CDT report – The Federal Government’s Power of the Purse: Enacting Procurement Policies and Practices to Support Responsible AI Use

Related Organizations

Table of Contents

• Executive Summary 4
• A Brief History of Federal AI Procurement Governance 11
• Current Federal AI Procurement Priorities 15
• Unique Risks and Governance Challenges of AI Procurement 18
• A. The federal government lacks a common definition of AI. 19
• B. Individual federal agencies lack capacity to properly evaluate vendors and rely on vendors, including bias identification and mitigation. 21
• C. Individual agencies lack necessary resources to ensure appropriate human oversight and intervention. 24
• D. Cross-government policymaking and resourcing bodies lack necessary expertise, and struggle to provide sufficient detail to support individual agencies. 25
• E. Procurement processes do not typically accommodate the continuous monitoring and post-award management required for responsible AI. 26
• F. AI raises particular difficulties for individual agencies in identifying and addressing contractor performance issues. 28
• G. Vendors and federal agencies are unclear about redress process responsibilities. 29
• Recommendations 30
• A. Incorporate responsible AI considerations at the acquisition planning stage. 31
• B. Include references to AI risks in the Federal Acquisition Regulations (FAR). 35
• C. Increase attention and capacity to perform pre-award vendor evaluation and post-award vendor monitoring through guidance and standardization. 36
• D. Clarify and strengthen transparency, reporting, and oversight requirements and provide guidance to facilitate compliance. 40
• E. Increase federal workforce capacity to ensure agencies are prepared to evaluate and manage risks throughout AI procurement. 48
• Conclusion 51
• Appendix A: History of Approaches to Federal AI Procurement 52
• 2018-2019 Federal Actions 53
• 2020 Federal Actions 55
• 2021 Federal Actions 59
• 2022 Federal Actions 59
• 2023 Federal Actions 62
• Appendix B: AI Risks Compared To Traditional Software Risks 65
• Appendix C: Glossary Of Acronyms and Abbreviations 68