For more than seven decades, the scientific community has relied on a foundational model to decode the human experience through the lens of linguistics. Since the 1950s, the VAD model—centered on valence, arousal, and dominance—has served as the gold standard for psychologists and linguists attempting to quantify the emotional weight of our vocabulary. However, a groundbreaking study conducted by researchers at the University of Vermont (UVM) has challenged this long-standing paradigm, suggesting that the architecture of human language is built upon a more primitive, survival-oriented framework: the prioritization of safety over danger.
The study, published in the journal Science Advances, introduces a methodology dubbed "ousiometrics"—derived from the Greek word ousia, meaning essence. By employing advanced computational techniques to analyze the essential meaning of language, the UVM team has revealed that our words are not merely vessels for emotional expression, but are instead structured around three fundamental dimensions: power, danger, and order.
The Evolution of Linguistic Theory
The VAD model was pioneered by psychologist Charles Osgood in the mid-20th century. It posits that human communication can be mapped across three distinct emotional axes: valence (positive vs. negative), arousal (excited vs. calm), and dominance (controlling vs. submissive). This framework has been the bedrock for sentiment analysis in modern artificial intelligence, helping software interpret the emotional tone of everything from social media posts to corporate communications.
However, the UVM researchers hypothesized that the VAD model failed to capture the full breadth of human communication. By examining billions of individual word uses across more than 20,000 unique terms, the team sought to determine if these emotional dimensions were truly independent. Their analysis revealed a startling conclusion: the VAD dimensions are deeply intertwined, and when re-examined through the lenses of power (weak vs. powerful), danger (safe vs. dangerous), and structure (ordered vs. chaotic), the predictive power of the linguistic model increases significantly. While the traditional VAD approach explains approximately 72% of the variation in word meaning, the new ousiometric framework accounts for more than 90%.
Chronology and Methodology
The research project, which received support from the US National Science Foundation, Google, and MassMutual, represents a multi-year effort to modernize the computational study of language. The team’s approach was rooted in the distinction between "types" (the categorical identity of a word) and "tokens" (the frequency with which a word appears in real-world usage).
Previous studies often assigned equal weight to every word in a dictionary, regardless of how rarely they appeared in daily conversation. The UVM team, led by Peter Dodds and Chris Danforth of the UVM College of Engineering and Mathematical Sciences, argued that this approach skewed the results. By incorporating word frequency—essentially weighting words by their actual impact on human discourse—the researchers uncovered a consistent, hidden bias: humans disproportionately favor language associated with safety.
This discovery provides a scientific pivot point for the "Pollyanna principle," a long-observed phenomenon where human language tends to skew toward positive, rather than negative, sentiment. For decades, this was interpreted as an innate human optimism. The new findings suggest that this "positivity bias" is, in fact, a one-dimensional shadow of a much deeper, more fundamental "safety bias."
The Survival Imperative
The implications of the safety bias extend into the realm of evolutionary biology. If human language is indeed structured around the necessity of identifying threats and maintaining stability, it suggests that our vocabulary evolved as a tool for survival. Words do more than describe feelings; they function as a navigation system for the environment. By signaling safety, predictability, and control, language allows individuals to coordinate, establish trust, and avoid danger.
"Expressions of safety are crucial to all language," noted Peter Dodds, the senior author of the study. The research suggests that when we describe something as "good," we are often unconsciously signaling that it is "safe" or "stable." This interpretation shifts the understanding of human interaction: we are not just creatures who seek happiness, but beings whose primary linguistic output is designed to signal that we are in a secure environment.
Implications for Artificial Intelligence
The tech industry is perhaps the most immediate stakeholder in this discovery. As large language models (LLMs) and automated moderation tools become more integrated into the fabric of daily life, their ability to accurately parse human intent is paramount. Most current AI systems are trained on sentiment analysis models that rely on the older VAD framework.
If these models are missing the "safety-danger" dimension, they may be misinterpreting the underlying intent of human communication. By integrating the ousiometric framework, developers could build AI that better understands the nuance of trust, risk, and stability. This is especially critical in fields such as mental health diagnostics, where AI might be used to monitor for signs of distress, or in automated news moderation, where identifying the difference between a neutral report and a veiled threat is essential.
Measuring the Essence of Storytelling
To prove the utility of their framework, the researchers developed the "ousiometer," a digital tool capable of mapping the "ousiometric trajectory" of any text. Unlike the team’s earlier "hedonometer," which tracks the happiness index of a text, the ousiometer maps how the core meaning of a narrative shifts over time.
In a test case, the researchers applied the tool to Victor Hugo’s Les Misérables. The resulting visualization—a complex, multi-colored structure resembling a protein—mapped the novel’s narrative arc across the four contrasting pairs of dangerous/safe, weak/powerful, gentle/aggressive, and bad/good. This allows researchers to see, in granular detail, exactly how a story’s underlying sense of security or peril changes, offering a new, objective lens through which to analyze literature and historical documents.
Broader Impact and Future Research
The interdisciplinary nature of this study—involving experts from the Santa Fe Institute, the Complexity Science Hub in Austria, and the University of California, Berkeley—underscores the gravity of these findings. The research challenges the traditional boundaries of psychology, linguistics, and neuroscience.
Neuroscientists, in particular, may find the results compelling. Given the human brain’s heightened sensitivity to threats—often described as the "negativity bias" in neurobiology—the presence of a corresponding linguistic "safety bias" suggests a profound alignment between our biological hardware and our linguistic software. Further research will be required to establish a direct causal link between evolutionary pressures and this linguistic structure, but the groundwork is now firmly in place.
As we move toward an era where human-AI interaction is the norm, the UVM team’s work provides a necessary recalibration. By acknowledging that human language is fundamentally a system for navigating the tension between chaos and order, safety and danger, we gain a deeper understanding of what it means to communicate. We are not just sharing emotions; we are constantly, and often unconsciously, mapping the boundaries of our own survival. The ousiometric approach provides the map for that process, ensuring that as our technology grows more sophisticated, it remains grounded in the realities of the human condition.
