AI models produce different opinion dynamics under the same discussion rules

Language models are increasingly used as synthetic participants in social-science experiments. Researchers give them personas, place them in discussions and observe how their expressed opinions change. But does the prompt determine the behavior, or does the underlying model help determine the outcome?

A preprint by Nataša Djurdjevac Conrad, Sebastian Pokutta and Christof Schütte, posted to arXiv on November 2, 2026, examines that question across six language models. Its central finding is that otherwise matched simulated discussions can produce substantially different opinion patterns depending on the model. These are results about artificial participants, not a demonstration that the simulations reproduce human opinion change.

Measuring opinions separately from modeling them

The main experiments used ten agents discussing government subsidies for electric vehicles over 20 rounds. Each agent received a fixed, one-sentence persona and an initial position. In the standard discussion condition, every agent heard every other agent’s statements from the preceding round.

After composing statements, agents answered six survey items on a seven-point agreement scale. Three items pointed toward support for subsidies and three toward opposition. The researchers combined those answers into a stance score between −1 and +1.

This measurement procedure was independent of the mathematical models subsequently fitted to the trajectories. Those models included descriptions of social averaging, attachment to initial positions, movement toward a common destination and pole-directed movement. Separating measurement from fitting allowed the authors to compare alternative descriptions rather than building a particular dynamics rule into the opinion score.

Matching rules, different outcomes

With the firm-but-uncommitted personas, every model reduced disagreement during full-group discussion. The extent varied considerably.

Starting from a population dispersion of 0.47, the final dispersion across five replicates was 0.08 for claude-sonnet-4-5 and 0.12 for gpt-5.6-sol. By comparison, gpt-4.1 ended at 0.34 and claude-sonnet-5 at 0.36. Dispersion here means the standard deviation of agents’ stance scores: lower values indicate opinions closer together, not necessarily better reasoning or greater accuracy.

The best-fitting mathematical description also differed between models. However, these fits describe the observed trajectories; they do not establish the language models’ internal mechanisms. For claude-sonnet-4-5, the leading candidate descriptions were effectively indistinguishable under the authors’ selection criterion.

Keeping a side is not keeping an exact opinion

A separate experiment assigned five strongly supportive and five strongly opposing personas. Without discussion, every agent in each of the five models tested retained its assigned side for 20 rounds. Their precise positions could still move: sign retention was not equivalent to complete stability.

Discussion produced sharper differences. The authors report that gpt-4.1 and gemini-3.7-flash maintained separated opinion clusters, while claude-sonnet-5 retained them partially. For gpt-5.6-sol and claude-opus-5, discussion reduced separation and increased average displacement from initial positions relative to the no-discussion controls. Some agents crossed to the opposite side.

The authors call the loss of an assigned stance during discussion “alignment drift.” In this paper, that term concerns fidelity to a simulated persona’s policy position, not a general measure of AI safety.

There is an important qualification: the no-discussion control removed both peer statements and the statement-composition phase. It therefore does not isolate exposure to peers from every other difference between the conditions.

Attention and identity change the picture

When agents selected which peers to hear based on their prose, gpt-4.1 preferentially retained peers with more similar positions and maintained greater separation than under full-group discussion. Claude-sonnet-5 showed a weaker similar-attention pattern. Other models did not uniformly behave this way.

An imposed group-identity condition increased gpt-4.1’s dispersion above its starting level. Neutral-start controls did not spontaneously polarize. The findings therefore concern maintaining or amplifying existing divisions, rather than establishing that these models create polarization from an initially neutral population.

A simulation variable, not an interchangeable component

Experiments on a four-day working week showed a broadly similar cross-model ordering, although individual rankings and fitted model labels could change. Larger, longer runs also supported the authors’ qualitative interpretation without preserving every exact mathematical classification.

The study remains limited to two economic-policy issues, English prompts, small populations, proprietary models and three to five replicates. Each simulated society used one model, and a single opinion axis omitted multidimensional trade-offs and emotional dynamics. Model snapshots and prompting choices may also affect the profiles.

The practical implication is methodological: researchers should report and ideally vary the underlying language model, rather than assuming that a persona prompt makes different models interchangeable.