The endless scroll of social media sometimes feels like a cognitive gym where words flex and vocabulary grows heavier with every new comment. Among the din of memes, debates, and breaking news, a quiet pattern starts to emerge: the more people invest themselves in a conversation, the more their words seem to swell in complexity. A team of researchers from several Italian universities and NYU Abu Dhabi set out to listen closely to that pattern, not by counting likes or shares but by analyzing the language itself. Their study focuses on language produced by influencers on Twitter across three globally charged topics COVID-19, the COP26 climate summit, and the Russia Ukraine war. The aim isn’t to pick winners or losers in a political battle but to understand how involvement reshapes the very fabric of online speech. The work is a collaboration among the University of Padova and Sapienza University of Rome, NYU Abu Dhabi, Ca’ Foscari University of Venice, and other partners, led by the paper’s first author Eleonora Amadori from Padova.
Think of language as a social instrument: the same sentence can cut like a blade in one hand and be a lullaby in another. The researchers wanted to know what happens to that instrument when people pour themselves into online debates. They gathered a vast trove of tweets from thousands of influencers, spanning three different arenas that have shaped public discourse in recent years. Their question was deceptively simple: does the way we talk change when we’re more involved, more invested, or more exposed to a topic that carries strong opinions and credible or questionable information? The answer, delivered through a blend of linguistic metrics and network analysis, is that involvement does not just shift tone; it reconfigures vocabulary, structure, and the communities that share certain linguistic habits.
The language of involvement
In three separate topics COVID-19, COP26 and Ukraine, the researchers tracked around 1.36 million tweets by more than 3,000 influencers. They labeled each profile for political leaning on a Left Center Right spectrum and for factual reliability as Reliable or Questionable. They then read the texts through three lenses of linguistic complexity: vocabulary richness, repetitiveness, and readability. It’s a triptych designed to capture different faces of complexity without collapsing them into a single number. The first striking pattern is straightforward on the surface but rich in implication: individual accounts tend to use more complex language than organizations. In other words, when a person speaks for themselves, their sentences and word choices tend to carry more weight, novelty, and nuance than the more regulated or bureaucratic voice of an institution or media outlet.
Beyond that baseline, the data reveal a subtler map. Across the datasets, centers of political gravity—profiles without an extreme partisan tilt—display higher lexical complexity than their more polarized peers. Left and Right voices tended to bundle their messages into shorter, perhaps more conventional phraseology, while Center voices wandered a bit more widely in their vocabulary. The effect is not universal across every topic, but COVID-19 shows the clearest signal: the more a profile leans toward a defined stance, the more likely its language harbors a larger vocabulary and a more varied syntax. The researchers stress that this is a trend, not a universal law; context matters, and the topic’s nature—whether it’s a science-heavy debate or a political one—modulates how language complexity colors the conversation.
What about reliability and negativity? Here the surprises pile up. Profiles flagged as questionable in terms of reliability tended to use more lexically complex language, at least in the COVID-19 dataset. And across topics, users who produced more negative or offensive content also tended to deploy a richer vocabulary. In the moment of tension, rhetoric thickens. The authors describe a tendency for audiences and content to converge toward a shared jargon when aligned on a political stance or a reliability profile, suggesting that debated topics don’t just polarize opinions; they corral language into distinctive regional dialects of online discourse.
Measuring complexity in the digital language
Measuring how complex a tweet is might sound like a job for a literary professor with a red pen, but the researchers built a practical, multi-faceted toolkit. They used three complementary metrics: a lexical richness score to capture how varied the vocabulary is; a compression-based measure that gauges how much text compresses when you remove redundancy; and a readability score that estimates how easy the text is to read. In a digital environment where micro-sentences and emojis abound, these measures help separate the density of ideas from mere length. The pre-processing step is crucial: they strip emojis and hashtags, reduce words to roots, and normalize text to make fair comparisons across accounts with wildly different styles.
To understand who is speaking and with what weight they bring to a discussion, the study relies on automated labeling to assign political leaning and reliability. The authors validate these labels against an external benchmark so the results don’t drift into the fog of subjective interpretation. They also classify sentiment and the presence of offensive language with classifiers trained on Twitter data, then roll these signals up to the user level to see how an individual’s overall negativity or offensiveness relates to language complexity. While no automated system is perfect, the researchers report substantial agreement with established benchmarks, lending credibility to the surprisingly persistent patterns they uncover.
