University of Vermont Researchers Discover a Hidden Safety Bias in Human Language That Upends a 70-Year-Old Scientific Framework

For more than seven decades, the scientific community has relied on a foundational psychological framework to understand how humans encode and interpret meaning through language. Developed in the 1950s by pioneering psychologist Charles Osgood, the VAD model—which evaluates words across three emotional dimensions of valence (positive versus negative), arousal (excited versus calm), and dominance (controlling versus submissive)—has long served as the bedrock for linguistics, behavioral psychology, and artificial intelligence sentiment analysis. However, a groundbreaking study conducted by a multidisciplinary team of researchers at the University of Vermont (UVM) has challenged this deeply entrenched consensus. Published in the journal Science Advances, the new research reveals that human language is structured not primarily by emotion, but by three more fundamental dimensions: power, danger, and order.

The study introduces an innovative computational approach dubbed "ousiometrics"—derived from "ouisa," the Ancient Greek root for "essence"—which measures the essential meaning of language using advanced mathematical methods. By analyzing billions of individual word uses across a vast corpus of real-world texts, the Vermont researchers uncovered a striking, previously unrecognized pattern: human communication demonstrates a pervasive, overwhelming preference for safety over danger. This hidden safety bias not only redefines how scientists view the architecture of communication but also offers a radical new explanation for long-standing linguistic phenomena, potentially transforming the development of artificial intelligence, psychological research, and automated content moderation.

A Decades-Old Paradigm Shift in Linguistics and Psychology

To appreciate the magnitude of the University of Vermont’s discovery, one must examine the historical context of linguistic and psychological modeling. Since the mid-20th century, Charles Osgood’s VAD model dominated academic disciplines by asserting that human emotional responses and semantic evaluations could be adequately mapped onto valence, arousal, and dominance scales. This triad became the standard lens through which researchers analyzed everything from clinical psychiatric evaluations to automated natural language processing (NLP) algorithms.

Yet, cracks in this seventy-year-old assumption became apparent as computational power advanced. Led by Peter Dodds, director of UVM’s Complex Systems Institute, and Chris Danforth, professors in the university’s College of Engineering and Mathematical Sciences, a ten-member research team set out to rigorously test the independence of the VAD model’s three dimensions. Collaborating with institutions including the Santa Fe Institute, the Complexity Science Hub in Austria, the Howard Hughes Medical Center, the University of California, Berkeley, the University of Adelaide, and MassMutual Data Science, the team embarked on an unprecedented computational analysis.

Rather than treating words as isolated dictionary entries, the researchers examined billions of word tokens—every single instance a word appeared in context—across a diverse array of written and spoken materials. Their datasets spanned centuries of literature, including Jane Austen’s novels and Arthur Conan Doyle’s Sherlock Holmes mysteries, alongside contemporary sources such as the New York Times, Wikipedia, Twitter feeds, and transcripts of talk radio broadcasts.

By applying advanced mathematical modeling to this vast linguistic corpus, the researchers discovered that the three traditional dimensions of the VAD model are not, in fact, independent. Instead, word meaning is more accurately and comprehensively captured by three distinct axes: power (weak versus powerful), danger (safe versus dangerous), and structure (ordered versus chaotic). While the traditional VAD framework accounted for approximately 72% of the variation in measured word meaning, the newly proposed framework accounted for more than 90%, representing a massive leap in descriptive accuracy and statistical robustness.

Unmasking the Safety Bias and Reinterpreting the Pollyanna Principle

One of the most profound revelations of the UVM study emerged when the researchers shifted their focus from word categories ("types") to actual usage frequency ("tokens"). While earlier linguistic analyses frequently assigned equal weight to every word in a dictionary, the Vermont team incorporated frequency of use into their algorithms. This methodological choice unmasked a powerful, hidden preference embedded within everyday human communication.

Across books, journalism, social media platforms, and spoken dialogues, words associated with safety appeared with exponentially greater frequency than those associated with danger. This discovery provides a radically revised explanation for the Pollyanna principle—a well-documented psychological phenomenon describing the human tendency to use positive words more frequently than negative ones.

For generations, scientists interpreted the Pollyanna principle as direct evidence of an innate human bias toward positive emotional expression, happiness, and optimism. However, the ousiometric analysis suggests that emotional positivity is merely a byproduct of a much deeper, more primal imperative. According to the study’s findings, the apparent positivity bias in human language is, in fact, a one-dimensional projection of an underlying safety bias.

"Expressions of safety are crucial to all language," noted Peter Dodds, the study’s senior author. Far from simply expressing joy or approval, many common positive words function primarily to convey security, stability, and predictability. Julia Zimmerman, a postdoctoral researcher in UVM’s Computational Story Lab and co-author of the study, emphasized the universal relevance of these newly isolated dimensions. "Power, danger, and structure," Zimmerman stated, "are relevant to every person that’s ever lived."

