Didn’t you get the memo?
Last week a House committee hearing looked at AI assistance in the work of the Congressional Research Service (CRS). CRS is effectively a Congressional agency tasked with providing analysis, data, and in-person consultation to Congress-at-large, often prompted by individual (confidential) requests. But their most widely read outputs are the official bill summaries on Congress.gov.1
They’re straightforward, fairly neutral rundowns of a bill’s text used as the starting point for most of the public (and even Congressional offices) to ingest a proposal’s content. Malcontents can quibble, but there’s no implied editorializing or normative subtext to these summaries. And no analysis of budget or other impact.
And yet for some time now, there’s been a long and growing lag between a bill’s introduction and when it’s summary appears.
Sen. Blackburn’s S. 233 went over 520 days from intro to posted summary (not coincidentally the day it reported out of committee). In the AI space, Sen. Young’s CREATE AI Act is without a summary for about two months, but Rep. Obernolte’s version has one—though introduced earlier and now passed by committee (that fact also not yet reflected).
Some gap is forgivable — Congress introduces lots of bills,2 some very long and complicated. Even texts meant to be duplicative (“companion” in leg.-speak) require review before copy-pasting summaries.3 Also the chambers have their own steps to officially record and distribute official text before seen by CRS, and committees making changes are another bottleneck. But it hasn’t been getting better as you’d think with better tech.
In testimony on The Congressional Research Service and the Future of AI-Enabled Policy Analysis, the CRS Director noted:
[U]sing six different Large Language Models (LLMs)[] for drafting bill summaries…[they] failed to meet CRS standards for accuracy, coherence, relevance, and objectivity (less than 3% of 3000 summaries were acceptable).4
I’ll stipulate “CRS standards” are robust based on my direct and extensive experience.5 Regardless of characterization of the standards, 3% does not account for all the {null set} summaries. That is 3% can be accurate within the sample, but a human CRS summary is only any good if it exists, which it does not for much of legislative text.6 An ungenerous phrasing would be that all the missing summaries are 0% accurate.
If you wanted to evaluate a person’s productivity to, say, dig a tunnel by driving steel versus some new mechanized machine, you would need to measure along two dimensions: yes, the length of mountain bored in a given time, but also ability to dig accurately where you intend.
[data]
So granted, a 3% accurate summary isn’t much more useful than a blank page, likely less. But the key is whether it can be applied to speed up the “0%” (i.e., blank summaries) stack. The testimony doesn’t describe the methodology but with such a large sample I assume the AI outputs were compared to existing human-written summaries, not new ones. It was also done in FY24 so no doubt excludes an incredible amount of improvement and useability.
Going forward she notes:
in early fiscal 2026, CRS contracted to develop detailed requirements, technology options, and an implementation plan (including cost estimates and efficiency gains) for the integration of AI-enhanced tools into each step in the bill summary workflow. Based on that work, CRS has a prioritized implementation plan for seven unique AI use cases across the bill summary workflow that, if fully implemented, is expected to provide a significant increase in the number of bill summaries produced by current staff.
Breaking down the production workflow is the right idea. And they should be commended for the work thus far so I don’t want to be entirely hard on them. But they’ve got a lot of catching up to do.
Institutional roadblocks
There are however binding constraints outside of their immediate control slowing their progress.
The “IT guys”
Any legislative office - personal office, committee, CRS, etc. - is restricted in their access to software and other IT assets. The Sergeant-at-Arms (Senate) and Office of the Chief Administrative Officer (House) are responsible for whitelisting vendors and products, including today’s most popular LLMs. Entities within Congress for the most part can’t install or use applications, sometimes even visit certain websites, unless they’ve been cleared. Approved products might include modifications or fixed settings reflecting internal policies.
That some entity does this is understandable. Most of Congress is networked, and sensitive/classified or other high-value information exists throughout. Even non-sensitive information is valuable, and an aggressive DDoS or wiper attack could effectively shut down the country’s Article I branch.
Though it does mean a forward-thinking office is not at liberty to move at it’s own speed.7 Optimally balancing safety with speed (and other dimensions) isn’t easy. As I’ve written elsewhere when it comes to cybersecurity the incentives are biased against security.
Budget
Ever the case, and always boring, but lack of funding is a meaningful barrier here. And I don’t just mean for software licenses. Upstream at the IT evaluation itself, additional efficient and expert staff will accelerate evaluation, allow for more specialization - and because they have the most direct contact with vendors - tailor applications and use cases for legislative work.8
This applies to the legislative staff proper as well. I’ve seen first hand new capabilities actually made available to offices, but without any capacity to scale the learning curve let alone integrate in their own work structure. A super-fast security evaluation doesn’t mean anything if nobody takes advantage. For this and a bunch of other reasons, Congress needs more staff, a point I’ve stressed repeatedly, including in testimony before this very same committee.
