CRS, AI, and Congress's Information Problem
New tools could make "Congress's think tank" more efficient, but questions about its effectiveness run much deeper
CRS OVERSIGHT
The Committee on House Administration will examine technological modernization at the Congressional Research Service in a hearing Thursday at 10:15 AM. The focus will be the “future of AI-enabled policy analysis,” which Acting Librarian of Congress Robert Newlen (who served in many roles at CRS over his distinguished career) has discussed as a major area of interest in reducing some of the workload at CRS. As of today, witnesses haven’t been announced, but we imagine CRS Director Karen Donfried, who was appointed to that role in 2024, will be testifying.
The role that technology should play at CRS is deeply tied up with CRS’s history, its management culture, and Congress’s research needs in the era of AI. These factors are worth exploring at some length before any consideration of technological modernization to optimize CRS for the new information era.
The intended role for CRS
The role of the Congressional Research Service was set in statute through the Legislative Reorganization Act of 1970. This legislation reorganized its predecessor, the Legislative Reference Service, which had been established inside the Library of Congress in 1914 to support reference questions from Congress. Over time, members needed support analyzing legislation and policy, a role to which LRS was not suited.
The Legislative Reorganization Act created a larger office with a renewed mission to provide both chambers with high-quality analysis of public policy and legislation, not only reference assistance.
Its new director was authorized to hire specialists who could, as the legislative text described, analyze legislative or agency action proposals, determine the advisability of their enactment, project their results, and present alternative methods for the same result.1 It also empowered CRS specialists to produce research reports on their own initiative that could inform congressional action on particular problems and provide confidential guidance to individual offices. It also mandated the production of summaries of bills reaching the floor.
The design of CRS reflects the centrality of committees at this point in congressional history. CRS was mandated to prepare a list for each congressional committee at the beginning of each Congress of subjects and policy areas they may want to pursue. CRS staff would routinely go on long details with committees, and staff would build close and enduring relationships with members.
Although materials requested by congressional offices would remain confidential, CRS staff were not shuttered away in the Library. They were expected to be experts on their topics and routinely would attend conferences and publish scholarly works like academic professionals. CRS nonconfidential reports were frequently made publicly available, and CRS published a magazine, the CRS Review, to provide a digest of its policy research and analysis that was publicly available through the GPO store. All of this was intended to improve discoverability of its products by staff and improve the information ecosystem surrounding Congress.
Political change and the onset of “neutrality”
Changes in the power dynamics within Congress began to affect CRS beginning in the 1980s. Leadership in the House was becoming more powerful and some of the work of legislating was being moved out of committees. The Republican Revolution after the 1994 midterms accelerated the shift to a leadership-centered model of House organization, creating challenges for offices and organizational structures that provided independent policy analysis like the Government Accountability Office, Office of Technology Assessment, and congressional caucuses among others.
The Gingrich revolution promised to cut the size of government and started with the legislative branch, pledging a one-third reduction in the number of committee staff. Speaker Gingrich cut spending for committee staff in excess of 20%, eliminated the Office of Technology Assessment, cut staff at GAO and CRS by more than 20%, defunded the Joint Committee on Printing, and destroyed the operating model for legislative service organizations. Indeed, there are anecdotes that suggest there was a moment where Gingrich was considering eliminating CRS under the theory that partisan think tanks could provide Congress with the same kind of policy support. Some Republicans defended these offices and agencies against dissolution or decimation, but leadership generally got its way. Whatever Gingrich’s intent may have been, the result was to move expertise and power to House leadership and away from the committees and rank and file.
In response to these threats, CRS management began changing how CRS engaged with Congress. Former senior specialist Harold Relyea described CRS managers as beginning to minimize CRS’s public profile, particularly in talking to the press. In 2004, CRS issued a memorandum directing that staff, when speaking publicly, must adopt a position of “neutrality.” This poorly defined directive was immediately criticized by CRS staff as requiring staff to avoid reaching an expert opinion–to “take no positions and reach no conclusions.”
