Why It Matters

The Congressional Research Service (CRS) revealed that less than 3% of the 3000 AI-generated bill summaries met its accuracy standards, exposing a fundamental challenge as lawmakers push the agency to adopt artificial intelligence tools. The House Administration Committee held a hearing on Thursday, June 25, and CRS Director Karen Donfried warned that AI cannot yet replace human expertise in legislative analysis, even as congressional staff increasingly expect faster, AI-enhanced research.

The Trump Administration has aggressively promoted AI development through executive orders and legislative frameworks, yet the CRS testimony exposed a widening gap between what congressional staff want and what AI can reliably deliver.

The Big Picture

The Congressional Research Service is Congress's nonpartisan research arm, employing approximately 400 policy analysts and attorneys. For two years, CRS tested six different large language models, generating 3,000 AI-created bill summaries. The results were sobering: less than 3% met CRS standards for accuracy, coherence, relevance, and objectivity.

This hearing comes as major technology companies, startups, and think tanks rush to develop AI-enabled policy synthesis and legislative analysis tools. Some AI chatbots pull information from the internet without providing citations, a critical flaw for legislative work requiring complete sourcing.

Donfried testified that the agency is pursuing a cautious path. CRS requested a 1.6 million dollar budget increase for fiscal year 2027, including five permanent staff members with data science and AI development experience. The agency plans a 12-month evaluation in fiscal 2027 to compare five AI tools: ChatGPT, Claude, Google AI, Perplexity, and MS Copilot.

The hearing fits into a broader Trump Administration push on AI policy. The White House released its national policy framework for artificial intelligence March 20 as formal legislative recommendations.

What They're Saying

  • Karen Donfried, CRS Director: "Through the responsible use of AI, CRS will strive to benefit from what AI tools can do best to provide even better support to you. Chatbots may provide you with the fastest answer to a question you have; CRS will provide you with the answer you can trust."

The challenge is structural. Congressional staff are beginning to expect richer, more tailored analysis to support their work as AI capabilities advance. Yet CRS staff expertise, knowledge of the legislative process, and institutional memory remain the service's most valuable assets. Human oversight is essential to review, double-check, and confirm sources for any AI-generated analysis.

CRS is developing principles to govern the responsible integration of AI tools into its workflow and an implementation plan to identify priority use cases for AI development. The agency evaluates AI-enhanced tools based on whether they increase productivity in developing products for Congress and align with CRS values. CRS has offered training to staff on topics ranging from ethics to prompt engineering.

Political Stakes

For the Administration

The hearing exposed a gap between policy ambitions and technical reality. The White House wants federal agencies to modernize through AI adoption, yet CRS's experience suggests that government agencies cannot outsource their core functions to commercial AI tools without losing control over data security and institutional independence.

For Congress

The stakes involve the quality of legislative decision-making. If CRS cannot reliably use AI to accelerate bill analysis, congressional staff may turn to commercial tools like ChatGPT that lack citations and institutional oversight. This could degrade the information environment on which lawmakers depend.

The Other Side

Critics of excessive caution argue that CRS is being too risk-averse. The First Branch Forecast, a congressional analysis platform, raised a structural concern about whether AI tools should be owned by the legislative branch or if it is appropriate to have data and technology reside with major AI companies.

This reflects a genuine policy debate. Free versions of AI chatbots are not closed systems and cannot be used for confidential congressional material. CRS can only create a closed system with higher-tier versions of AI models. Building proprietary AI infrastructure within Congress would be expensive and technically complex, but it could protect legislative confidentiality while enabling faster analysis.

The Bottom Line

Congress wants AI to move faster, but CRS says it cannot move faster without sacrificing accuracy—and Congress depends on that accuracy more than it depends on speed.

Access the Legis1 platform for comprehensive political news, data, and insights.

Spot something wrong? Report an issue with this article