R0058/2026-04-03/C001/SRC01/E01¶
83.1% global homophily measurement from Figure 1 caption and network analysis
URL: https://arxiv.org/html/2512.10058
Extract¶
Figure 1 caption states: "Literature on AI Ethics (left, in blue) and AI Safety (right, in red) is densely insular (83.1% homophily), with a wide gap sparsely populated by mixed-methods papers (shades of purple). Sample of top 1000 nodes by degree."
Supporting statistics from the paper body:
- Global homophily: 83.1% — "over four out of every five collaborations occur between authors who are exclusively focused on either safety or ethics"
- Safety researcher within-group collaboration: 73.5%
- Ethics researcher within-group collaboration: 68.2%
- Paper-based network homophily: 71.2%
The homophily measurement is based on analysis of 6,442 papers by 20,690 authors across 12 major ML/NLP conferences from 2020-2025. Statistical significance confirmed via label-shuffle null models (500-2000 replications, p<0.001).
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Supports | Directly confirms the 83% homophily component of the claim. The measured value (83.1%) matches the claimed value (83%) to within rounding. |
| H2 | Supports | Confirms the first component is accurate, consistent with partial correctness |
| H3 | Contradicts | Robust methodology with null model controls makes it difficult to argue the homophily measurement is materially wrong |
Context¶
The 83.1% is a "global" homophily measure — the fraction of all edges in the co-authorship network that connect same-category researchers. The subcategory rates (73.5% safety, 68.2% ethics) are lower because they measure within-group rates for each community separately. The paper-based network homophily (71.2%) is lower because it measures at the paper level rather than the author level.
Notes¶
The claim states "83%" which is a fair rounding of 83.1%. This component of the claim is accurately sourced.