Skip to content

R0058/2026-04-03/C001/SRC02/E01
Research R0058 — Candidate evidence test
Run 2026-04-03
Claim C001
Source SRC02
Evidence SRC02-E01
Type Statistical

94% of AI ethics institutions in the largest connected component; research ecosystem appears highly connected

URL: https://www.degruyterbrill.com/document/doi/10.1515/ijdlg-2025-0007/html

Extract

Key findings from 6,084 articles (2015-2025) in the Web of Science:

  • 94% of 123 countries/institutions belong to the largest connected component in the co-country network
  • 16 distinct keyword clusters organized into three thematic categories: Frontiers of AI Innovation, Socio-Ethical Dimensions, and Transformative Applications
  • Top institutions: University of London (157 publications), University of Oxford (143), Harvard University (139)
  • The field demonstrates "robust international collaboration" and is described as an "inherently interdisciplinary" research ecosystem

Relevance to Hypotheses

Hypothesis Relationship Strength
H1 N/A This study examines AI ethics broadly, not the specific safety-ethics divide. It does not confirm or deny the 83% figure.
H2 Supports The high connectivity at the institutional level is consistent with the broader research ecosystem being connected even while the specific safety-ethics subfields remain siloed
H3 Supports Could be read as evidence that the divide is overstated — if 94% of institutions are connected, the communities may be less siloed than Roytburg & Miller suggest. However, this measures a different network (countries/institutions vs. individual co-authorships) and a different scope (AI ethics broadly vs. safety-ethics specifically).

Context

This study and Roytburg & Miller measure fundamentally different things. Qiu et al. examine AI ethics as a broad field using Web of Science data, finding a well-connected global ecosystem. Roytburg & Miller specifically examine the safety-ethics divide within AI alignment using ML/NLP conference data. The two findings are not contradictory — the broader field can be well-connected while specific subfield boundaries remain sharp.

Notes

The 94% connected component finding applies to country-level co-authorship networks, not individual researcher networks. This is a coarser-grained measurement that would naturally show higher connectivity.