TLDR: A new study reveals that unscripted spoken English is showing a moderate but significant increase in words commonly associated with Large Language Models (LLMs) since ChatGPT’s release in 2022. This suggests that AI’s linguistic patterns might be influencing human language use, raising questions about natural language evolution versus AI-driven shifts and the “seep-in” effect of AI language into human communication.
Language is constantly evolving, a dynamic process shaped by social, cultural, and cognitive factors. Throughout history, technological advancements have played a significant role in this evolution, from the printing press standardizing written language to the telephone and internet influencing spoken communication. In recent years, a new force has emerged: Large Language Models (LLMs) like ChatGPT. These AI models have already brought about noticeable shifts in written language, particularly in academic and scientific fields, often characterized by a distinct lexical style. This divergence between AI output and human norms is often termed ‘misalignment’.
While many attribute these changes to the direct use of AI tools for text generation, a crucial question remains: do these shifts reflect a deeper change in human language itself? A recent research paper, Model Misalignment and Language Change: Traces of AI-Associated Language in Unscripted Spoken English, delves into this very question. The study, conducted by Bryce Anderson, Riley Galpin, and Tom S. Juzek from Florida State University, aimed to determine if human lexical choices in unscripted spoken language are increasingly converging towards the patterns characteristic of LLMs.
Investigating Unscripted Speech
To explore this, the researchers built a substantial dataset of 22.1 million words from unscripted spoken language, specifically from conversational science and technology podcasts. This focus on unscripted speech was a deliberate choice to minimize the presence of AI-generated content, which is increasingly prevalent in written texts. The dataset was divided into two periods: pre-2022 (mostly 2019-2021) and post-2022 (2023-2025), with the year 2022 excluded as it marked ChatGPT’s release and served as a transitional period.
The study focused on 20 specific words previously identified in literature as commonly overused by LLMs, such as “surpass,” “boast,” “meticulous,” “strategically,” “garner,” “intricate,” and “delve.” They analyzed the frequency of these words (occurrences per million) before and after 2022. As a baseline for comparison, they also tracked the usage of 87 common synonyms for these AI-associated words.
Key Findings: A Moderate Convergence
The results revealed a moderate yet statistically significant increase in the usage of the LLM-associated words in unscripted spoken English post-2022. Words like “surpass,” “boast,” “strategically,” “align,” and “significant” showed notable increases. Interestingly, some words frequently linked to LLM overuse in written contexts, such as “delve” and “intricate,” did not show a significant increase in spoken language, and words like “realm” and “crucial” actually decreased.
In stark contrast, the baseline synonym words showed no significant overall directional shift, with increases and decreases roughly balanced. This suggests that the observed uptick in AI-associated words is not merely a general trend in language but is specific to the vocabulary favored by LLMs.
Also Read:
- Evaluating Large Language Models’ Grasp of Nonbinary Pronouns
- Bridging the Cognitive Divide: Why AI’s Goals Differ from Human Intentions
Implications and Open Questions
The study’s findings suggest that human word choices are, to a moderate degree, converging with those of Large Language Models. While the effect in spoken language appears less pronounced than in written texts (where direct tool usage might be a stronger factor), this convergence in unscripted speech is significant. It raises the intriguing possibility of a “seep-in” effect, where AI linguistic patterns might be subconsciously adopted by humans through repeated exposure, even if they don’t reflect the speaker’s original preferences.
However, the question of causation remains open. It’s possible that LLMs are simply amplifying natural language trends that were already underway. Many of the words now considered overused by AI were already increasing in frequency prior to 2022. Regardless of the ultimate cause, the influence of machine-generated language on human communication is a significant development.
This research also highlights broader societal implications, particularly concerning model alignment. If AI’s linguistic preferences, potentially stemming from training procedures like Learning from Human Feedback, begin to shape human communication norms, it mirrors ethical concerns about misaligned models influencing social or moral beliefs. Furthermore, the rise of AI-generated content introduces the challenge of “human-authorship indeterminacy,” making it increasingly difficult to determine if a text was genuinely produced by a human, which has profound implications for linguistic research and other fields relying on language data as a proxy for human behavior.
In conclusion, while more extensive datasets and further analysis are needed, this study provides compelling early evidence that AI is not just a tool for language generation but may be subtly reshaping the very fabric of human spoken language. This marks a potentially profound shift in the trajectory of language evolution, prompting crucial discussions about the future of human-AI interaction and its linguistic consequences.
