Who gets to test AI, and who controls the results?
The authors examine who can run AI safety tests and share the results. They map 12 African projects and ten safety institutes in three joint exercises. Some groups can add tasks or labels while others control access to the AI or publication. These cases do not cover every evaluator or show a measured drop in harm.
What the source reports
The paper offers a way to ask who has power in an AI safety test. The authors examine access, the quality of the test, independence, and who makes final choices. They map 12 evaluation projects in Africa. They also document the roles of ten safety institutes across three joint exercises. A group might help write a task or label a result yet still rely on someone else to run the AI or publish the findings. That distinction matters when a test is called independent. The cases are not a count of all AI evaluators. They also do not measure whether this approach reduced harm. The paper maps control over testing; it does not prove a safety outcome.
Original source
- Title
- Who Can Evaluate AI Independently? A Framework for Analysing Evaluation Infrastructure
- Author
- Sumaya Nur Adan, Joanna Wiaterek, Nourhan Abdelrhim, Sydney Reis, Robert F. Trager, Benjamin Rosman, Marina Jirotka
- Publication
- arXiv cs.CY
- Date
- Friday, November 6, 2026