The Congressional Research Service is testing artificial intelligence on bill summaries to address the backlog, but fewer than 3% of the results meet the criteria, the director said Thursday.
To improve this, Director Karen Donfried told a House Administration Committee hearing that $1.6 million is needed on a recurring basis to upgrade the service to a more professional and sensitive model and staff five data scientists and an AI developer.
“AI is advancing rapidly, and CRS sees the potential for AI to streamline our workflows and enhance our services to Congress,” he said during the hearing. “However, the use of AI also comes with risks such as outdated information, illusions, bias, and distortion.”
CRS tested six models on about 1,000 banknotes over two years, and most failed to meet standards for accuracy, consistency, relevance and objectivity, Donfried said. The funding will likely help train the service on legislative data and upgrade it to a model that can input sensitive information, she added.
Lawmakers were “stunned” to hear about the unreliability of AI, and Rep. Julie Johnson, D-Texas, said, “I think it really highlights that as policymakers, we can’t have a false sense of security about the accuracy of AI. We have to put strong guardrails in place and be vigilant about the data that goes into these models.”
Throughout the hearing, Donfried reiterated that while AI could help CRS, it is not yet ready to replace staff. She said it’s important to keep tabs on “highly skilled people,” and what they need now is help with overviewing bills that haven’t made it to the floor.
“In the age of AI, I believe the role of CRS’s highly skilled workforce will become even more important as Congress seeks trusted sources of information in an increasingly unreliable information environment,” she said.
Congress is already using AI. A recent study by POPVOX Inc. found that AI-related spending is increasing across the House. Several tools are in use in both chambers, including ChatGPT, Gemini, and Copilot, and the Congressional Data Task Force recently announced that it is working on developing several AI-powered tools for staff to use.
On Wednesday, Rep. Anna Paulina Luna, R-Florida, said “most of the staff” were using AI after one of the revised briefs submitted to the House Rules Committee included the phrase “Claude reacted.”
Members also used AI during the hearing. He said Rep. Greg Murphy, R-N.C., “asked AI” if CRS should be replaced with AI, and the answer was “probably not.” Panelists discussed use cases for AI in the office and beyond, but some, like Virginia Republican Rep. Morgan Griffith, were concerned about potential partisan reactions from chatbots.
“I love AI. I think it has great potential,” Griffiths said. “We should be skeptical.”
Rep. Joe Morrell (D.Y.), the committee’s ranking member, said he’s not sure that using AI to speed up legislation is “of course a good thing, but that’s the world we live in.”
“We have not made the mistake of thinking that technology can replace a skilled and dedicated workforce. So, as I think you would agree, CRS’s investments in AI will make analysts more efficient and free them to do things that only analysts can do, and should never replace them,” he said.
Donfried said the service is developing plans for AI governance and prioritized use cases, as well as using AI to generate code for creating graphics and analyzing public comments on regulations. He said CRS and the Library of Congress’ Office of the Chief Information Officer plan to test five AI models in 2027: ChatGPT, Claude, Google AI, Perplexity, and Microsoft Copilot.
