Hidden safety bias in human language challenges 70 years of linguistic theory

University of Vermont researchers propose new framework, ousiometrics, based on power, danger and structure

By LineZotpaper AI staff

Published24 October 2026

Read Time2 min

Researchers at the University of Vermont have discovered a consistent preference for safety over danger in human language, a finding that overturns a widely accepted model of word meaning that has stood for more than seven decades. The study, published in Science Advances, introduces a mathematical approach called ousiometrics and shows that the traditional valence-arousal-dominance (VAD) framework, developed by psychologist Charles Osgood in the 1950s, may miss a deeper structure rooted in survival priorities.

Since the 1950s, much of the scientific understanding of word meaning has relied on the VAD model, which describes meaning through three emotional dimensions: valence (positive vs. negative), arousal (excited vs. calm), and dominance (controlling vs. submissive). The University of Vermont team analysed billions of uses of more than 20,000 words across a wide range of real-world texts, including books, news reporting, social media and spoken language. Using advanced computational techniques, they found that the three VAD dimensions are not independent, and that word meaning can be more effectively described through power (weak vs. powerful), danger (safe vs. dangerous) and structure (ordered vs. chaotic).

The new ousiometric framework accounted for more than 90% of the variation in measured word meaning, compared with approximately 72% for the traditional VAD approach. The researchers also uncovered a widespread bias: words associated with safety appear far more frequently than those associated with danger. This offers a new explanation for the so-called Pollyanna principle, the tendency for language to lean toward positive words.

"The Pollyanna principle's positivity bias is, in fact, a one-dimensional projection of an underlying safety bias," the study concludes.

"This is a big observation that comes out of this work," said Peter Dodds, director of the UVM Complex Systems Institute and senior author. "Expressions of safety are crucial to all language."

Julia Zimmerman, a postdoctoral researcher and co-author, said the three newly identified dimensions are "relevant to every person that's ever lived."

The researchers suggest that the bias may reflect evolutionary pressures: distinguishing safety from danger is essential to survival, and language may be shaped by that priority. However, the study does not establish a direct evolutionary cause.

Beyond linguistics, the findings have implications for artificial intelligence. Many AI systems use sentiment analysis or the VAD framework to interpret emotional tone. Incorporating power, danger and structure could help models understand language more accurately, particularly in areas involving trust, risk and decision-making.

Analysis

Why This Matters

The discovery challenges a foundational model of word meaning used in psychology, linguistics and AI for over 70 years, potentially reshaping how researchers understand human communication.
If confirmed, the ousiometric framework could improve AI systems that rely on sentiment analysis, content moderation and large language models, making them more attuned to underlying signals of safety and threat.
The findings raise broader questions about whether language structure reflects biological survival priorities, opening a new avenue for neuroscience and evolutionary psychology.

Background

The VAD model, developed by psychologist Charles Osgood and colleagues in the 1950s, has been a staple of research into how words carry emotional meaning. It treats meaning as composed of three independent emotional dimensions: valence (good–bad), arousal (calm–excited), and dominance (submissive–controlling). The model has been widely used in sentiment analysis, linguistics and psychology. The new study proposes that these dimensions are not independent and that a different set, power, danger and structure, better captures the organisation of word meaning. The research was supported by the US National Science Foundation, Google, MassMutual and other funders.

What to Watch

Whether the ousiometric framework is replicated in other languages and cultures, as the current study focused on English.
How AI developers respond: if major language models or sentiment analysis tools begin incorporating power, danger and structure as primary dimensions.
Any follow-up research testing the hypothesised link between the safety bias and evolutionary survival mechanisms, which the study notes is not yet established.

Sources

A hidden bias in human language challenges 70 years of science · ScienceDaily

Checked

11 claims in this story checked against 1 source: 8 supported.

Show each claim

SupportedThe headline says a hidden safety bias challenges 70 years of linguistic theory.Latest Science News -- ScienceDaily: “A hidden bias in human language challenges 70 years of science. Scientists have uncovered a hidden survival-driven pattern in human language that overturns 70 years of assumptions about how words carry meaning.”
Partly supportedThe opening says the finding overturns the widely accepted VAD model.Latest Science News -- ScienceDaily: “Their findings suggest that this long-established model overlooks important aspects of how meaning is organized.”
SupportedThe study concludes that the Pollyanna principle’s positivity bias is a one-dimensional projection of an underlying safety bias.Latest Science News -- ScienceDaily: ““The Pollyanna principle's positivity bias,” the study concludes. “is, in fact, a one-dimensional projection of an underlying safety bias.””
SupportedPeter Dodds said, “This is a big observation that comes out of this work,” and called expressions of safety crucial to all language.Latest Science News -- ScienceDaily: ““This is a big observation that comes out of this work,” said Peter Dodds, director of the UVM's Complex Systems Institute and senior author of the study. “Expressions of safety are crucial to all language.””
SupportedJulia Zimmerman said the dimensions are “relevant to every person that's ever lived.”Latest Science News -- ScienceDaily: ““Power, danger, and structure,” she said, “are relevant to every person that's ever lived.””
Partly supportedThe VAD framework was developed by psychologist Charles Osgood in the 1950s.Latest Science News -- ScienceDaily: “Since the 1950s, much of the scientific understanding of word meaning has relied on a framework developed through pioneering research by psychologist Charles Osgood and others.”
SupportedThe team analysed billions of uses of more than 20,000 words.Latest Science News -- ScienceDaily: “Their analysis examined billions of individual uses of more than 20,000 words across a wide range of real-world texts.”
SupportedThe ousiometric framework explained more than 90% of variation in measured word meaning, compared with about 72% for VAD.Latest Science News -- ScienceDaily: “The research ... found that this new framework accounted for more than 90% of the variation in measured word meaning. The traditional VAD approach explained approximately 72%.”
SupportedWords associated with safety appeared much more frequently than words associated with danger across books, news reporting, social media and spoken language.Latest Science News -- ScienceDaily: “Across books, news reporting, social media, and spoken language, words associated with safety appeared much more frequently than those associated with danger.”
SupportedThe study does not establish that evolutionary pressures directly caused the safety bias.Latest Science News -- ScienceDaily: “Although the results are consistent with a possible connection between language and survival, they do not establish that evolutionary pressures directly caused the safety bias.”
Partly supportedMany AI systems use sentiment analysis or the VAD framework to interpret emotional tone.Latest Science News -- ScienceDaily: “Many artificial intelligence systems use sentiment analysis to determine whether language expresses positive, negative, or neutral feelings. Some approaches also draw on the traditional VAD framework to measure emotional meaning.”