A particular moment of inflection was a hullabaloo in 2006 when CRS experts provided their expert views in a public forum. In this case, one analyst accurately described his CRS report to a journalist on the limitations of protections to whistleblowers, prompting blowback from a committee chair. The CRS director apparently ordered the analyst to apologize to his division manager or be terminated. He ultimately ended up reassigned elsewhere in the Library, performing similar duties.
There were other such incidents. In 2009, CRS fired former military commissions prosecutor Morris Davis for expressing his personal views in the Wall Street Journal criticizing the Obama Administration’s policy on detainees. In 2012, CRS withdrew a report by analyst Thomas Hungerford on the effects of various tax policy proposals when some members of Congress didn’t like his conclusions. It belatedly reissued that report with the conclusions intact and the insertion of watered down language.
According to former CRS analyst Kevin Kosar, “there was a change in the quality of our internal reports” as a result. How serious was the problem? “Reaching conclusions—the job of an expert—became verboten. In fact, for a time, CRS analysts were told not to end their reports with a section titled ‘conclusion.’”
The problem remained pervasive and ingrained in management practices. In 2019, CRS’s union pointed to agency staff engaging in self-censorship out of fear of management reaction. “Management fear of Member objections overrode the nonpartisan, expert analysis and judgment of their analysts,” the union president testified to CHA. “Authors see this trend and scale back analytically ambitious products on their own. In some cases, analysts are prevented in CRS’s review process from synthesizing new perspectives on issues, and are told to instead focus only on what others have said. The chilling effect influences the design of new projects and the review of draft products by peers, supervisors, division management, and the CRS Review Office.”
Pressures from Congress and its own management subsumed much of the initial intent behind the creation of CRS as a wayfinder and analytic spare brain for the House and Senate. Reports became less analytical and more descriptive and staff were discouraged from engaging with their academic peers. Much of its mandated work, meanwhile, was being largely ignored, like those lists of issue areas for committees to consider exploring. As Kosar noted in Washington Monthly shortly after leaving to enter the think tank world in 2015, what had been “set up as a professorial policy analysis shop, now spends a lot of time answering constituent requests.” It also had become “a remarkably risk-averse organization, tenaciously clinging to old ways and habits.”
CRS in crisis
When the COVID-19 pandemic hit Congress, calcified management and decades of underinvestment in technology placed additional strain on the agency. Robert Newlen testified in 2024 about how “there was a lot of concern about IT support in our divisions and offices and frustration with getting things resolved very quickly.” There were many examples of basic tools not working properly, including an antiquated version of Microsoft Word, Zoom accounts that would cut out mid-meeting with staff, cell phone dead zones that made it impossible to call staff back, and failed attempts to procure data resources critical to its work.2
Managerial problems under the direction of Mary Mazanec had become acute. In a 2023 CHA oversight hearing, CRS’s union had expressed deep concern with CRS’s leadership, calling it a “five-alarm fire,” and pointing to high attrition, low morale, poor communication, and a lack of trust and confidence in senior leadership. In the American Law Division, staff turnover had hit 19% over a three year period without reaction. When asked at the hearing what was one thing the committee could do to make CRS best serve Congress, Kosar answered, “By my assessment, most policies that affect the service of CRS to the Congress come from the front office of the agency. If you want to effect the most change in the most immediate way, you change leadership.”
Two months after this hearing, Mazanec departed under a cloud. Newlen, who started his career at CRS, had served in positions throughout the Library, and stepped down in 2017 to run a private foundation, was brought back and appointed acting CRS director. He testified before the House at its most recent hearing on CRS, in 2024, supporting the modernization of CRS and the use of federal data. Newlen once described himself as a loving critic of CRS; he now serves as the highly respected acting Librarian of Congress. In the wake of the hearing with Newlen, the Washington Monthly published a guide to reviving CRS, prescribing remedies to strengthen the agency and arguing it is an essential component of a strong Congress.
The State of CRS
What we saw at the 2023 hearing was that Congress had a CRS management team driven into a decades-long defensive crouch and a workforce driven by irreconcilable requirements to provide useful analysis to Congress and to avoid saying anything that could offend them. Moreover, the technology available to CRS staff was antiquated and many of their statutory requirements no longer fit Congress’s needs. On top of all of that, CRS leadership at that time was utterly incapable of meeting the challenges before them and created a culture of fear among CRS staff and learned helplessness.
