The Congressional Research Service (CRS) recently tested artificial intelligence (AI) models to assist in generating legislative bill summaries, an effort aimed at alleviating significant operational backlogs. According to testimony by CRS Director Dr. Karen Donfried before the House Administration Committee on June 25, 2026, the two-year testing phase evaluated six large language models (LLMs) on approximately 1,000 bills. The results indicated that less than 3% of the 3,000 AI-generated summaries met the agency’s standards for accuracy, coherence, relevance, and objectivity (Banks, 2026; Will, 2026). This article outlines the factual pattern of the CRS AI testing initiative, the agency’s subsequent funding requests, and the broader context of AI adoption within the United States Congress.
Congressional Research Service AI Testing and Results
The CRS, an agency within the Library of Congress that provides nonpartisan policy and legal analysis to committees and Members of Congress, faces a substantial backlog of bill summaries. During the 118th Congress alone, more than 14,000 bills were introduced, exceeding the capacity of the 12-person CRS staff responsible for summaries (Bracken, 2026).
To address this backlog, CRS explored the use of generative AI. The testing process involved feeding legislative text into six different LLMs over a two-year period, generating approximately 3,000 AI-produced summaries. The output was largely deemed insufficient. Dr. Donfried noted that the models struggled with the specific analytical requirements of CRS summaries, which involve analyzing how a proposed bill would alter existing law, rather than merely summarizing the text itself (Banks, 2026).
The five core values CRS pledges to Congress — nonpartisan, objective, authoritative, confidential, and timely — formed the evaluative framework against which AI outputs were measured. Dr. Donfried testified that the tension between these values, particularly between speed and authoritativeness, presents a structural challenge for AI integration (Committee on House Administration, 2026).
Following the low success rate of generating complete summaries, CRS adjusted its approach to focus on discrete steps within the bill summary workflow. For example, the agency is exploring the use of AI to prioritize bills advancing to the House or Senate floor, thereby directing analysts to focus their efforts on the most urgent legislation (Committee on House Administration, 2026).
Library of Congress Fiscal 2027 Budget Request for AI
In response to the limitations of existing commercial models and the need to handle confidential legislative data securely, the Library of Congress submitted a $5.4 million budget request for Fiscal Year 2027 to develop a centralized, enterprise AI platform. Within this request, $1.6 million is designated specifically for CRS to hire five data scientists and AI developers and to acquire specialized, confidential models (Will, 2026).
The proposed AI Enterprise Platform aims to establish a secure environment under the Library’s control. The platform would focus on training two AI service models: one on legislative data for CRS, and another on bibliographic data and collections workflows for the Library Collections and Services Group (LCSG) (Library of Congress, 2026).
| Funding Component | Amount | Purpose |
|---|---|---|
| Library of Congress AI Platform | $5.4 million | Centralized, secure enterprise AI platform for data management and model training |
| CRS Specific Allocation | $1.6 million (recurring) | Hire 5 data scientists/AI developers; acquire specialized/confidential models |
Table 1: Library of Congress FY2027 AI Budget Request Components
Robert Newlen, the acting Librarian of Congress, testified before the Senate Appropriations Legislative Branch Subcommittee in May 2026 that without a proprietary, secure platform, the Library and Congress would “be left behind.” He also disclosed that he had spoken with OpenAI CEO Sam Altman regarding the Library’s initial efforts to use AI for applications serving Congress and the American public (Bracken, 2026). Newlen stated that he made clear to Altman that everything the LOC does with the technology must produce accurate, authoritative, and nonpartisan information for lawmakers.
In FY2027, CRS and the Library of Congress Office of the Chief Information Officer (OCIO) are planning to test five AI models: ChatGPT, Claude, Google AI, Perplexity, and Microsoft Copilot (Will, 2026).
The June 25, 2026 House Administration Committee Hearing
The hearing, titled “The Congressional Research Service and the Future of AI-Enabled Policy Analysis,” was held in Room 1310 of the Longworth House Office Building at 10:15 a.m. ET. Dr. Donfried was the sole witness (Committee on House Administration, 2026).
Committee Chairman Bryan Steil (WI-01) questioned Donfried on how CRS balances the speed of AI-assisted responses against the accuracy and depth required by the agency’s authorizing statute. Donfried stated that CRS’s five core values represent a “promise” to Congress and that the agency always marries those values in its outputs (Committee on House Administration, 2026).
Rep. Stephanie Bice (OK-05) asked whether CRS was considering building a custom LLM specific to legislative language. Donfried responded that the near-term focus is on building a workflow tool that can increase the volume of bill summaries produced, and that a custom LLM trained on legislative data would be a subsequent step contingent on receiving the requested AI platform funding (Committee on House Administration, 2026).
Rep. Julie Johnson (D-TX) stated that the low reliability results “really highlights that we, as policymakers, can’t have a false sense of security of the accuracy of AI.” Rep. Morgan Griffith (R-VA) expressed concern about potentially partisan responses by chatbots. Rep. Greg Murphy (R-NC) stated he “asked AI” during the hearing whether CRS should be replaced with AI, and the AI responded “probably not.” Committee Ranking Member Rep. Joe Morelle (D-NY) stated that CRS’s investment in AI “should make analysts more efficient, free them to do what only they can do, and should never, ever replace them” (Will, 2026).
Donfried stated throughout the hearing that AI is “not yet ready to replace staff” and that keeping a “highly skilled human in the loop” is essential. She also noted that CRS is using AI to produce code for making graphics and to analyze public comments on regulations (Will, 2026).
