A new tool drawing on billions of uses of more than 20,000 words and diverse real-world texts claims to have found that human language is systematically biased, but not against things. It is instead biased toward safety and that has impacted everything from psychology claims to how Large Language Models (LLMs, colloquially called Artificial Intelligence and AI).

Language scholars in the 1950s distilled it into three basic emotional dimensions, listed as valence (positive compared to negative), arousal (excited compared to calm), and dominance (controlling compared to submissive) but then new work says those basic emotions are not independent and mask a better organizing principle: power (vs. weak), danger (versus safe), and structure (ordered vs. chaotic).

How many of you noticed versus was written out once and abbreviated the other two times? An LLM tool can't see the difference and psychologists won't hear it but it's there and this new "ousiometrics" (ouisa is Greek for 'essence') tool is touted as better explaining variance in meaning than legacy approaches.

It reveals bias that went undetected before; strongly favoring words associated with safety over those associated with danger. If so, this reframes the"Pollyanna principle" - human language skews positive - as a one-dimensional projection of an underlying safety bias.


Ousiograms showing the analytic sequence moving from the valence-arousal-dominance (VAD) framework (top row) to the goodness-aggression-structure (GAS) and power-danger-structure (PDS) frameworks (second and third rows). Row 1 (A to C): Ousiograms for the three pairs of variables Va, Ar, and Dm for ~20,000 words in the VAD NRC lexicon (19) [(B) corresponds to Fig. 1]. We determine the ellipses by using singular value decomposition (SVD) in each plane, ignoring the third dimension. The ill fit of the VAD framework is apparent for the misalignments of ellipse axes. Word annotations along the edges of the nine pairwise distributions, coupled with ranked words lists by component size (figs. S29 to S33), enable interpretation of the new frameworks of GAS and PDS. Row 2 (D to F): We perform SVD on the full matrix formed by the Va, Ar, and Dm scores, and identify goodness Gd, aggression Ag, and structure St, with the first two dimensions accounting for more than 90% of explained variance. Row 3 (G to I): Rotating the goodness-aggression plane by π, they uncover a framework with {weak ⇔ powerful} and {safe ⇔ dangerous}. The GA and PD dimensions interlink to form an interpretable circumplex model GPADS. See Fig. 3 for a larger, more detailed power-danger ousiogram. As any lexicon reflects only the possible but not the used language (types versus tokens), whether or not the VAD, GAS, or PDS frameworks are sensible must be tested by considering real corpora. See the “Assessing the failure of the VAD framework” section and Eqs. 2 and 4 for interpretation of the VAD, GAS, and PDS relationships.

This might even provide a science basis to language; that communication is shaped by evolutionary pressures tied to survival. Words are not just emotional signals, humans constantly addressing situations, people, and actions as safe or dangerous could be foundational to emotion. The Pollyanna principle, bias in language toward expressions low aggression are instead shadows of an underlying linguistic safety bias.

Tools like ChatGPT were created using those legacy Valence-Arousal-Dominance (VAD) frameworks so if this is accurate it could explain why they systematically misinterpret language and are instead being fancy autocomplete tools.

Psychology could also step closer to being science if interpretations of emotion, perception, and behavior are revisited in a more accurate framework.

One example in their analysis traces the “ousiometric trajectory” for an English translation of Victor Hugo’s Les Misérables. Like all great stories it deals with dangerous vs. safe, weak vs. powerful, gentle vs. aggressive, and bad vs. good, so to understand it in their framework they create abstract entities (“types”) and words as they're used (“tokens”). “Apple” is a type and every time the word “apple” is used in a sentence is a token.

An LLM will treat words as equally important but in the ousiometrics analysis usage frequency was able to reveal patterns, like the safety bias, that only emerge in real-world language.

Citation: Peter Sheridan Dodds, Thayer Alshaabi, Mikaela Irene Fudolig, Julia Witte Zimmerman , Juniper Lovato , Shawn Beaulieu, Joshua R. Minot , Michael V. Arnold, Andrew J. Reagan, Christopher M. Danforth, 'Ousiometrics: The essence of meaning aligns with a power-danger-structure framework instead of valence-arousal-dominance', Science Advances, May 6 2026; Vol 12, Issue 19, https://www.science.org/doi/10.1126/sciadv.adr4039