How is CRS doing now? CRS and the Library’s annual reports, and their testimony before appropriators, provide some insights. Generally speaking, it is difficult to know. So far, our understanding is anecdotal and confidential, except for the public-facing testimony of CRS’s union. That’s why this hearing is so timely.
What we can gather about CRS activities from its annual reports, which AGI has collected, are the metrics of how busy the place is (very!), but little more. For example, in its FY 2024 report, the most recent one available to the public, CRS reported 75,000 custom responses to requests from congressional offices, publication of about 1,000 new reports, 1,700 report updates, 6,500 bill summaries, 2,500 confidential memos, 1,850 in-person consultations, and nearly a million hits on the crs.gov website. These figures have been consistent over the last decade.
In 1970, Congress started asking for various types of products, and ever since, CRS has produced them. But are they useful and used by staff?
Congress and the Information Environment
To answer that question, we must understand how congressional staff seek and use information to solve problems. This is a much larger topic than we have space for here, but in brief, any answer would depend on how experienced the staffer is, where they work, how quickly they need an answer, and their purpose: objective research, confirming an opinion, making a public communication, or satisfying a request.
With a few exceptions, no one has placed researchers in congressional offices to watch what they really do. Relevant literature on this topic is out of date or limited to specific domains (like science). In the US, congressional staff are discouraged from answering surveys, making reliable data scarce. CRS conducted internal staff climate surveys,3 but as testimony from CRS’s union in 2019 indicated, leadership tampered with the report and the version posted on CRS’s website was missing some of its recommendations.
If CRS were interested in such user-centered research, a 2017 effort by the British Parliamentary Office of Science and Technology entitled “The Role of Research in the UK Parliament“ could be a model. It surveyed 157 people, providing context for how reference services work. It was not, however, a time-and-motion study.
We think approaching reference and analysis work from the perspective of a variety of staff backgrounds is essential. For our BillMap project, which sought to build a tool to provide contextual information to staff trying to evaluate legislation, we created archetypes for the various kinds of users we expected—a personal office legislative assistant, a senior professional staffer in a committee, etc.—and conducted in-depth interviews to understand how they engaged with their information environment.
We need to understand that staffers are trying to learn about and solve a problem under tremendous time pressure. They are not interested or able to dig around CRS’s website, then go to GAO’s website, then the website for the national academy of science, partisan think tanks, etc. Staff will go to CRS’s website, as the data shows, but this is not their primary discovery path. (Indeed, going to CRS’s website becomes even more disfavored when staff realize that CRS doesn’t make all of its reports available on its website—that more of them are discoverable on the open web than CRS’s intranet.) Furthermore, CRS reports can be quite lengthy, requiring a time commitment to discover if it’s even useful.
Written products are one type of user experience for congressional staff. They also can schedule time for consultation with a CRS analyst. Even these experiences can be variable, as not all researchers have the same level of expertise. That challenge may have gotten worse as CRS’s union in 2019 pointed to the hiring of less experienced CRS analysts in recent years.
A 2019 National Academy of Public Administration report classified the various kinds of support needs that a staffer may have — quick turnaround, networking, consultative, short term report, technology assessment, and horizon scanning — and recommended that Congress have a single place staffers can go to that is an intermediary between all the sources they may wish to use but is expert at providing that information in the ways that Congress can use. CRS could be that intermediary, if properly reformed and retooled. As this hearing will undoubtedly touch upon, artificial intelligence could be, too. But even if it could, should it be in an institution driven by politics and collective problem solving by humans?
CRS and AI
We have no doubt that AI technology can be a productivity booster inside CRS, freeing analysts to have more time to the type of work befitting their qualifications. CRS stood up an artificial intelligence working group in May 2023 to identify “best practices and operational processes necessary for the adoption of AI technologies” and share awareness of AI use across the Library. Acting Librarian Newlen has mentioned in Legislative Branch Appropriations Subcommittee hearings of interest in CRS developing its own Large Language Models with congressional data and it will be interesting to hear the office’s current perspective on use cases.
