Assessing Risks and Impacts of AI (ARIA) is a research program by the National Institute of Standards and Technology (NIST) aimed at developing evaluation methods and criteria that assess AI’s risks and impacts in real-world scenarios. NIST recently released the ARIA Evaluation Design Document, which details the nature, structure, and objectives of the program.
ARIA's method includes a three-layered evaluation process, as seen in the flowchart below, which culminates in the AI system being assigned a score on the Contextual Robustness Index (CoRIx). The CoRIx score, which is ARIA’s measurement instrument and suite of metrics, indicates whether AI systems can maintain robust and trustworthy functionality ranging over a variety of deployment conditions. It builds off of NIST's AI Risk Management Framework (AI RMF), utilizing a total of ten measurement dimensions, including the seven trustworthy characteristics enumerated in the AI RMF.
Red teaming AI systems, the practice of rigorously challenging a system to identify any potential weak points or flaws, is an emerging discipline in the field of machine learning. It will be important to monitor ARIA's development and standardization of such practices, especially because in some cases, engaging in evaluations like red teaming can be a prerequisite for receiving the benefits of indemnification by a foundation model provider.
Evaluation of AI applications starts in ARIA’s three-level testbed, in which each level uses a different testing approach to explore potential risks and impacts: (1) Model testing: confirm claimed capabilities (2) Red teaming: stress test and attempt to induce risks (3) Field testing: examine positive and negative impacts that may arise under regular use
ai-challenges.nist.gov/...
