A LinkedIn post sharing a FedScoop report describes the Congressional Research Service’s early use of artificial intelligence to help address its bill-summary workload. The story is less about an imminent automated authoring system than a legislative-technology pilot that exposed a large gap between general-purpose model output and CRS’s standards for authoritative, nonpartisan analysis. 1 2
FedScoop reported that CRS Director Karen Donfried told the House Administration Committee that the service had tested six models over two years on about 1,000 bills and that fewer than 3% of outputs met its standards for accuracy, coherence, relevance and objectivity. That specific percentage is reported testimony. The committee’s official account separately confirms that Donfried said the six large language models were not successful at producing bill summaries and that CRS is reassessing the workflow task by task. 2 3
A narrower use case than automatic drafting
At the June 25 hearing, Donfried said CRS was not looking to have AI produce bill summaries “out of the gate.” Instead, she described breaking the process into discrete steps. One potential use would help analysts identify which bills are likely to reach the floor, so staff can focus their attention where it is most needed. 3
That distinction matters. CRS is Congress’s nonpartisan research arm, and its official account says the service’s core commitments include objectivity, authority, confidentiality and timeliness. Donfried told lawmakers that speed cannot come at the expense of accuracy. 3
“Maybe AI can’t help us produce a bill summary out of the gate,” Donfried said, according to the committee’s official hearing summary. “What are all the discrete steps that our colleagues undertake to produce that bill summary?” 3
| Issue | What the record supports | Important limit |
|---|---|---|
| Pilot results | FedScoop reported that Donfried cited testing of six models on roughly 1,000 bills, with fewer than 3% meeting CRS standards. 2 | The reviewed official committee summary confirms the models were not successful, but does not independently reproduce the exact percentage or full test methodology. |
| Near-term use | The committee summary says CRS is considering AI support for workflow steps, including prioritizing bills likely to reach the floor. 3 | It does not describe a deployed system that independently writes CRS bill summaries. |
| Specialized model | Donfried said a Library platform trained on legislative data could eventually position CRS to produce bill summaries. 3 | This is contingent on funding and future development, not confirmation of an operational custom model. |
| Staffing and platform request | The Library’s FY2027 request includes $1.622 million and five FTEs for CRS, within a broader $5.446 million AI-platform request. 4 | A budget request is not an enacted appropriation or an operational capability. |
A budget request for secure, domain-specific work
The Library of Congress’s fiscal 2027 budget justification describes a broader proposal for an AI enterprise platform. It would use a secure environment to test and evaluate AI services, initially with legislative data and then with bibliographic data. The request says the platform could support domain-trained models, improve quality control and operate under Library oversight. 4
For CRS, the proposal requests $1.622 million and five full-time positions: one GS-15 data scientist and four GS-14 data scientists. The request also includes funding for AI productivity software, cloud hosting and training. The full enterprise-platform proposal totals $5.446 million and seven positions across CRS and the Library Collections and Services Group, along with Office of the Chief Information Officer support. 4
Those figures are a fiscal 2027 request. They should not be treated as enacted funding, a completed procurement or proof that CRS already operates a confidential, specialized legislative model.
The human-in-the-loop conclusion
The most consequential lesson from the hearing was not that Congress has rejected AI. CRS is testing possible uses for research, analysis and workflow management. The official budget request explicitly describes experimentation, evaluation and responsible deployment in a secure, controlled environment. 3 4
But the initiative is designed around human review rather than replacement. Donfried’s testimony, as reported by FedScoop and summarized by the committee, framed skilled analysts as the source of the trusted judgment Congress needs when models can produce outdated information, hallucinations, bias or distortions. 2 3
For now, the public record supports a narrower conclusion: CRS is pursuing AI as a way to improve triage, research tools and analyst capacity. It has not established a reliable automated system for producing the bill summaries that lawmakers rely on.
References
AI generated