To our minds, however, a more pressing question is not how AI can accelerate the delivery of work products that CRS already produces, but whether CRS products and services truly are meeting users’ needs in the current information environment context. Can AI accelerate the delivery of analyses built upon expertise from across the various legislative branch agencies and synthesized with official data from congressional databases (like committee reports) and beyond?
And, if AI can do that work directly, or at least do some of that work, what risks, biases, dangers, and costs is Congress taking on by putting front and center an automated synthesizing tool that is essentially a black box with the capability of hallucinating? What exactly do we mean by AI? Are we talking about deterministic tools like natural language processing, or probabilistic tools like small or large language models? Moreover, should these tools be owned by the legislative branch or is it appropriate to have data and technology also reside with the major AI companies?
It is our view that Congress has for decades underinvested in the tools and techniques that don’t require probabilistic AI.4 Not all of CRS’s reports are digitized and made available online—or even available directly to congressional staff—even as the older reports are relevant today.5 Not all of Congress’s committee reports are digitized and available online, a project that won’t be completed for at least another four years. Websites like Congress.gov should be rethought based on the different kinds of users and their use cases, providing contextualized, timely data in user-friendly interfaces. The Library and other information intermediaries should focus on delivering information to staff at the time that it’s needed, not following procedures that deliver “authoritative” information long after it has gone stale. (We acknowledge that major improvements have been made to aspects of Congress.gov over the last 15 years, and the welcome efforts to engage with internal and external stakeholders, but we also acknowledge that not all components of the Library are rowing in the same direction.)
This doesn’t mean there isn’t room for AI. It can speed up the discovery and contextualization of information. But for it to be deployed to the best effect, it needs authoritative, complete, comprehensive, clean, well-organized data that has been digitized and made available in bulk, through APIs, or via Model Context Protocol (a standard that lets AI applications connect to external datasets).
In the context of CRS, analysts should be able to search across the entirety of materials generated by the legislative branch to be able to find information that answers questions by congressional staff. Because they are experts, they should be empowered to direct the creation of and to review, refine, and vet automated AI responses and to compare it against their knowledge and experience to make sure that any response they give to staff is as close to timely, complete, and reliable as humanly possible. A human should still be in the loop for analytical products. And for finding aids, AI can facilitate the surfacing of information for agency and congressional staff, as well as the general public.
Moreover, AI can facilitate the production of information assets that meet the needs of congressional staff. Maybe the product shouldn’t be a report, but a dashboard. Or a dataset. Or a model. Or a series of interactive graphs. Or a podcast.
This question of how to do all this is in many ways a managerial, not technical one, touching on how CRS leadership structures feedback mechanisms from users, approaches and implements innovative ideas from staff, and measures its impact on serving Congress’s needs. It is also a financial one, dependent on whether Congress provides sufficient, sustained funds to make this transformation occur. The woeful underfunding of CRS in particular, and the legislative branch in general, over the last few decades makes it hard to invest in new approaches when the agency is on life support.
CRS and Congress
We suggest the Committee on House Administration think about that context when discussing technological modernization because faster delivery of materials that are not fully aligned to congressional use cases or can be obtained elsewhere would underdeliver on what AI promises to provide: a comprehensive single touchpoint for congressional staff to get answers to their questions regardless of where the data resides.
To attain that promise, congressional stakeholders have to guide CRS out of its risk aversion. Even with a new agency director and a capable Librarian of Congress, some of the old managerial culture (and dare we say personnel) who punish innovation, rely on process over outcomes, and prevent problems from bubbling to the surface for resolution, remain in situ. Addressing the needs of congressional users must come first.
Other parliamentary research organizations can be a helpful model, or at least a comparator. Some do a much better job in public outreach, engagement with outside experts, and explaining the limitations of their research efforts. The United Kingdom Parliament’s Parliamentary Office of Science and Technology is the gold standard in public openness and two-way engagement, and the nonprofit Hansard Society regularly publishes podcasts on the workings of Parliament. The European Union’s Parliamentary Research Service publishes their reports, holds online briefings, uses social media, and even has an app.
