Human language hides a safety bias that rewrites 70 years of meaning science
A University of Vermont study finds that word meaning is organized around power, danger, and order — not emotion — and that our language is systematically skewed toward safety, possibly echoing ancient survival instincts.
For more than 70 years, scientists studying the meaning of words have relied on a model built around emotion. A new analysis of billions of word uses suggests they have been measuring the wrong thing — and that buried inside everyday language is a powerful, previously undetected preference for safety over danger.
The study, published May 6 in Science Advances by researchers at the University of Vermont, introduces a framework called 'ousiometrics' — the quantitative study of essential meaning — and argues that the emotional model dominating psychology, linguistics, and artificial intelligence since the 1950s is fundamentally incomplete.
The dominant framework, known as the VAD model, describes word meaning through three emotional dimensions: valence (positive vs. negative), arousal (excited vs. calm), and dominance (controlling vs. submissive). Developed through pioneering work by psychologist Charles Osgood and others, it became the backbone of sentiment analysis in AI and decades of research into emotion, perception, and behavior.
What the Vermont team found, after analyzing billions of uses of more than 20,000 words drawn from books, news reporting, social media, and spoken language, was that the VAD model's three dimensions do not operate independently of one another — a core assumption of the framework that their data did not support. More importantly, the researchers identified a different set of dimensions that does a better job of capturing what words mean.
Those dimensions are: power (weak vs. powerful), danger (safe vs. dangerous), and structure (ordered vs. chaotic). The three dimensions together explained more than 90% of the variation in measured word meaning. The traditional VAD approach explained approximately 72%.
The most unexpected finding came when the team examined how words actually appear in real-world communication. Across every type of text they studied, words associated with safety appeared far more frequently than those associated with danger — a consistent, widespread pattern the researchers call a 'safety bias'.
This is a big observation that comes out of this work. Expressions of safety are crucial to all language.— Peter Dodds, director of UVM's Complex Systems Institute and senior author of the study
The discovery reframes a long-standing observation in linguistics called the Pollyanna principle — the well-documented tendency for human language to skew toward positive rather than negative words. For decades, researchers interpreted this as evidence that people simply favor positive emotional expression. The new analysis offers a different explanation.
The Pollyanna principle's positivity bias is, in fact, a one-dimensional projection of an underlying safety bias.— the study
In other words, what looked like a preference for happiness may actually be a preference for security. Many words that register as positive, the researchers argue, do so largely because they convey safety, stability, or predictability — not because they express joy or approval.
Power, danger, and structure are relevant to every person that's ever lived.— Julia Zimmerman, postdoctoral researcher in UVM's Computational Story Lab and study co-author
Linguists have long noticed that language tilts toward expressions of goodness and low aggression, a pattern the researchers also address. The team writes that these tendencies are now understood to be 'shadows of an underlying linguistic safety bias'.
The study raises the possibility that language's structure reflects evolutionary pressures tied to survival — that words function not just as emotional signals but as tools for recognizing threats, evaluating uncertainty, establishing trust, and coordinating behavior. The researchers are careful to note, however, that the results do not establish that evolutionary pressures directly caused the safety bias; the findings are consistent with that interpretation but do not prove it.
The implications reach well beyond linguistics. Many AI systems — including large language models and automated content moderation tools — rely on sentiment analysis grounded in VAD-like frameworks. If that framework misrepresents how meaning is actually organized, those systems may be systematically misreading human language. Incorporating power, danger, and structure into AI design could produce more accurate models, particularly in applications involving risk assessment, trust, and decision-making, according to the researchers.
Decades of psychological research built on the VAD model may need to be reexamined in light of the findings, given that the foundational dimensions of meaning appear to differ from what was previously assumed — with potential implications for how researchers interpret emotion, perception, and behavior. The researchers also note that in neurobiology, the discovery resonates with what is known about the brain's sensitivity to threat and safety, suggesting a potential bridge between neural processes and the structure of language itself.
To carry out the analysis, the team developed a new instrument called an 'ousiometer,' designed to measure the essential meaning of large-scale texts quickly and at scale. Funding for the research came from the US National Science Foundation, Google, MassMutual, and additional sources.
The findings are currently tied to the specific study and text collections analyzed. Broader replication across languages and contexts is ongoing, according to reporting on the study.
Why it matters — The VAD model underpins decades of psychology research and the sentiment-analysis tools inside AI systems used by billions of people — if it is measuring the wrong dimensions of meaning, both the science and the technology built on it may need to be rebuilt from a different foundation.
⚠ Not yet confirmed
- Broader replication of the findings across languages and contexts is ongoing.
- The safety bias may reflect evolutionary pressures tied to survival.
Reported by sciencedaily.com, eurekalert.org
- Charles E. OsgoodAmerican psychologist (1916–1991)