Beliefs lie at the heart of human actions, as they profoundly influence our decision-making and behaviours.
When they looked at the belief map created using LLMs, researchers made some interesting discoveries.
First, they found that "relative dissonance" significantly influences people's decision-making. This essentially means that when online users encounter new information or beliefs, they tend to choose or accept those that cause them less "discomfort" or are most aligned with their existing beliefs.
More importantly, people's belief choices are shaped not just by how close a belief is to their own, but by how much closer the belief is compared to its competing belief.
When two opposing beliefs on a certain issue are equally distant, people are just as likely to choose either one. However, when one belief is clearly closer than the other, people are far more likely to choose it.
The effect observed is .... "relative dissonance." This term essentially suggests that people's decisions are influenced by the relative gap between beliefs that are closer and further from their own. Specifically, the researchers found that the greater this gap is, the stronger a person's preference is for beliefs more aligned with their own.
In other words, people are not only avoiding disagreement, but actively minimizing the difference in disagreement between available options.
This finding highlights that decision-making is not driven by absolute distance alone, but by the relative discomfort of accepting a belief that feels much further away, echoing key ideas from cognitive dissonance theory.
The findings of this recent study could have various implications. First, they provide an explanation for why some information is readily accepted by some people and strongly rejected by others, shedding new light on the processes underpinning the formation and maintenance of social perspectives.
This work also offers guidance on how messages should be constructed to be effectively delivered to a target audience, by carefully considering their existing beliefs. It could also inform the refined design of policies or campaigns aimed at encouraging behavioral change in various fields, such as health or environmental initiatives, by better understanding the intricate interplay of beliefs.
Byunghwee Lee et al, A semantic embedding space based on large language models for modelling human beliefs, Nature Human Behaviour (2025). DOI: 10.1038/s41562-025-02228-z