Broader Congressional AI Adoption and Policies
The CRS testing occurs alongside broader, though measured, adoption of AI tools across the legislative branch.
House and Senate Policies
Both chambers of Congress have established internal guidance for AI use, though these policies are not publicly accessible. The House Chief Administrative Office (CAO) issued “HITPOL 8” in September 2024, which outlines approved use cases categorized into four tiers: Generally Permissible, Requires Management Approval, Requires HIR Approval, and Prohibited. The policy mandates human oversight, prohibits the input of constituent Personally Identifiable Information (PII) into AI tools, and forbids using AI for “finalizing legislation” (Wilson, 2026).
The House policy conditionally approves tools such as OpenAI ChatGPT Pro, Anthropic Claude Pro, Google Gemini, and Microsoft Copilot. The Senate established a two-tier framework in October 2025, authorizing Microsoft Copilot Chat, Google Workspace + Gemini, and OpenAI ChatGPT Enterprise for use with official Senate data (Wilson, 2026).
| Chamber | Policy Issued | Approved Tools | Key Restriction |
|---|---|---|---|
| House | September 2024 (HITPOL 8) | ChatGPT Pro, Claude Pro, Gemini, Copilot | No constituent PII; no finalizing legislation |
| Senate | October 2025 (Tier 2 Framework) | Copilot Chat, Gemini, ChatGPT Enterprise | Tier 2 = authorized for official Senate data |
Table 2: Congressional AI Policy Frameworks (as of March 2026)
Microsoft Copilot Deployment
In late 2025, the House signed an enterprise agreement with Microsoft to provide 6,000 Copilot licenses to Member offices and committees. The rollout began in 2026, requiring offices to establish internal use policies prior to deployment. Speaker Mike Johnson (R-LA) announced the deployment at the annual Congressional Hackathon in September 2025, stating that Congress would “deploy artificial intelligence” across the institution (Nazzaro, 2025; Wilson, 2026).
AI Expenditure Trends in the House
A review of House of Representatives expenditure records by POPVOX Foundation indicated a rise in identifiable AI software spending in the first quarter of 2026. Matched AI-tool spending totaled $156,692.88 between January 2025 and March 2026, with the Jan–Mar 2026 quarter accounting for $70,984.87 — approximately 3.5 times the level of Jan–Mar 2025. OpenAI/ChatGPT accounted for approximately 72% of the matched total. Member offices accounted for roughly two-thirds of the identified AI spending (POPVOX Foundation, 2026).
| AI Tool Category | Total Spend | Rows |
|---|---|---|
| OpenAI / ChatGPT | $112,442.23 | 916 |
| Otter.ai | $19,882.66 | 87 |
| Anthropic / Claude | $13,395.62 | 38 |
| Descript | $9,179.38 | 91 |
| Other AI apps | $1,492.99 | 7 |
| xAI / Grok | $300.00 | 1 |
Table 3: House AI Tool Spending by Category, January 2025 – March 2026 (Source: POPVOX Foundation, 2026)
AI Use Incident During the Hearing
On June 25, 2026, Rep. Anna Paulina Luna (R-FL) stated that “most staff” use AI after one of her amendment summaries submitted to the House Rules Committee included the phrase “Claude responded” (Will, 2026).
Congressional Data Task Force
The Congressional Data Task Force, a working group within the legislative branch innovation ecosystem, has indicated it is developing several AI-powered tools for staff use (U.S. Government Publishing Office, n.d.).
Background: CRS and the Library of Congress
The Congressional Research Service was established by legislation passed in 1914, signed into law by President Woodrow Wilson. Originally called the Legislative Reference Service, it was renamed the Congressional Research Service in 1970. The director of CRS serves by appointment of, and under the general direction of, the Librarian of Congress (Library of Congress, 2024).
Background: Dr. Karen Donfried
Dr. Karen Donfried was appointed Director of the Congressional Research Service effective September 23, 2024, by Librarian of Congress Carla Hayden. She brings more than three decades of experience in foreign policy, having previously served as Assistant Secretary of State for European and Eurasian Affairs, President of the German Marshall Fund of the United States, a member of the National Security Council, and as National Intelligence Officer for Europe on the National Intelligence Council. Prior to her State Department service, Donfried served as an analyst and specialist in European Affairs at CRS for 10 years. She holds bachelor’s degrees in government and German from Wesleyan University, a Magister from the University of Munich, and a Master of Arts in law and diplomacy and a Ph.D. from the Fletcher School at Tufts University (Library of Congress, 2024).
Background: Dan Spokojny and fp21
Dan Spokojny, who shared the FedScoop article on LinkedIn, is the founder and CEO of fp21, a nonprofit, nonpartisan think tank dedicated to transforming the processes and institutions of U.S. foreign policy. He is a former U.S. Foreign Service Officer and foreign policy legislative staffer in Congress, currently a Ph.D. candidate in political science and international relations at the University of California, Berkeley. He teaches courses on foreign policy decision-making at Johns Hopkins School of Advanced International Affairs and the George Washington University Elliott School of International Affairs (Spokojny, n.d.; fp21, n.d.). His LinkedIn profile is at linkedin.com/in/spokojny.
References
fp21. (n.d.). Executive Team. Retrieved June 29, 2026.
Spokojny, D. (2026, June 27). LinkedIn Post. LinkedIn.
AI generated