New Language Framework Challenges 70 Years of Science, Reveals Hidden Safety Bias

This research challenges a scientific assumption held for more than 70 years in psychology, linguistics, and artificial intelligence. The new framework could reshape how human communication is interpreted and how AI systems are built[1][9].

What the Records Show

The study, published in Science Advances on May 6, 2026, by a team from the University of Vermont, introduces "ousiometrics" as a new way to understand language meaning[1][4]. This framework proposes that the essence of meaning is explained by three independent dimensions: power (weak vs. powerful), danger (safe vs. dangerous), and structure (ordered vs. chaotic)[5][7]. This new model can account for over 90% of the variance in meaning[5]. This significantly exceeds the 72% explained by the Valence, Arousal, and Dominance (VAD) framework[1][5]. The VAD framework, developed by Charles Osgood and colleagues in the mid-20th century, has been widely used in fields like psychology, linguistics, and artificial intelligence for over 70 years[9][7].

The Vermont team analyzed billions of word uses across more than 20,000 words in diverse real-world texts, including Wikipedia pages, Google Books, The New York Times, and Sherlock Holmes stories[5][9]. Their analysis found the VAD framework insufficient, with senior author Peter Dodds noting, "The compass was off"[1][5]. A central discovery of the research is a strong, hidden bias in human language towards safety[1][2][5]. Researchers suggest this bias is rooted in ancient survival instincts[4]. To aid their research, the team developed an "ousiometer," a computational tool that measures essential meaning in large texts[9]. The study received support from the US National Science Foundation, Google, and MassMutual[9].

What Changed and For Whom

The findings have implications across several disciplines. In psychology, research previously based on the VAD model may require re-evaluation[9]. For linguistics, the study redefines the fundamental structure of language, suggesting meaning is grounded in survival-relevant distinctions rather than primarily emotional tone[9]. In artificial intelligence, this new framework offers an important approach for accurately interpreting human communication[9]. The study also indicates a connection to neurobiology, with the brain's focus on threat and safety potentially mirrored in the structure of language itself[3][9]. The team's work suggests that the "Pollyanna principle," which describes language's tendency to skew positive, is not merely emotional but reflects an underlying safety bias, indicating predictability in a shared environment[2].

What the Researchers Say

Peter Dodds, director of the University of Vermont's Complex Systems Institute and senior author of the study, observed, "The compass was off," regarding the previous approach to language meaning[5]. Dodds also stated that "Expressions of safety are crucial to all language"[2]. Study co-author Julia Zimmerman, a postdoctoral researcher in UVM's Computational Story Lab, highlighted the universal applicability of the new dimensions, asserting, "Power, danger, and structure are relevant to every person that's ever lived"[2].

What we could not confirm

• The specific long-term impact on the development of AI language models.
• The exact timeline for a re-evaluation of psychology research based on the VAD model.
• Whether the "ousiometer" tool will be made publicly available for broader use.