In a voting-rights trial with thousands of pages of evidence, generative AI tools offered a glimpse of how technology might ease the judiciary’s heaviest burdens.
E-discovery tools that harness the power of artificial intelligence (AI) assist attorneys somewhat regularly.1 But my recent experience presiding over a bench trial in La Union Del Pueblo Entero v. Abbott showed me that generative AI (GenAI) can help judges in several crucial ways that go far beyond discovery.
La Union Del Pueblo Entero was a consolidated action brought by advocacy groups, voters, and an election official challenging two dozen provisions of an omnibus election law enacted in Texas in 2021, known as S.B. 1, under various federal civil rights statutes and the U.S. Constitution. In all, there were approximately 80 witnesses (both live and by deposition testimony), nearly 1,000 exhibits, and more than 5,000 pages of trial transcripts.
This case provided an opportunity to evaluate how GenAI might help a judge in a complex and document-intensive case. This article explores how GenAI can help locate documents and summarize witness testimony, whether GenAI tools are currently capable of completing a rough draft of a judicial opinion, and how GenAI can review party submissions.
Locating Documents: Saving Trees and Cutting Time
The first challenge facing my chambers was simply how to locate the exhibits and relevant portions of testimony as we began researching and drafting findings of fact and conclusions of law (FFCL).
While the federal court’s electronic case filing system (ECF) permits full-text searches of docket filings, its functionality is limited. First, it relies on keyword search, which tends to be both over- and under-inclusive. Second, many documents filed on ECF have not been run through optical character recognition (OCR) software, rendering them unsearchable. Third, ECF searches do not permit judges to limit their queries to specific documents on the docket or to search documents that have been submitted to the court but not yet filed. Thus, a search for a party name, for example, would likely turn up every document filed in the case — not a particularly helpful exercise.
I teach an e-discovery law school course, and over the past 10 years, many vendors have graciously provided my students with access to their review platforms. I give my students requests for production and require them to use the relevant platform to provide the number of “hits” they believe are responsive. Given the difficulty my chambers would have in locating relevant exhibits, I immediately thought of turning to an e-discovery provider to ingest, index, and grant me access to a review platform that I could use to locate documents (or portions of documents) needed during the drafting process.
Merlin,2 an e-discovery vendor, agreed to help me, and — after the court’s IT department vetted the company for cybersecurity concerns — we uploaded about a thousand documents representing all the trial testimony, admitted exhibits, the parties’ briefings and proposed FFCL, and my earlier orders in the case to a secure site created specifically for our project.
Merlin’s OCR and integrated keyword and algorithmic search capabilities (Search 2.0) allowed my chambers to quickly locate key documents and testimony without crafting complicated searches. Like many e-discovery tools I have used, Merlin permits users to tag documents by relevant fields.
Given the evolution of GenAI in recent years, I was interested in testing out other uses for these tools in my chambers, including tasks — such as summarizing evidence and drafting factual findings — and reviewing our work product.
One note, however: Judges and attorneys considering using these GenAI tools should be cautious in uploading any data that is confidential, privileged, or otherwise contains sensitive data (e.g., financial or health information). In particular, users should ensure that the provider’s large language model (LLM) does not use any nonpublic data to train its system and that the provider has adequate cybersecurity measures in place.3
Summarizing Text: Generating Overviews of Witness Testimony
GenAI has a unique ability to summarize text data. Although GenAI summarization can struggle with some types of information, judges and lawyers may find GenAI to be useful for analyzing trial and deposition testimony. Deposition and transcript summaries are a standard way to extract, condense, and organize information from testimony that may be spread over the course of hundreds — or thousands — of pages and multiple days of examination. While useful, they can be time-consuming (and expensive) to create and, depending on how they are structured, can produce inflexible results.
To conduct this task manually, the judge or lead counsel must determine how the transcript will be summarized (e.g., chronologically, by witness, or by topic). Then, a more junior attorney or legal assistant must locate the portions of the transcript they intend to summarize, typically by running keyword searches against transcripts. Finally, of course, they read and summarize the portions of the testimony they identified through their keyword searches, including pin cites to the pages and lines of the transcript along the way.
Read more
Judging AI: How U.S. Judges Can Harness Generative AI Without Compromising Justice