A second methodological pillar is the network analysis. Influencers don’t speak in a vacuum, and the researchers model the landscape as a bipartite network that links influencers to the types of words they use. They then project this bipartite graph onto the influencer layer in a way that corrects for biases—especially the tendency of louder accounts to appear more similar purely by virtue of activity level. The result is a refined map of which influencers share a meaningful vocabulary, filtered so that the connections aren’t just noise. When they detect communities within this network, the communities tend to align with political leaning and reliability, revealing how language acts as a social bond among like-minded voices.
A social map of online influence
When you glance at the networks the researchers uncover, a familiar pattern emerges: language clusters tend to cluster people. Communities with similar political leanings or reliability profiles tend to use similar word choices, and those clusters hold together more tightly in topics that lean on scientific discourse, like COVID-19 and COP26. The Ukraine dataset shows a looser linguistic seam, perhaps reflecting the more politically charged and emotionally charged nature of that debate. Still, across all three events, the strongest backbone of connection runs along the axis of involvement. The more a profile is engaged, the more its language diverges from the ordinary and coalesces with others who share a similar stance or credibility profile.
Among the structures the team mapped, some patterns are especially resonant. The relationship between complexity and negativity suggests that sharper, more biting language isn’t necessarily shorter or simpler. In fact, people who post more offensive or negative content often do so with a higher lexical load, weaving more nuanced constructions into their messages. This isn’t a simple good-vs-bad story; it challenges a common intuition that more heated debates produce simpler, more stripped-down language. The data imply that heated, ideologically charged conversations may be the very spaces where language becomes more elaborate, not less.
Another striking feature is topic dependence. The COVID-19 and COP26 analyses show clearer modular structures tied to political and reliability dimensions, signaling that scientific or policy-oriented debates may sharpen language in a way social-political conversations do not. That doesn’t mean the Ukraine debate is less important; it means its linguistic ecology differs, possibly because the conversation includes more personal, identity-based dimensions and a wider variety of information sources. The researchers interpret this as evidence that scientific topics tend to crystallize language into more distinct ecological niches, whereas more purely political or conflict-related topics yield a more dispersed linguistic landscape.
What this means for public discourse
If language is a fingerprint of involvement, then online debates reveal more than opinions; they reveal the social mechanics of engagement. The study suggests that the most linguistically ambitious speakers are often those who are most involved, whether that involvement comes from personal identity, political conviction, or a sense of responsibility for the information they share. In practical terms, this means complexity can be a signal rather than a judgment. A tweet that packs a punch with varied vocabulary and intricate sentence structure might reflect a highly engaged voice trying to persuade, explain, or defend a viewpoint with care and precision. This challenges the old dichotomy that simple language equals clarity and complex language equals opacity. In the wild world of online discourse, complexity can simultaneously signal nuance and risk.
The findings also carry a note of caution for the design of digital platforms and for media literacy efforts. If mis/disinformation campaigns can ride on language that feels credible because it is lexically rich and rhetorically dense, then moderation and education must account for this complexity. Simple heuristics that equate brevity with truth or hostility with negativity can miss a richer, subtler reality. The study invites us to cultivate a more nuanced literacy, where readers train themselves to parse not just what is being said but how it is being said and who is saying it. Language becomes a map of social dynamics, and reading it well requires attention to the same factors that shape a public debate in the first place: involvement, credibility, and the political currents that carry words from mouth to screen to mind.
Yet the research doesn’t advocate surrender to a doom loop of ever-increasing linguistic complexity. Instead it provides a framework for understanding how influence works in the modern information ecosystem. By shedding light on how language diverges across account types, political stances, and reliability signals, the study offers a catalog of motifs that platforms, educators, journalists, and researchers can study to better understand how ideas spread, how communities form, and how public discourse can be steered toward more constructive exchange. The language isn’t just decorative; it’s a running log of how people choose to connect with one another in real time, under pressure from events that shape the global agenda.
As a closing thought, this work reminds us that online language is a living artifact of human involvement. The more a topic becomes a shared concern, the more the vocabulary evolves into a social instrument that can cut as deeply as it can clarify. Language does not simply reflect what we think; it reveals how deeply we care, how carefully we argue, and how keenly we participate in a public life that increasingly unfolds in a global space of tweets, threads, and debates that never truly end.
