This paper tackles a critical but often overlooked question in AI governance: Who actually has the power to evaluate AI? While independent evaluation is hailed as a cornerstone of AI safety, the "plumbing"—the compute, model access, and funding—is concentrated in very few hands.
TL;DR
The authors introduce the PAAI framework (Power, Access, Adequacy, Independence) to analyze how infrastructure arrangements dictate evaluative authority. By studying African AI projects and international safety testing, they demonstrate that simply participating in an evaluation does not grant the independence required for credible, long-term AI governance.
What Problem Does This Paper Address?
AI safety depends on third-party "audits" or evaluations to find risks before systems are deployed. However, the resources needed for these evaluations (massive compute clusters and access to proprietary model weights) are controlled by the developers themselves or a few wealthy nations.
This creates a "control gap." Even when researchers from the Global South or smaller labs are invited to participate, they often lack the authority to set the research agenda, run the models themselves, or publish findings without a provider's approval. The paper argues that without infrastructure authority, "independent" evaluation is often an illusion.
Methods
The authors synthesize Science and Technology Studies (STS) and political economy into the PAAI framework, which breaks down evaluation into four dimensions:
- Power: Divided into Design Authority (who defines the harms and metrics?) and Infrastructure Authority (who controls the compute and model access?).
- Access: The specific terms of use (e.g., duration of API credits, geographic restrictions).
- Adequacy: Whether the setup is technically rigorous and contextually fit for the local setting.
- Independence: The ability to choose targets and publish results without needing "goodwill" from providers.
The researchers applied this framework using structured thematic analysis on 12 African AI evaluation projects and coding the roles of 10 member institutes across three international AI Safety Institute (AISI) joint testing exercises.
Datasets
The paper does not introduce a new machine learning dataset but rather a meta-dataset of evaluation projects:
- 12 African AI projects (2022–2025): Including benchmarks like AfroBench and Uhura, and domain-specific audits in healthcare.
- 29 AISI Institute-by-Exercise records: Documenting the roles of countries like Singapore, the UK, the US, and Kenya in global safety testing.
Key Insights and Ideas
- Infrastructure is Authority: The most striking finding is that design authority alone never yielded more than "partial" independence. If you write the test but someone else runs the model, they ultimately control the evaluation's continuity and depth.
- The Paradox of Participation: Formal inclusion often masks structural exclusion. African actors frequently contribute "essential labor"—like translating prompts or annotating data—while external institutions retain control over the frameworks and analysis.
- Project-Based Funding is a Bottleneck: Independent evaluation requires standing infrastructure. Relying on one-off grants or expiring API credits (which the authors found in all 12 African cases) prevents the longitudinal monitoring necessary for safety.
- Framework Co-Design as Governance: When safety frameworks are designed in the West, they often fail to capture local risks (e.g., specific types of fraud prevalent in Kenyan mobile money contexts).
Key Results
The authors identify three recurring configurations of control:
- Extractive Collaboration: External leads hold all authority; local actors provide labor/data.
- Dependent Innovation: Local actors design the evaluation but rely on precarious, external infrastructure.
- Constrained Independence: Local actors hold both design and infrastructure authority but lack the resources to scale.
In the AISI joint testing exercises, a clear divide emerged: the "Leads" (Singapore, UK) ran the models and built the tools, while others (like Kenya) moved from annotation to task design but never achieved documented model execution. This suggests that even in high-level diplomatic networks, infrastructure remains a gatekeeper.
Limitations and Future Work
The study is limited by its reliance on publicly available documentation. Private contract terms for API access or compute grants are often undisclosed, meaning some aspects of "Access" and "Independence" had to be inferred from observable indicators. Furthermore, the authors acknowledge that they cannot always distinguish between a participant choosing a specialized role (like translation) versus being constrained into it by a lack of resources.
Future work will involve fieldwork and interviews with practitioners in Africa and AISI participants to understand how these power dynamics are negotiated behind the scenes.
Summary updated