In truth, to the degree staff adopt use of new tools and apps, it’s through private use, like at home or on their own smartphone which is most of the time used for work anyway. But again, the scaling up of overall efficiency of Congress only works if it’s widely interoperable. An agriculture L.A. might figure out a way to speed up their own work, which is great, but it’s ad hoc and can’t apply widely.
Norms
As you can imagine, as its a debate happening across schools and offices nationwide, it’s unclear in Congress what the “rules” are with respect to AI. Each Senator and Representative sets the internal rules for their office for most issues: pay, org structure, schedule, writing style, and on down. So they could certainly declare, subject to availability, extensive use of AI is perfectly okay for their staff. But when that work crosses into other offices or the chamber at large, it’s unclear whether this could result in serious issues.
Also there are “shared services” resources that are essential to the legislative process on which every office relies. For example the Legislative Counsel office.
While any member of Congress is free to draft a bill on their own saying whatever they like, they typically request this central counsel office to draft on their behalf. Beyond substantive policy reasons, there are peculiar formatting and syntax requirements which if not adhered to can jeopardize consideration of the bill. But say an office decides to draft it’s own anyway, using a leading LLM. What would the public think? Rep. Lieu actually did this, but it was a non-binding resolution and anyway never seriously considered.
Every bill requires some form of signature from the sponsoring member, so technically they are assuming ownership of the language regardless of origin. But this is not a settled question. The Leg. Counsel office itself would certainly benefit from affirmative use of AI tools as they are always working at capacity and forced to triage requests. But again, are offices comfortable with that?
Actually the largest unexploited opportunity for improvement through automation and AI is in constituent services. Every office devotes significant labor to this responsibility, core to the legislative prerogative. Much of the work is merely intake and info management. I wouldn’t be so offended if an AI agent for instance was the first point of contact for my constituent concern. I suspect most people wouldn’t be so forgiving. Further down the services process, even if a human is the intake vehicle, using AI to record, annotate, store, and distribute info is an obvious value-add of advanced software. Again, how much of the public would be okay with that if they knew, even if it improved service?
In my time I would’ve benefited enormously from auto- or semi-automated reconciliation of various drafts (including proposed drafts used for negotiation). Much of this, which took up a great deal of our time, is properly understood as part of substantive legislating, but it’s primarily the “pre-work” before people get into a room and try to come to a consensus based on their elected boss’s preferences. “I’m still going through the redlines” is a routine holdup to productive negotiation from going forward. While I wish I could offload that task to a computer, I also ponder the degree I would (or should) trust it. Small excerpts of text or a subtle contextual misinterpretation can make for a big effect in policy. If I had mistakenly accepted a proposed change because of software error, and it passed into law….well, I wouldn’t be fired because I’d have already quit out of shame.
[I know the folks at Keeping Pace (POPVOX Foundation) and First Branch Forecast are more knowledgeable and have thoughts on this issue]
CRS is actually a department within the larger Library of Congress. The latter is responsible for operating the actual Congress.gov website which has its whole own list of deficiencies which any staffer could tell you about.
Some 19,000 every Congress. https://www.usnews.com/news/national-news/articles/2024-12-24/the-118th-congress-by-the-numbers
Getting into the weeds, House/Senate companions can have identical substantive meaning, but idiosyncrasies in each chamber’s drafting protocols result in differing texts.
p. 4. https://docs.house.gov/meetings/HA/HA00/20260625/119397/HMTG-119-HA00-Wstate-DonfriedK-20260625.pdf
Not wanting to run past this point, let me emphasize CRS is an incredible asset within the Legislative branch and highly undervalued. I myself failed to fully exploit this asset as a staffer until perhaps nearing the end of my time there, and I was certainly above the median user. Part of this is these high standards but also the intrinsic incentive structure and historic mission.
I attempted to determine average “time to summary” and share of bills currently without summary but was unsuccessful in pulling the necessary data (with assistance of Claude Code)
(1) requested access to Congress.gov's public API, which provides structured access to legislative data including bill introduction dates and CRS summary posting dates. (2) with API key, attempted to query the summaries endpoint directly. Those requests were blocked by the site's bot-detection systems. (3) attempted to work around this with locally run script (Python) using the same API key, which would have paginated through thousands of bills and computed the lag for each one. That approach was not completed.
The exception is the small footprint of SCIFs. It’s possible a rogue office could set up an air-gapped system in contravention of the rules, but even that probably limits the upside productivity ceiling because it won’t be interoperable with every other office.
Technically these offices are overseen by Congressional committees (Senate Committee on Rules and Administration and House Committee on Administration) who set the rules. In my experience these committees are not generally in the business of micro-managing the IT work of these offices and set fairly general policies from which to operate. But neither do they regularly take an active in pushing them, or finding ways to increase efficiency or re-evaluate risk tolerances.