A paper on opinion dynamics, co-authored by Prof. Katarzyna Sznajd-Weron from the Faculty of Management, has been published in the world’s most prestigious review journal in physics. It is the first paper in the history of “Reviews of Modern Physics” to feature an affiliation with Wrocław University of Science and Technology.

The article “Opinion dynamics: Statistical physics and beyond” was prepared by an international team of researchers affiliated with institutions in eight countries. They combined tools from statistical physics with insights from sociology, psychology, economics, computer science and network science.

A common language for many disciplines

For many years, research on opinion dynamics developed along parallel tracks, often using different concepts and methods. The paper organises this body of knowledge and proposes a common language for researchers from various disciplines.

The authors deliberately move away from the traditional division of the literature into classes of models. Instead, they structure the field around two levels: macroscopic social phenomena, such as consensus, fragmentation and polarisation, and microscopic psychological mechanisms responsible for changes in opinion.

This approach makes it possible to connect research conducted across different disciplines and using different methods.

Wrocław Tech in “Reviews of Modern Physics” for the first time

“Reviews of Modern Physics” is the flagship review journal of the American Physical Society. It publishes studies that not only summarise current knowledge but also organise entire fields of research and identify the most important problems and directions for future work. In 2025, the journal published just 32 articles, while its Journal Impact Factor was 48.9.

The publication is a major success for Wrocław University of Science and Technology and confirms that our researchers contribute to the international academic debate and help shape new directions of research.

People instead of atoms

The article explores opinion dynamics, examining how individuals’ views change under the influence of others and how numerous such interactions give rise to phenomena affecting entire societies.

To study these processes, the authors use tools from statistical physics, among other methods.

“From the perspective of physics, the starting point is a question very similar to those asked in statistical physics: how does the behaviour of a large system emerge from numerous simple interactions between its elements? The difference is that instead of atoms, we are dealing with people and their opinions,” explains Prof. Katarzyna Sznajd-Weron from the Faculty of Management.

This does not mean equating a person with an atom, but applying a similar approach to analysing large systems. Researchers investigate how the behaviour of individuals and the relationships between them affect the group as a whole.

In such models, each person is represented by an agent, a simplified model of a member of a community. The agent holds a particular opinion. This may be a “yes” or “no” choice, a position on a scale ranging from opposition to support, or a set of views on several issues.

The model defines the rules governing how agents influence one another. They may follow the majority, oppose it, seek contact with people who share similar views or maintain their existing position. The pattern of relationships between individuals also matters.

After many such interactions, the simulation shows what happens to the group as a whole: whether it reaches consensus, divides into opposing camps, fragments into several groups or forms echo chambers.

“This does not mean that we can accurately predict the decisions of a particular person,” Prof. Sznajd-Weron points out. “Models are used to study patterns emerging in large groups and the mechanisms that lead to particular social phenomena.”

Not every division is polarisation

Models can produce a variety of phenomena that are often confused in everyday language. An important part of the publication is therefore devoted to clarifying key concepts.

Polarisation may refer, among other things, to the emergence of two opposing camps, a growing distance between viewpoints or increasing hostility between groups. Fragmentation, on the other hand, means that society splits into a larger number of groups that do not necessarily form two opposing camps. The isolation of part of a social network is also not enough to constitute an echo chamber. An echo chamber emerges when people are surrounded primarily by individuals and content that reinforce their existing beliefs.

Drawing precise distinctions between these phenomena makes it possible to determine whether research teams using different methods are actually analysing the same problem.

Models must be confronted with reality

“When constructing a model, we base its assumptions on what psychology and other social sciences tell us about human behaviour,” says Prof. Katarzyna Sznajd-Weron. “We then examine whether the model’s results are consistent with phenomena that we actually observe in society. We can also go a step further and use the model to design an experiment involving people that directly tests its predictions. In this sense, models can also inspire new experiments,” she adds.

The authors therefore devote considerable attention to laboratory and online experiments, election results, polls, surveys and large-scale social media data.

Comparing models with such data makes it possible to assess which mechanisms may lead to the observed phenomena. Researchers can therefore ask “what if?” questions and analyse the conditions that encourage consensus or reinforce social divisions.

AI as a social laboratory

The agents described earlier operate according to rules defined by researchers. Agents powered by large language models offer new possibilities. In a simulated environment, they can formulate statements, exchange arguments and respond to other participants. However, their greater complexity does not mean that they accurately reproduce human behaviour. Language models may replicate biases present in their training data, oversimplify differences between people and create groups that are more homogeneous than those found in reality.

“Agents based on large language models play a different role from those used in classical agent-based models,” says Prof. Katarzyna Sznajd-Weron. “In a classical model, we define the rules governing agents’ behaviour and their interactions, so we can investigate which phenomena result from particular mechanisms. We do not have the same knowledge in the case of LLM-based agents, which means that they answer different questions. Such studies are closer to social experiments, except that they are conducted on artificial participants. Sometimes we use them to reproduce human behaviour, while at other times we are interested in communities composed entirely of AI agents and the phenomena that emerge within them. In the first case, we must of course remember that AI agents are not people,” concludes Prof. Sznajd-Weron.

Starnini M., Baumann F., Galla T., Garcia D., Iñiguez G., Karsai M., Lorenz J., Sznajd-Weron K., „Opinion dynamics: Statistical physics and beyond”, Reviews of Modern Physics, 2026.