Research note
Homophily under model substitution
A scaled-down Claude Haiku 4.5 replication of He et al. (2026), written for quick technical review.
Research Question
He et al. (2026) report that GPT-3.5-powered agents on Chirper.ai formed homophilous social structures without being instructed to do so. This short replication asks whether the core mechanism generalises across model families and time: if the same kind of engagement environment is rebuilt with Claude Haiku 4.5 in May 2026, do agents still preferentially engage with semantically similar others?
Design
I ran a topic simulation rather than a faithful clone of Chirper. The completed run, full-100x12, used 100 Claude Haiku 4.5 agents over 12 rounds on the topic: Should universities ban AI in coursework?
Agents were balanced across five latent perspectives. Each round, every agent wrote a short post and chose 1-3 posts from a mixed feed to like, follow, or ignore. The prompt asked agents to decide in character, but never told them to seek similar agents.
Post histories were embedded with sentence-transformers/all-MiniLM-L6-v2. Weighted engagement graphs were analysed with Louvain communities, modularity, assortativity by detected community, and a 100-iteration degree-preserving bootstrap null.
Results
| Metric | He et al. English subset | Claude Haiku simulation |
|---|---|---|
| Agents / duration | 17,746 / 28 days | 100 / 12 rounds |
| Final communities | - | 7 |
| Final modularity | 0.38 | 0.131 |
| Bootstrap 95% null interval | - | [0.101, 0.117] |
| Bootstrap p-value | < .001 | 0.000 |
| Final assortativity | 0.61 | 0.069 |
| Content-engagement correlation | significant in paper | 0.018 |
Classification: reproduces, with attenuated magnitude. The effect is much smaller than the original Chirper result, but the final modularity is clearly above the bootstrap null and the other metrics remain weakly positive.
Interpretation
The result matters because it tests the part of the finding that is important for synthetic-audience systems: interaction can change a simulated population. A synthetic audience is not only a static panel of persona cards; it is a social process that can concentrate attention and amplify similarity over time.
For a system such as Radiant, the product implication is a diagnostic one. A simulated buyer committee, shareholder base, or policy audience may begin diverse but become less diverse through repeated rounds of interaction. Surfacing modularity, assortativity, and content-similarity engagement over time would make that drift visible.
Limitations
This is a scaled-down replication, not a full reproduction. It uses one seed topic, 100 agents, and 12 rounds, while He et al. analysed tens of thousands of agents over 28 days on a richer social platform. The feed is a compact topic simulation approximation rather than a live Twitter-like system. Community detection also differs: I used Louvain because it is a modern weighted-graph default, while the paper reports label propagation and fast-greedy variants.