20,000 words, billions of uses, and a 70-year model that did not hold up
What happened
A team at the University of Vermont says the standard model of what word meaning is built from has been wrong for about seventy years — and has proposed a replacement. The study was published in the journal Science Advances and reported more widely on 24 October 2026. The ten-person team was led by Peter Sheridan Dodds and Christopher M. Danforth, both professors at Vermont. Since the 1950s, the dominant model has described the essential meaning of a word along three dimensions, usually shortened to VAD: - valence — whether something is pleasant or unpleasant - arousal — whether it is calm or exciting - dominance — whether it is controlling or controlled That framework grew out of work associated with the psychologist Charles Osgood, and it has shaped research in psychology, linguistics and, more recently, artificial intelligence. The Vermont team argues those three dimensions are not actually independent of each other, which is a problem if you are using them as separate axes. They propose a different set of three: - power — weak or powerful - danger — safe or dangerous - structure — ordered or chaotic The test was how much of the variation in measured word meaning each framework could account for. The new framework explained more than 90%. VAD explained about 72%. The analysis drew on billions of word uses across more than 20,000 words, taken from a wide range of real-world text: books including Les Miserables, Jane Austen's novels and the Sherlock Holmes stories, alongside the New York Times, Wikipedia, Twitter and talk radio transcripts. One finding emerged from how the team handled frequency. Instead of treating every word equally, they weighted words by how often people actually use them. Doing that revealed that words to do with safety appear far more often than words to do with danger — across books, news, social media and speech alike. The team gave their approach a name, ousiometrics, from the Greek for essence, and built tools to go with it: an "ousiometer" for scoring text, and "ousiometric trajectories" for tracking how meaning shifts across a story. Lead author Peter Dodds said: "Expressions of safety are crucial to all language." His colleague Julia Zimmerman put the case for the new dimensions this way: "Power, danger, and structure," she said, "are relevant to every person that's ever lived"
Why this matters
This sounds abstract. It is not, because these frameworks are already running in things people use every day. Sentiment analysis — software that reads text and decides whether it is positive or negative — is built on models of this kind. It is used to sort customer reviews, flag content on social platforms, screen job applications, and monitor public mood. Large language models are trained on text that has often been labelled using similar schemes. If the underlying dimensions are the wrong ones, or are not really independent, then everything built on top inherits the problem. There is also a well-known puzzle this study claims to explain. Researchers have long observed that human language skews positive. Pleasant words outnumber unpleasant ones in almost every language studied. This is called the Pollyanna principle, after a relentlessly cheerful children's book character. The Vermont team's argument is that positivity was never the real pattern. What they say is going on is a preference for expressions of safety, and the positivity bias is simply what you see when you squash that into a single pleasant-to-unpleasant axis. That is a more interesting claim than it first appears. Positivity is hard to explain. A bias toward talking about safety is easier to connect to how people actually live: checking that things are all right, telling each other it is fine, describing what is secure and what is not. For teaching and for writing, the three dimensions are also simply useful to think with. Power, danger and structure are a decent checklist for why a sentence lands the way it does.
What the evidence actually says
The scale of the data is the strongest part of this study, and it is worth saying what the numbers do and do not cover. Billions of word uses across more than 20,000 words is a very large sample. It spans novels, newspapers, an encyclopaedia, social media and spoken radio, which is a genuinely broad mix of registers. That breadth is important, because a pattern that only showed up on Twitter would mean much less. The 90% versus 72% comparison is the central result. It says the new three dimensions account for more of the variation in measured word meaning than the old three do. That is a real improvement, and a substantial one. But two caveats matter. First, this is about English. The paper's data are English-language sources, and the Pollyanna principle has been reported across many languages. Whether power, danger and structure outperform VAD in Mandarin, Arabic or Swahili is a question for other studies. Second, the evolutionary explanation is explicitly not demonstrated. The authors note that their results are consistent with a link between language and survival — that humans might be built to talk about safety because talking about safety kept people alive. They are careful to say they have not shown that evolutionary pressure caused the bias. It is a story that fits, nothing more. The implications the team suggests for artificial intelligence, psychology and neuroscience are proposals rather than findings. Nobody has yet rebuilt a sentiment model on these dimensions and shown it works better in practice. The study was funded by the US National Science Foundation, Google, MassMutual and others. Google's involvement is worth noting given the AI relevance, though there is nothing to suggest it affected the analysis.
What to know
Researchers at the University of Vermont, writing in Science Advances, argue that the long-standing valence-arousal-dominance model of word meaning should be replaced by power, danger and structure. Their framework explained more than 90% of the variation in measured word meaning, against about 72% for the older model. The analysis covered billions of word uses across more than 20,000 words, drawn from novels, the New York Times, Wikipedia, Twitter and talk radio. Weighting words by how often people actually use them revealed that safety-related words far outnumber danger-related ones. The team says this, rather than simple positivity, explains the long-observed Pollyanna principle. The data are English-language only, and the suggestion that this reflects evolutionary pressure is explicitly not demonstrated. Implications for artificial intelligence and psychology are proposed, not yet tested. The team has released tools, including an "ousiometer" for scoring text. Sources: Science Advances, "Ousiometrics: The essence of meaning aligns with a power-danger-structure framework instead of valence-arousal-dominance", Dodds P S et al., 2026, volume 12 issue 19, DOI 10.1126/sciadv.adr4039, https://doi.org/10.1126/sciadv.adr4039; ScienceDaily, "A hidden bias in human language challenges 70 years of science", 24 October 2026, https://www.sciencedaily.com/releases/2026/10/261023023738.htm; University of Vermont, https://www.uvm.edu. This article is for general information. It covers language and data research and contains no health advice; for any health question, speak to a doctor, pharmacist or other qualified healthcare professional.
Related articles
7,300 years on, a supervolcano's tank is filling again
Scientists have mapped the magma reservoir beneath Kikai caldera, a mostly underwater volcano south of Japan, and found it is refilling with fresh molten rock a…
A crow-sized bird with dinosaur teeth that gave up flying 150 million years ago
Palaeontologists in China have described a crow-sized bird from the Late Jurassic that appears to have lost the ability to fly — making it the earliest known ex…
Five sections, no new rules: the violence guidance schools were just sent
On 22 October 2026 the Department for Education published guidance called "Understanding and preventing serious violence in education". It was updated the follo…