Vermont's Ousiometer Overturns 70-Year-Old VAD Theory

TL;DR

• A University of Vermont team has overturned the 70-year-old VAD model of linguistic meaning, the foundation of modern sentiment analysis and affective computing.
• Their new Power–Danger–Structure (PDS) framework — embedded in a broader GPADS circumplex — explains ~90% of variance in meaning vs ~72% for VAD.
• A custom instrument, the ousiometer, measures essential meaning across any text corpus over time.
• Human language carries a deep safety bias: the famous Pollyanna positivity effect is a shadow of a more fundamental survival orientation.
• The architecture of meaning mirrors the evolutionary problems early humans had to solve — who is powerful, what is dangerous, what is ordered — strengthening the case that language emerged as a survival tool.

Executive Summary

A University of Vermont research team has published what may be the most consequential revision to the theory of linguistic meaning since Charles Osgood's seminal work in the 1950s. Using a purpose-built instrument they call the ousiometer, the team analysed billions of word usages across more than 20,000 English words and demonstrated that the long-dominant Valence–Arousal–Dominance (VAD) model — the cornerstone of sentiment analysis, affective computing, and a generation of NLP systems — is structurally flawed.

In its place they propose a Power–Danger–Structure (PDS) framework, embedded within a broader Goodness–Power–Aggression–Danger–Structure (GPADS) circumplex. The new model explains roughly 90% of the variance in meaning, compared with about 72% for VAD. Embedded in the data is a striking discovery: human language is not merely positively biased (the long-observed "Pollyanna principle") — it is systematically biased toward safety. Positivity, the authors argue, is a one-dimensional shadow of a deeper survival-oriented orientation.

For the study of language origins, this matters profoundly: it suggests the architecture of meaning itself is organised around the evolutionary problems our ancestors had to solve — who is powerful, what is dangerous, and what is ordered — rather than around abstract emotional categories.

Background: The 70-Year Reign of VAD

Since Osgood's mid-twentieth-century semantic differential studies, researchers across psychology, linguistics, and artificial intelligence have largely converged on three orthogonal dimensions to describe the essential meaning of words:

VAD became the backbone of:

• Sentiment-analysis pipelines used across social media monitoring, finance, and marketing
• Affective computing and emotion-aware AI
• Lexicon engineering (ANEW, Warriner et al., NRC-VAD, and dozens more)
• A vast empirical literature in psycholinguistics

Its assumed orthogonality — that the three axes are independent — was foundational. If VAD is wrong, a great deal of downstream science and engineering is built on a shaky base.

The Vermont Discovery

Dodds, Danforth, and their colleagues at the Computational Story Lab revisited the question using modern computational scale. Their corpus included:

• More than 20,000 English word types with human-annotated affective ratings
• Billions of word tokens drawn from books (including Jane Austen and Conan Doyle), the New York Times, Wikipedia, talk-radio transcripts, and Twitter
• Automatically annotated histograms they call "ousiograms"

Their analysis produced two headline findings.

1. VAD's dimensions are not actually independent

Re-examining both word types and token usage, the team found systematic correlations between valence, arousal, and dominance — the dimensions collapse and intermix when measured at scale. The classical model obscures rather than reveals the underlying structure of meaning.

2. A better framework: Power–Danger–Structure (PDS)

The team identified three genuinely independent dimensions:

Embedded within a wider circumplex — GPADS (Goodness–Power–Aggression–Danger–Structure) — PDS is described as the minimal framework that adequately represents essential meaning. It explains roughly 90% of the variance in word meaning across corpora, a sizeable jump from VAD's ~72%.

The Safety Bias

The most viral finding is empirical rather than theoretical: across every corpus they examined, language strongly favours words associated with safety over those associated with danger.

This recasts the well-known Pollyanna principle — the observation, dating to Boucher and Osgood (1969), that human language skews positive. The Vermont team argues:

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

In other words: humans don't talk in positive terms because they are optimists. They talk in safe terms because the deep organising logic of meaning is survival-oriented. Positivity is a side-effect; safety is the signal.

