A hidden bias in human language challenges 70 years of science
A groundbreaking study challenges 70 years of assumptions about how human language works. By analyzing billions of word uses, scientists discovered that meaning is organized around power, danger, and order rather than emotion alone. Even more surprisingly, human language shows a powerful hidden preference for safety, suggesting that our words may reflect ancient survival instincts.
Key points
- University of Vermont researchers found that human language consistently favors safety over danger, a pattern they call a hidden safety bias.
- The new framework of power, danger and structure explained more than 90% of variation in word meaning, versus about 72% for the 70-year-old VAD model.
- The team says the well-known Pollyanna positivity bias may actually reflect a deeper preference for safety in language.
Researchers at the University of Vermont have identified a pattern in human language that they say challenges a scientific theory accepted for more than 70 years. The study, published in Science Advances, introduces a method called ousiometrics, which uses mathematical techniques to measure the essential meaning of language. Its most striking finding is that the words people use show a consistent preference for safety over danger.
The work was supported by the US National Science Foundation, Google, MassMutual and other funders. The ten-member team was led by Peter Dodds and Chris Danforth, professors in UVM's College of Engineering and Mathematical Sciences.
A new model of meaning
Since the 1950s, word meaning has largely been described through the VAD model, which uses valence, arousal and dominance. The Vermont team analyzed billions of uses of more than 20,000 words and found those three dimensions are not independent, as previously assumed. Instead, meaning was better described by power, danger and structure.
The new framework accounted for more than 90% of the variation in measured word meaning, while the traditional VAD approach explained approximately 72%. The researchers also built a tool called the ousiometer, which scores the essential meaning of large bodies of text and can track how meaning shifts as a story unfolds.
Expressions of safety are crucial to all language.
— Peter Dodds
Safety bias and AI
Across books, news, social media and spoken language, safety-related words appeared far more often than danger-related ones. The team argues the familiar Pollyanna principle, long read as a preference for positive words, largely reflects an underlying preference for safety. The results are consistent with a possible link between language and survival, but do not establish that evolutionary pressures caused the bias.
The findings may affect AI systems that rely on sentiment analysis or the VAD framework, which the researchers say could overlook important distinctions. They suggest adding power, danger and structure could help models interpret language more accurately.
