CRS tells House Administration Committee AI fell short on bill summaries; requests $1.6 million and staff to build secure platform

A Congressional Research Service representative told the House Administration Committee that tests of AI tools produced far fewer usable outputs than expected and that CRS is seeking $1.6 million and five permanent staff to build capacity for responsible AI use.

"AI is advancing rapidly, and CRS sees its potential to streamline our workflow and enhance our service to Congress," the CRS representative said in an opening statement to Chairman Steele and committee members, while also warning that "the use of AI ... carries risks of outdated information, hallucinations, bias, and distortions."

The witness described CRS’s recent experimentation: over a two-year period CRS tested six different large language models, feeding approximately 1,000 bills into those models and producing roughly 3,000 AI-generated bill summaries. "Less than 3% of the summaries met CRS standards for accuracy, coherence, relevance, and objectivity," the representative said. That result, the witness added, shaped a plan to redesign parts of the bill-summary workflow so AI could safely augment staff work.

CRS said it is working in close partnership with the Library of Congress Office of the Chief Information Officer (OCIO) to develop governing principles and an implementation plan that identifies priority use cases that must both increase productivity and align with CRS values. The representative emphasized CRS’s five core values—confidentiality, objectivity, nonpartisanship, authoritativeness, and timeliness—and argued a "highly skilled human in the loop" is essential to maintain those standards.

As part of its fiscal 2027 budget proposal, CRS requested an increase of $1.6 million that would fund five permanent positions with data science and AI development expertise. The witness also described an OCIO request to support development of a dedicated, secure cloud environment "to develop AI solutions tailored to CRS unique requirements," which CRS said would be critical for testing and deployment.

The CRS representative further outlined planned staff training on AI ethics, generative AI, and prompt engineering, and said CRS intends to run a 12-month evaluation in fiscal 2027 comparing five AI tools—ChatGPT, Claude, Google AI, Perplexity, and Copilot. The overall message was that while AI can process large volumes of information quickly and help CRS experts, the current outputs often fall short of CRS standards and require human review and quality control before release to Congress.

The representative closed by reiterating CRS’s mission "to support Congress with objective, nonpartisan, authoritative, confidential, and timely research and analysis" and thanked the committee for the opportunity to testify. CRS did not announce a decision timeline beyond the proposed fiscal 2027 evaluation and budget request.