Motivation
The use of algorithms in public policy has expanded dramatically in recent decades. They currently play an active part in informing policymakers in criminal justice decisions, welfare fraud detection, allocation of public resources, public education decisions, national defense strategy, and other areas of high societal relevance. However, standards of accountability reflecting current legal obligations and societal concerns have lagged behind their extensive use and influence. To address this gap in the research, the CRASA Project seeks to determine: How can standards be established for evaluating the accountability of public policy algorithms?
Many government institutions are already struggling with this question. While there have been attempts to increase community input into decisions regarding the use of algorithms, establish auditing standards, or even ban some algorithms altogether, there is little guidance for best practices in evaluating the algorithms. This project proposes a multidisciplinary, community-based participatory research program for analyzing methods to balance societal needs for accountability, current legal standards, and practical issues of algorithm auditing.
Goals and Objectives
The goal of this research program is to develop an algorithm accountability benchmark that can be applied to a variety of public policy algorithms and can be used by governments, advocacy groups, and corporations for design and evaluation.
Objective 1
Discover and interview stakeholders to document the needs of public policy actors. Conduct interviews with various stakeholders in different public policy areas focusing on the Harris County, TX community and convene a Community Advisory Board (CAB) to advise and oversee the research.
Objective 2
Perform a comprehensive review of legal and regulatory frameworks that shape the accountability of AI algorithms to be used by the auditing tools and processes; and informing members and community partners.
Objective 3
Develop an algorithm accountability benchmark that will provide specific standards of the information needed for the community, internal auditors, external auditors, and the legal system to evaluate algorithms and develop an appropriate balance between community, public, and industry interests.
Objective 4
Conduct behavioral experiments to build a body of behavioral evidence and assess the elements of the algorithm accountability benchmark to align design principles of algorithms and auditing processes with social concerns.
Objective 5
Develop software scoring tools and processes along the dimensions of the algorithmic accountability benchmark and employ them to assess public policy applications in criminal recidivism assessment and facial recognition.
The Benchmark
The goal of the accountability benchmark is to evaluate how well an AI system meets the regulatory standards and best practices. The criteria within the benchmark will be based on comprehensive legal analysis, broad community input, scientific evaluation of public preferences, and rigorous evaluation of real-world algorithms. They will be driven by concepts such as sociotechnical context, interpretability and explainability, transparency, nondiscrimination, robustness, and privacy and security. The benchmark will serve as a reference for auditing of AI systems.
Broader Impacts
The major mid-to long-term outcome of the CRASA project is the establishment of the accountability benchmark and the associated toolkit to be adopted by AI developers in a multitude of industries. Positive socio-economic and workforce impacts are also designed.
Socio-Economic
The engagement in the community-based participatory research framework will bring in stakeholders from a variety of backgrounds who will be able to adapt and spread the resulting ideas from this research.
- The Community Advisory Board members will serve as champions for our project to the communities they represent.
- The benchmark can be widely adopted by governments interested in leveraging the power of AI in an accountable manner.
- The benchmark is also designed to set expectations and standards for private industry.
Workforce
The CRASA project will engage in a three-pronged strategy for education and workforce development.
- Producing a set of educational materials that can be easily accessed by legal professionals and the general public on algorithms in public decision-making will be produced.
- Developing a multidisciplinary undergraduate/graduate course for students on the ethics of artificial intelligence that will be taught in several popular programs at the University of Houston.
- Training the next generation of scholars interested in accountability of AI in a collaborative atmosphere through hiring and training of research assistants.