Evolutionary Pressures and the Architecture of Survival

While the study’s empirical findings are robust, the researchers exercise cautious scientific restraint regarding causation, noting that their results do not definitively prove that evolutionary pressures directly engineered human language. Nevertheless, the correlation between linguistic structure and evolutionary survival is an irresistible and compelling avenue of inquiry.

From an evolutionary standpoint, the human brain is hyper-vigilant, finely tuned to detect threats, evaluate uncertainty, establish trust within social groups, and coordinate cooperative actions. Language evolved not merely as an artistic medium for expressing abstract feelings, but as a vital survival tool. Communicating whether a physical environment, a social situation, or a potential food source is safe or dangerous can mean the difference between life and death.

The UVM findings suggest that this foundational biological priority has indelibly shaped the semantic architecture of human speech. Throughout daily life, humans constantly exchange information regarding threat levels, structural order, and relative power dynamics. By demonstrating that language is fundamentally organized around safety, power, and structure rather than abstract emotional states, the research aligns human linguistics more closely with evolutionary biology and neurobiology. Neuroscientists have long understood that the human brain prioritizes threat detection and safety signaling; this study provides the first large-scale linguistic evidence that these biological imperatives are mirrored in the very fabric of our vocabulary.

Methodological Innovation: The Birth of the Ousiometer

To map these complex linguistic landscapes, the research team developed a sophisticated analytical instrument they named the "ousiometer." Building upon their previously invented "hedonometer"—a tool designed to measure the macroscopic emotional happiness score of large text bodies—the ousiometer evaluates the essential meaning of extensive collections of written and spoken language.

The utility of the ousiometer was demonstrated through its application to radically different textual genres. The research team tracked the shifting semantic trajectories of complex narratives, such as an English translation of Victor Hugo’s monumental novel Les Misérables. By charting what the researchers termed an "ousiometric trajectory," they were able to visualize how the underlying meaning of a story shifts across narrative arcs, moving along a multidimensional grid defined by four contrasting pairs: dangerous/safe, weak/powerful, gentle/aggressive, and bad/good. The resulting data visualizations resemble complex, multicolored protein structures, offering literary scholars and computational scientists a novel window into narrative progression and thematic evolution.

Furthermore, the team’s strict methodological distinction between word types and word tokens proved essential. By accounting for how frequently specific words are deployed in real-world communication, the researchers avoided the analytical distortions that have historically obscured underlying linguistic structures. Words that appear millions of times in daily discourse exert a disproportionate influence on human culture and cognition; factoring this reality into the ousiometer’s algorithms was the key to unlocking the hidden safety bias.

Broader Implications for Artificial Intelligence, Psychology, and Society

The implications of this study extend far beyond theoretical linguistics, promising disruptive impacts across several applied scientific and technological fields.

First and foremost is the field of artificial intelligence and natural language processing. Modern large language models (LLMs), automated content moderation systems, and sentiment analysis tools rely heavily on algorithms designed to gauge emotional tone, often utilizing legacy frameworks like the VAD model. If these AI systems are built upon an incomplete or flawed understanding of semantic structure—mistaking safety biases for pure emotional valence—they risk misinterpreting human intent, nuance, and risk assessment. Integrating the dimensions of power, danger, and structure into AI architecture could yield systems that interpret human communication with unprecedented accuracy, providing deeper transparency into how machines assess trust, risk, and decision-making.

In the realm of psychology, the findings invite a critical re-evaluation of decades of behavioral research that relied upon the VAD model to study human perception, emotional regulation, and social behavior. If fundamental human motivations are more accurately framed through the lenses of power dynamics, structural order, and safety assurance, psychological models of emotion and motivation may require revision.

In public spheres, recognizing the pervasive safety bias in language could transform methodologies in political communication, public health messaging, and media studies. Understanding how narratives frame safety versus danger can illuminate how information spreads, how public fear is mobilized, and how societal consensus is built.

Conclusion: A Fundamental Reimagining of Human Communication

Supported by grants from the U.S. National Science Foundation, Google, MassMutual, and other funding bodies, the University of Vermont study represents a watershed moment in the scientific study of human language. By dismantling an accepted paradigm that has stood for over seventy years, Dodds, Danforth, and their international colleagues have offered a profound new lens through which to view human communication.

Human language, far from being a mere reflection of fleeting emotional states, emerges from this research as a sophisticated, evolutionarily resonant map of our deepest existential priorities. Structured around the fundamental axes of power, danger, and order, and anchored by a pervasive, underlying preference for safety, our vocabulary reflects the enduring human struggle to navigate an uncertain world, establish communal trust, and ensure survival. As researchers, computer scientists, and psychologists begin to integrate these insights, our understanding of what words mean—and why we use them—will be permanently transformed.