The Ousiometer

To make these patterns measurable at scale, the team built the ousiometer — an instrument for sensing the essential meaning of texts over time. The name draws on the Ancient Greek ousia ("essence"), the root of the English word essence.

The ousiometer extends the team's earlier hedonometer (a "happiness meter" they have run on Twitter and other corpora for over a decade). Where the hedonometer outputs a single scalar of average positivity, the ousiometer maps text along the PDS axes and a four-quadrant grid of opposing pairs:

• dangerous ↔ safe
• weak ↔ powerful
• gentle ↔ aggressive
• bad ↔ good

Applied to a long-form text — say, a novel — the ousiometer traces a path through that grid, distilling how the essential meaning of the work evolves chapter by chapter. The authors describe the resulting trajectories as resembling multicoloured proteins — tangled, structured, and characteristic of the work.

Why This Matters for Language Origins

The reason this study scores so highly on the language-origins lens is that PDS is not merely a better statistical fit — it is a theoretically motivated one. The three dimensions map directly onto problems any social, embodied organism must solve:

1. Power — assessing capacity (mine, yours, theirs) to act on the world
2. Danger — detecting threat to survival and wellbeing
3. Structure — perceiving and creating order in a chaotic environment

If the architecture of meaning aligns with these survival-relevant distinctions, it strengthens the hypothesis that language evolved primarily as a tool for navigating a social and physical world of threats, allies, and patterns — not as an abstract emotional signalling system. It also dovetails with two other 2026 findings (the Stone Age symbol study and the 135,000-year genomic synthesis) to suggest that the content of early language was shaped by the same evolutionary pressures that shaped the capacity for language.

Implications

For linguistics and psycholinguistics

• A large body of work built on VAD lexicons will need re-evaluation.
• Theories of semantic primitives (Wierzbicka, Jackendoff) gain a new, empirically grounded interlocutor.
• The "Pollyanna principle" gets a deeper mechanistic explanation.

For AI, NLP, and sentiment analysis

• Sentiment pipelines built on VAD lexicons (ANEW, NRC-VAD, Warriner) may be systematically mis-specified.
• Foundation models and RLHF pipelines that rely on affect labelling could benefit from PDS-aligned re-annotation.
• New opportunities for safety-aware model evaluation: the ousiometer offers a principled way to measure whether a generated text drifts toward danger, disorder, or coercive power.

For cognitive science and evolutionary psychology

• PDS aligns suggestively with established threat-detection, dominance-perception, and order-perception literatures.
• The safety bias offers a quantitative anchor for theories of why human communication is fundamentally cooperative.

For business and policy applications

• Brand language analysis, crisis comms, and public health messaging can be re-tooled around PDS rather than VAD.
• Trust and safety teams gain a more discriminating instrument for monitoring discourse drift.

Caveats and Open Questions

• The corpus is English-dominant. Whether PDS generalises across typologically diverse languages is the next obvious test — and the most important one for any "language origins" claim.
• Cross-cultural replication will determine whether the safety bias is a human universal or an artefact of English-language corpora skewed toward Western, contemporary text.
• Integration with embodied and multimodal meaning (gesture, prosody, image) remains open.
• The relationship between PDS and non-linguistic affective systems (e.g., interoception, the dimensional theories of emotion) is suggestive but not yet mapped.

Bottom Line

If the VAD framework was the twentieth century's best guess at the dimensions of meaning, PDS looks like the twenty-first century's better answer. It explains more variance, sits on a more coherent evolutionary foundation, and yields a usable instrument — the ousiometer — for measuring meaning at scale. For anyone interested in why language exists at all, the Vermont team has made one of the most interesting empirical claims of the decade: meaning is organised around survival, not sentiment.

Sources

• Dodds, P. S., Danforth, C. M., et al. (2026). "Ousiometrics: The essence of meaning aligns with a power-danger-structure framework instead of valence-arousal-dominance." Science Advances, 12(9), eadr4039. DOI: 10.1126/sciadv.adr4039
• Phys.org coverage (6 May 2026): "Human language shows deep safety bias, challenging 70-year-old theory"
• University of Vermont — Computational Story Lab: paper landing page