Watch opinions spread
Replay how statements about one question changed across a 5×5 network of humans, of AI agents, or of both, and paint your own network to see how polarization and neighbour agreement are measured.
Explainer This page is not an experiment. It replays published runs and lets you try the measures, and nothing you do here is recorded or sent anywhere.
In the opinion diffusion experiments, statements on the question Does red meat cause cancer and cardiovascular disease? travel across the 25 nodes of a 5×5 grid. At every round each node is filled by a new participant, human or AI, who sees the statements at the neighbouring nodes and at their own node from the round before, picks the one the group would most likely agree with, and rewrites it. A language model then labels each new statement as positive, negative or neutral. The network starts split in two: statements supporting the claim on top, statements opposing it below.
- + positive
- − negative
- 0 neutral
- polarization
- neighbour agreement (NCI)
The paper’s figure shows iterations 0, 1, 3, 5, 7 and 8, so the replay steps through those. The runs also passed through iterations 2, 4 and 6, so the chart’s straight lines skip over some turns.
Two numbers for a network
Polarization is the variance of the opinions, scoring positive as +1, negative as −1 and neutral as 0. It is highest when the network splits evenly between the two sides and zero when everyone says the same thing. Neighbour agreement (the Neighbors Correlation Index, NCI) is the correlation between each person’s opinion and the average of their neighbours’. A network can be highly polarized and still have high neighbour agreement if the two sides form separate camps. Try it: click a node to cycle its opinion, or use the arrow keys and Enter.
Two camps: high polarization, high agreement. Checkerboard: about the same polarization, with every neighbour disagreeing. Everyone agrees: no polarization, and agreement is undefined because there is nothing to correlate.
What the study found
Polarization fell in all three conditions. The hybrid network ended with the lowest polarization and the highest neighbour agreement; the human-only network ended with the highest polarization and the lowest agreement. For the AI-only condition these comparisons use the average of 21 runs, so the single run replayed above can end lower on both. These results are preliminary: the human and hybrid networks were each run once, on one question, with one language model.
Where this comes from
Related work
- Hybrid human-AI societies balance creativity and diversity Mixed networks of people and AI keep more variety than either alone.
- Characterizing the interaction of cultural evolution mechanisms in experimental social networks How copying, selection and network structure combine in experimental networks.
- Using LLMs to advance the cognitive science of collectives The case for studying groups of people and language models together.