Khoury News
ChatGPT is adapting its speaking style to you. You may be adapting to it, too.
Many AI users are familiar with sycophancy, where chatbots fudge the truth in favor of what they think the user wants to hear. But new research indicates that AI isn't just inclined to agree with you — it can start to talk like you, too.
When people talk with one another, they often unconsciously begin to sound a little more alike. They pick up each other’s vocabulary, sentence patterns, and even small functional words such as prepositions and conjunctions.
Linguists call this “linguistic accommodation” — the tendency for speakers to adapt to their language based on the person they are talking to. One form, called convergence, occurs when those linguistic patterns become increasingly similar over the course of a conversation or group of conversations.
Khoury Assistant Professor Terra Blevins has found that the same phenomenon happens when the other side of the conversation is with a large language model, or LLM. In her new paper, “Accommodation Goes Both Ways: Studying Linguistic Convergence Between Humans and Language Models,” Blevins analyzed thousands of ChatGPT conversations to examine how users and the chatbot adapt to each other.
The central finding: ChatGPT adapts to its users far more than users adapt to ChatGPT. Humans do accommodate the chatbot linguistically, but they do so about as much as they would for any person that they speak to.
For Blevins, the project fits into her broader research goal — to determine whether language models actually use language in the same way humans do.
“Instead of just assuming that models are good at language if they use it the way a human would, are there differences in the way they use languages?” Blevins said. “And how does that come out?”
For her most recent study, she turned to WildChat, a composite of roughly one million conversations between people and ChatGPT. She sampled conversations in eight languages — English, French, Spanish, Portuguese, Italian, Russian, Turkish, and Chinese — and compared them to recorded human-to-human conversations.
The differences were substantial. In English conversations, the model’s average convergence on functional words, such as prepositions, conjunctions, and pronouns, was nearly twice the users’ rate. On nouns, the difference was even larger; ChatGPT’s convergence scores were nearly three times those of human users.
“It was kind of surprising because not only did the model exhibit what we would characterize as accommodation if a human had generated the language, it actually over-accommodated,” Blevins said. “It was over-fitting to the conversation that it was supposed to respond to.”
Blevins initially wondered whether people might treat the chatbot differently from a human because they know it is not human. Instead, she found that users accommodated ChatGPT much as they would other humans; knowing that there was a chatbot on the other side didn’t change anything.
Blevins thinks this may have less to do with what people believe ChatGPT is and more to do with how fluently the AI communicates. Linguistic accommodation is largely unconscious, so if modern language models produce sufficiently coherent, conversational language, humans will perceive them as human enough, and the processes humans use with other people will kick in.
That possibility matters because chatbot conversations don’t necessarily end when the chat window closes. Blevins points to familiar examples of AI-associated language habits — including words such as “delve” or the use of em dashes — that people may encounter in model-generated text. If humans unconsciously accommodate the language of a chatbot during conversations, she says, it raises a larger question: Could repeated exposure eventually influence how people communicate outside those conversations?
“Seeing that, at least on the conversational level, we will adopt the patterns of these models has a lot of implications for what we will also incorporate into our language longer term,” Blevins said.
Blevins’ next step is a more controlled experiment: Have people produce language before and after interacting with a chatbot, then test them again later — perhaps a week afterward — to see whether any changes persist. That would help separate momentary conversational accommodation from longer-term linguistic change.
For everyday chatbot users, Blevins says, that distinction is important. There’s no evidence yet that talking to ChatGPT will permanently rewrite your vocabulary or prose style. But it does offer a reason to pay closer attention to those interactions. Because if humans naturally adjust their language to conversational partners, and AI systems are now among those partners, then the language of the machines may become part of the linguistic environment we unconsciously adapt to.
The Khoury Network: Be in the know
Subscribe now to our monthly newsletter for the latest stories and achievements of our students and faculty