Researchers fed 7.9 million speeches into AI—and what they found upends our understanding of language

When the meaning of a word shifts, do people of all ages follow the trend—or do younger generations lead while older speakers remain linguistically stuck in the past? A large-scale linguistic analysis published in the Proceedings of the National Academy of Sciences suggests that semantic change is more inclusive than previously believed. While younger individuals tend to be slightly quicker to adopt new meanings, older speakers typically follow within a few years, and in some cases, even lead the way.

This finding runs counter to a long-standing view in sociolinguistics that language evolves primarily through generational turnover. Instead, the results point to a more dynamic process in which speakers of all ages participate in real-time shifts in how words are used.

The researchers set out to test a foundational assumption in the study of language change: whether older individuals maintain stable linguistic patterns over their lives, or whether they update their language use in response to changes in the broader speech community.

For decades, sociolinguists have relied on the “apparent time” method, which compares the language of older and younger people at a single point in time to infer changes across generations. This method hinges on the idea that adult language use is relatively fixed. If, instead, older speakers are regularly adjusting to current trends, then these assumptions may not hold—particularly when it comes to how the meanings of words evolve.

Although prior research has largely supported the generational-change model, especially when it comes to pronunciation and grammatical structures, the question of whether word meanings follow the same pattern has remained relatively unexplored, especially at scale.

Free daily newsletter

“What led us to explore the topic was the fact that a simple question hadn’t really been answered yet — when words change meaning, do people of all ages follow?” said Gaurav Kamath, a PhD student in linguistics at McGill University and the lead author on the paper. “It’s an important question for language change more broadly, because (i) sociolinguists often assume that older speakers are a window into the past (which is true only if they DO NOT adopt changes), and (ii) it tells us something about our individual capacity to change how we speak, even as adults. Plus, language change is generally a fun, relatable thing to study.”

To address this gap, the research team analyzed more than 7.9 million U.S. Congressional speeches delivered between 1873 and 2010. These speeches were given by thousands of speakers whose ages were known at the time of each speech, providing a rare opportunity to track linguistic behavior over nearly 140 years while also controlling for speaker age.

The researchers focused on a set of approximately 100 words that were likely to have undergone meaning change during the 20th century. Examples include words like “monitor,” “articles,” “satellite,” and “outstanding.” Each of these words was examined for multiple possible meanings—referred to as “senses”—using advanced language models that predicted the context-based usage of each word. These predicted meanings were then grouped using clustering algorithms to identify distinct senses of each word.

For example, the word “articles” could refer to physical goods, legal provisions, or written stories. By analyzing the context in which the word appeared and modeling the rise or fall of each sense over time, the researchers could chart how meanings shifted across different time periods.

To determine whether age influenced the adoption of new meanings, the team used statistical models that predicted the likelihood a speaker would use a given word sense, based on both the year and the speaker’s age. These models estimated whether older speakers used outdated senses or whether they adopted newer senses at a slower or faster rate compared to their younger colleagues.

The researchers also performed a Bayesian meta-analysis to calculate an average age-related lag across all word senses. This allowed them to quantify just how much slower older speakers were to adopt new meanings, if at all.

Across the dataset, the researchers found that word meaning changes were overwhelmingly driven by a collective shift in usage across time rather than by generational replacement alone. While younger speakers tended to adopt newer meanings slightly earlier, older speakers were not far behind. On average, an older speaker lagged a younger speaker by about two to three years when it came to adopting a new word meaning.

In many cases, this lag was so minimal that older speakers could not be considered linguistically “behind.” For instance, an older member of Congress in the 1960s might use the newer sense of a word like “articles” only a few years after a younger colleague had already started doing so. In a minority of cases, older speakers actually led the shift—such as with the geopolitical sense of the word “satellite,” which gained prominence during the Cold War era.

“The main result, that older speakers are highly adaptable to new word meanings, was itself a surprise,” Kamath told PsyPost. “But the even bigger surprise was that for some of the words we looked at, we even found evidence of older speakers being the ones leading the change.”

The results provide evidence that meaning change tends to be a “zeitgeist” effect—a product of the cultural and temporal moment—rather than a strict generational handoff. Even at the individual level, speakers adjusted their usage over time. When examining a handful of prolific speakers who used the same word frequently across decades, the researchers observed noticeable within-person changes in how those words were used, tracking closely with broader shifts in usage patterns.

“In a nutshell, older people DO pick up new meanings of words,” Kamath explained. “Another way of putting it — this is evidence that your parents/grandparents are in fact capable of using words like “sick” (i.e. “cool”) or “model” (i.e. “AI model”) in their increasingly dominant new senses.”

These findings carry implications for how linguists model and interpret language change. If older speakers frequently adopt contemporary usages, then differences observed in cross-sectional data may not fully capture the speed or nature of ongoing change. In fact, apparent time comparisons may underestimate the extent of change already underway, as the linguistic behavior of older speakers quickly converges with that of younger ones.

The results also demonstrate the power of computational approaches to studying semantic change at scale. By leveraging large text corpora, speaker metadata, and advanced natural language processing models, researchers were able to draw conclusions that would be difficult to reach using smaller-scale observational studies.

“We think that this study shows the potential to use tools from Natural Language Processing (NLP) to study human language, and hope that it inspires further work that uses NLP tools for linguistic inquiry,” Kamath said.

But there are some limitations. The study focused exclusively on adult speakers, as membership in the U.S. Congress requires individuals to be at least 25 or 30 years old. Since teenagers and young adults are often the earliest adopters of linguistic innovation, this analysis may miss the very beginning of certain shifts in meaning.

The dataset also reflects a specific sociopolitical group—U.S. legislators—who tend to share certain demographic characteristics, especially in earlier decades. The results may not fully generalize to the broader population or to speakers outside the United States.

“The main limitation to keep in mind is that we looked at Congressional speeches,” Kamath said. “We relied on this genre of data because it was the only kind of data that allowed us to keep track of thousands of speakers’ ages over ~140 years. But the downside is that the speakers we studied (members of Congress) are not at all socially representative. Women and minorities are underrepresented, and just as importantly, our study did not include language from adolescents, who are typically at the forefront of language change.”

In addition, while the language models used in this study were generally effective at identifying distinct meanings, they are not infallible. Some errors in sense classification likely remain, particularly in cases where word usage is ambiguous or infrequent.

“The next steps would be to find a way to broaden the scope of this research, to address the limitations mentioned above,” Kamath said. “Can we expand beyond just North American English, and include a more balanced demographic sample? What about other languages and societies? And what about speech from adolescents?”

The study, “Semantic change in adults is not primarily a generational phenomenon,” was authored by Gaurav Kamath, Michelle Yang, Siva Reddy, Morgan Sonderegger, and Dallas Card.