chevron-down Created with Sketch Beta.
May 09, 2024 Feature

Using ChatGPT in Law School

Harsh Mahajan

Today, we stand on the precipice of a new era, where the burgeoning landscape of artificial intelligence (AI) promises to reshape the fabric of legal scholarship and practice. This transformation echoes a striking historical parallel: the emergence of the automobile. Just as early cars initially met with skepticism and even fear, while they revolutionized transportation, AI technologies are bound to drive a similar paradigm shift in the legal world.

My own journey as a law student has been deeply influenced by my background in software engineering. In today’s AI tools, I see not just a new utility but a new way of approaching long-established legal practices. This parallels one of my key takeaways from engineering: Always master tools designed to automate tasks rather than spending long hours doing them manually.

More than a technological leap, the rise of cars, or, as they were earlier called, “the horseless carriages,” necessitated a shift in perspective and adaptation. Early pioneers of cars didn’t just invent new machines; they had to convince the public of their utility and safety. Similarly, the proliferation of generative AI tools like ChatGPT calls for a reevaluation of how we approach legal writing. It isn’t simply about adding a new tool to our toolbox but embracing a new, AI-powered paradigm. This doesn’t diminish the importance of critical thinking and legal reasoning; it merely suggests that these essential skills can be complemented and augmented by the capabilities of AI.

Taking inspiration from my engineering background, this past fall semester, I chose an unconventional path for my two term papers: I wrote them with the assistance of ChatGPT, a generative AI tool. This wouldn’t have been possible without the support of my professors. One, an engineer himself, saw it as a valuable exercise in embracing technology impacting legal practice. He encouraged us to experiment in the safe haven of law school, where “failing” wouldn’t carry real-world consequences. My other professor, equally supportive, requested a disclaimer on the extent of AI assistance used. In the following sections, I’ll dive deeper into generative AI, large language models like ChatGPT, and my experience using ChatGPT for these term papers.

What Is ChatGPT?

The technical details of how ChatGPT works can get too convoluted too soon. So, in simple terms, ChatGPT is a tool that functions like a highly informed virtual assistant capable of generating text-based responses. It works by using a method called “transformer” technology, which is a bit like having a really good memory for patterns in language. ChatGPT has been “trained” by reading a vast amount of text, helping it learn how to piece together words and phrases in a way that makes sense. It is almost like it has read an enormous library of books and can recall and reuse that information to answer questions or write text.

That said, ChatGPT has its limitations. For one, it does not truly “understand” what it is talking about; it is more like it is really good at guessing based on patterns it has seen. Also, its knowledge is frozen at the point when it was last updated, currently 2021, so it will not know about recent events or the very latest trends. While it can write about a wide range of topics, it doesn’t have personal experiences or feelings, so its responses can sometimes lack a human touch. Additionally, because it learns from existing text, it can sometimes repeat biases or inaccuracies found in its training material. One of the largest pitfalls of ChatGPT is that it “hallucinates”: generating incorrect or misleading results due to insufficient training data, incorrect assumptions made by the model, or biases in the data used to train the model. We are all probably familiar with the issue of the New York lawyer who used ChatGPT to draft a legal brief for a case in federal district court, and ChatGPT ended up fabricating legal cases and citations. Such awkwardness can be avoided if every user understands the limitations of the tool they are using and utilizes the tool with those limitations in mind. To me, ChatGPT is meant to reduce human effort, not replace human intellect or creativity.

For law students, the tool’s ability to generate coherent, contextually relevant responses and learn from human feedback present an opportunity to enhance legal writing. Directing ChatGPT to generate documents by providing it with the necessary information can increase efficiency and help shave off significant periods of time that can be spent on other productive tasks. The ability to have a thesis statement drafted for you, on your direction, allows more time for editing the statement, rather than having to start from scratch.

ChatGPT, in my opinion, is about to bring about a significant shift in how legal writing can be conducted, promising increased efficiency and accuracy in these critical tasks.

ChatGPT: Another Tool in the Shed

Any new technology invariably reminds me of a lecture from my second year in software engineering. My professor sketched two rectangles on the blackboard, labeling one “Input/Output” and the other “Calculations,” connecting them with arrows to illustrate the basic functions of a computer. He went on to say, “Look at this idiot! It’s not smart, just fast!” His point was clear: Despite their speed, computers require human oversight to produce accurate outcomes.