Indeed, parliaments worldwide are struggling with how to incorporate AI, which we wrote about recently. They are trying to be thoughtful in improving their capacities without creating expensive dependencies and undue risks to their work.
CRS should be in dialogue with these and other parliaments, as well as with congressional and public stakeholders. Yes, public stakeholders. The public has a lot to say about the provision of congressional information. They are in routine conversation with congressional staff, and CRS should be in regular conversation with them. This 2020 civil society report is a roadmap of how some legislative information services could be improved.6 CRS does have a public mission, despite protestations by some to the contrary, and CRS should understand that making much of its information publicly available is how it feeds the congressional information ecosystem to help members and staff understand and solve their problems.
CHA may have to consider charter changes for CRS to accomplish some of the managerial and workplace culture change described here. The rapid development of AI technology is a great opportunity to help ensure that Congress has access to reliable, non-partisan, expert policy advice that synthesizes all available information in a timely way that meets congressional user needs.
We do not expect this hearing to be contentious, but curious. We are excited to see what’s in the future for CRS and for Congress.
APPROPRIATIONS
Once dead in the water, appropriations have restarted in the Senate. The Appropriations Committee will hold a four-part markup on Thursday beginning at 10:00 AM in 106 Dirksen. On the schedule is Legislative Appropriations, which we will do our best to summarize.
ODDS AND ENDS
Democracy Awards: The Congressional Management Foundation has announced the winners of its 2026 Democracy Awards, given in six categories to a Republican and Democratic office. The awards also honor two House and Senate chiefs of staff of the year.
Faction: Politico handicaps five members who may lead the House Freedom Caucus next Congress, in which many current members will be out of office.
Agenic reporting: The Department of Defense is using AI to produce some of the many, many mandated reports it owes to GAO. The legislative branch is contributing to the problem as the number of mandates has soared well past 1,000 and DoD staff have to come through legislation to discover new ones.
Not automatic: The Center for Budget and Policy Priorities’s Richard Kogan and Sam Berger assess the Lankford-Hassan proposal to avoid government shutdowns through an automatic continuing resolution and find it lacking. In particular, it strengthens the power of the executive branch because “The President could repeatedly and indefinitely veto any appropriations bill that provides more funding than he supports or includes legislative language aimed at ensuring he is complying with congressional intent in funding matters, without concern that it would result in a disruptive shutdown of key government operations.”
Baked in: The International Institute for Democracy and Electoral Assistance has a new report exploring “how rules on agendas, debate, discipline and legislative scrutiny can be turned into tools for silencing opposition, concentrating power and weakening democratic accountability” across parliaments.
Strong responses: Agora shares good practices for parliaments in addressing violence directed against female lawmakers.
This sketch of CRS draws heavily from three excellent articles written by former CRS analyst Harold Relyea, which are (as far as we know) the best history of CRS. They are: “Across the Hill: The congressional research service and providing research for congress—A retrospective on origins” (2010), “Across the hill: The congressional research service and providing research for congress — Considering the future,” (2012), and “Across the Hill: The congressional research service and providing research for congress — A retrospective on personal experience” (2012). In addition, Steven Aftergood of the Federation of American Scientists tracked CRS closely, covering its activities for his excellent blog, Secrecy News.
Newlen testified how CRS and the Library were able to resolve the problem he had identified. We note that they were addressed after Mazanec’s departure from the agency.
The surveys are focused on staff perceptions of their work environment. The issues of work environment and for whom CRS analysts are working are interconnected.
Congress has massively underinvested in the legislative branch writ large. We are looking at this issue through the lens of CRS, but this is an across-the-board problem.
This is not just an underinvestment problem, however. For years, CRS management has actively resisted and undermined this kind of modernization. Most recently, however, there has been a thaw, with some in management more open to embracing these approaches.
Some of the issues on the roadmap have been addressed since 2020. Those reforms were undertaken largely at the direction of Congress.