1 of 3

SUPERCHARGED SUMMARIES: THE PROBLEM

Congress introduces a lot of legislation. In the 117th Congress, 17,817; in the 118th (year-to-date), 9,416 according to GovTrack.us.

How does anyone keep track of this (staffers, civil society, public)? Congress.gov

​

​

​

Bill summaries are critical, but who authors them? The Congressional Research Service (CRS): Congress’ private think tank.

The issue: CRS does not provide summaries in a timely way!

​

DYLAN IRLBECK (DYLANIRLBECK@GMAIL.COM) AND PRANAY MITTAL (PRANAYM12@GMAIL.COM)

2 of 3

SUPERCHARGED SUMMARIES: A SOLUTION

  • 2023 has been the year of large language models (LLMs)!
  • Various properties of LLMs make them uniquely good at summarizing text: language “understanding”; contextual awareness; scalability; customization by task; and processing speed.
  • Our solution: in the absence of a CRS-authored bill summary, provide a GPT-4-generated short summary.
  • Prompt example: “Summarize the following bill in the style of a Congressional Research Service analyst, using one paragraph: <insert bill text>”
  • Copilot, not autopilot: CRS analysts would still approve these summaries, and the machine-generated text would be labeled as such in the interface.

DYLAN IRLBECK (DYLANIRLBECK@GMAIL.COM) AND PRANAY MITTAL (PRANAYM12@GMAIL.COM)

3 of 3

SUPERCHARGED SUMMARIES: LOOKING FORWARD

  • Short-term:
    • Generate new types of summaries: full section-by-section, focus on classes of outcomes (taxation, regulatory, technology).
    • Fine-tune the foundational model to summarize legislative text in the style of CRS (using prior bills/summaries as training data).
    • Incorporating AI-powered features across Congress.gov. For example, linking to bills with related outcomes (going beyond the current feature of related keywords / high-level terms).

DYLAN IRLBECK (DYLANIRLBECK@GMAIL.COM) AND PRANAY MITTAL (PRANAYM12@GMAIL.COM)

  • Long-term:
    • Where is the leverage for CRS (and Congressional staff, for that matter)?
    • Congress’ think tank – it ought to be thinking. I argue that summarizing is a low-leverage task that can and should be done mostly by machines.
    • Congress as an institution should be thinking about how to maximize its creative time: LLMs stand to improve many processes – which of your workflows could you supercharge?