megaphone. For centuries, new ideas, slang, and ways of speaking have traveled through us via books, radio, and the internet. This process is called cultural evolution: the way our shared habits and language change over time as we learn from each other. Usually, this happens person-to-person. But recently, a new player joined the game: Artificial Intelligence. Specifically, Large Language Models (LLMs) like ChatGPT. These are computer programs trained on mountains of human writing that can talk back to us, write stories, and answer questions. The big question scientists are asking is: If we talk to these AI machines, do we start talking like them? Do the machines start to shape our culture, or are we just using them as tools without changing ourselves?

This paper investigates whether AI is quietly rewriting the way humans speak. The researchers found that the answer is a resounding "yes." They discovered that when people chat with AI, they start using the AI's favorite words, and they keep using them even after the conversation is over. It's as if the AI is a very persuasive friend who whispers a new slang term into your ear, and suddenly, you can't stop saying it.

The AI's "Accent" and the Podcast Detective Work

To understand what happened, we first need to know that AI has a distinct "accent." Just like some people love using fancy words like "utilize" instead of "use," AI models have their own favorite vocabulary. The researchers found that words like delve, showcase, boast, intricacies, and meticulous are used far more often by ChatGPT than by regular humans.

To see if this AI "accent" was leaking into human speech, the team acted like linguistic detectives. They didn't just look at what people wrote; they looked at what people said. They analyzed a massive library of 737,083 hours of podcast conversations—over 824,634 episodes—covering topics from science and tech to sports and business. They were very careful to filter out scripted shows (like news readings) and only kept the spontaneous, chatty ones where people were talking off the cuff.

They treated the release of ChatGPT in November 2022 as a natural experiment. They used a clever statistical trick called a "synthetic control" to build a "ghost version" of what podcast word usage would have looked like if ChatGPT had never been released. Then, they compared the real world to this ghost world.

The result was a sudden, sharp spike. Right after ChatGPT launched, the word delve (and other AI favorites) jumped up in spontaneous human speech. In the Science & Technology category, the use of "delve" rose to about 44% higher than the ghost version predicted. This wasn't just a few people; it was a massive shift across thousands of conversations. The study suggests that humans are absorbing the AI's vocabulary and using it in their own unscripted thoughts.

The "Magic Mirror" Experiment

To prove that this wasn't just a coincidence or that people were just copying what they read online, the researchers ran a controlled experiment with 496 people.

Imagine a game where you have to describe a picture to a partner so they can guess which one it is. In this experiment, the "partner" was a chatbot. The researchers secretly programmed the chatbot to always use one specific word from a pair of synonyms. For example, if the picture showed a cup, the chatbot would always call it a "mug." If it showed a repair job, the chatbot would always say "to fix" instead of "to repair."

The participants played 12 rounds of this game. Then, they did a 3-minute distraction task (solving math puzzles) to clear their minds. After that, they were asked to describe new pictures they had never seen before, with no chatbot present.

The result? The participants kept using the chatbot's words. Even though the chatbot was gone and they had been distracted, they were 36 percentage points more likely to use the AI's word (like "mug") instead of their own usual word (like "cup"). Even more surprisingly, when asked to choose between the two words later, 63% of them picked the word the chatbot had used.

This suggests that the change wasn't just temporary mimicry. The AI's words had "entrenched" themselves in the participants' brains. It's like the AI tapped into their memory and planted a new habit that stuck, even after the interaction ended.

The Twist: The AI Changed Its Mind, and So Did We

Here is the most fascinating part of the story. The study didn't just show that people adopted AI words; it showed that the relationship is a two-way street.

The researchers noticed that the popularity of words like "delve" didn't stay high forever. After peaking in mid-2024, the usage of these words started to drop, and in some cases, fell below the original baseline. Why? Because people started realizing that these words sounded "robotic." Once the public realized that "delve" was a signature AI word, humans started avoiding it to sound more authentic.

At the same time, the AI companies themselves noticed this. Newer versions of the AI models (like GPT-4o and GPT-5) were updated to stop using "delve" so much. The AI changed its vocabulary, and the humans followed suit, dropping the word again.

This creates a loop: The AI influences humans, humans influence the AI (by training it on their new speech), and the AI changes again, which changes human speech again. The researchers call this a "coupled human-machine cultural process."

What This Means for Us

The paper concludes that we are no longer just talking to machines; we are talking with them in a way that changes who we are. The AI is becoming a "cultural model," a source of language that we learn from, just like we learn from our parents or teachers.

The study suggests that this could lead to a "homogenization" of language, where everyone starts sounding the same because we are all learning from the same few AI systems. It raises a big question: If a few companies control the AI that shapes our vocabulary, do they also have the power to shape our culture? The researchers found that this influence happens quietly, below the level of our conscious awareness. We don't realize we are adopting the AI's style; we just find ourselves saying "delve" instead of "explore," and we keep saying it long after the chatbot has stopped talking.

In short, the machines we built to help us communicate are now teaching us how to speak, and we are listening.