MATLAB (an abbreviation of MATrix LABoratory) is a coding language by MathWorks that enables one to conduct extensive mathematical computations expeditiously. However, should errors arise, it is obvious that the onus is on the user to validate the correctness of the code, the algorithms applied, and the results obtained. In engineering, I was taught how to use MATLAB and how to conduct mathematical computations with it, but if the result was wrong, I was the one responsible for generating a wrong output. Likewise, I believe that with ChatGPT, the onus is on the student to ensure that the result is correct. The program, here ChatGPT, simply generates what we tell it to generate. If it does not generate what we want, we should revise what we are telling ChatGPT.

Similarly, computer-aided design (CAD) has revolutionized 3D modeling, allowing users to not only conceptualize entire gear trains but also generate precise engineering drawings instantaneously. This stands in stark contrast to the bespoke era when drafters toiled for hours, meticulously crafting these drawings by hand. Presently, CAD facilitates the swift creation and modification of such drawings, saving drafters valuable time. In our Engineering Design class in the first year of engineering, we were taught how to use Autodesk Inventor, a type of CAD. The focus was more on how to design, visualize, and integrate parts in CAD. We were taught how to read engineering drawings, build the parts from the basis of the drawings, and make free-hand sketches of the parts. Sketches are free hand, not to scale, and primarily intended to convey the idea, while drawings are precise, follow conventions, and are to scale. So, when it came to producing the drawings of our parts, we would simply click a few buttons in Autodesk Inventor to produce the required drawings with the specified conventions. One can design complex gear trains using CAD, but likeMATLAB, here too, the onus is on the user to design a proper, functioning gear train.

In both cases, the user is responsible and gets the credit for the work. Similarly, I see ChatGPT as a tool that can save law students hours spent crafting the perfect thesis statement. The key is to learn how to use this tool well, to create a thesis and then refine it as needed. It is like using code to get the desired outcome: You guide ChatGPT to produce the draft you want.

How I Used ChatGPT to Draft

As someone transitioning from engineering to law, the task of writing term papers comprising over 6,000 words seemed a Herculean task. In engineering, our reports, spanning 40–60 pages, mainly consisted of diagrams, tables, and explanations of our project’s components. Our research focused on existing solutions to a problem, followed by how our project is a more efficient approach to solving the problem. Unlike in law school, our engineering classes did not require us to write term papers based on extensive academic research, a concept unfamiliar to me. Only a small subset of engineers, who engaged in research, were familiar with concepts like building a thesis statement. To me, conducting research and building a thesis statement from it, and then an entire research paper, were a novel challenge. In staying true to my engineering practice of learning how to use a tool effectively, I turned to ChatGPT to aid in drafting my term papers. In one paper, I received a B+, and in the other, which was a bit shorter, I received an A–. For both the papers, I received the grade to which the class was curved. I could have certainly done better, but for my first attempt, I was quite content. My process involved the following steps.

Conducting the Research

Recognizing that ChatGPT’s knowledge base only extends to publicly available information until 2021, I relied on traditional legal research tools such as Westlaw and LexisNexis. This step ensured I was aware of the latest research and able to identify and rectify any “hallucinations” in ChatGPT’s responses. This way, because I had read the research and was aware of the arguments made in each paper, I could tell when ChatGPT had hallucinated and generated a result whose argument or logic did not align with those made in the research papers.

Priming ChatGPT

Priming ChatGPT is akin to dressing appropriately for a specific activity. One would not wear ski gear for fly-fishing, nor waders for skiing. Similarly, setting up ChatGPT in the correct context is crucial. Priming ensures ChatGPT understands the context and user expectations better, leading to improved results.

One may start by telling ChatGPT the ideal person for the job. For example, in the context of drafting term papers, I began by setting the context, telling ChatGPT, “You are a legal academic adept at writing legal research papers, and you will assist me in drafting a paper on [your topic].” Envision the ideal person for your job and describe that person to ChatGPT. The more detailed the context provided, the better ChatGPT can adapt to the role. For example, detailing the target audience, the paper’s tone, key takeaways, and desired reader impact sharpensChatGPT’s focus.

Next, we want to communicate to ChatGPT its role in aiding us. For example, “I will give you the outline of my term paper, and I want you to give me feedback on the outline. Tell me if the structure makes sense or if I am missing any potential discussion points. Then, we shall work on drafting the section together.”

Always ensure to ask ChatGPT to repeat the task before anything. This will ensure that we are on the same page as ChatGPT. If not, we can either edit our message or have a conversation with ChatGPT and direct it in our preferred direction.

Another way to build the context for ChatGPT is to ask it a series of questions (or queries). Here, we are having a conversation with ChatGPT, and through our queries we are helping ChatGPT reach the context in which we want it to operate, quite like the Socratic Method in law schools. The Prompt Engineering Institute offers an in-depth discussion on this priming process. It calls this a “pyramid approach,” which encourages small, specific questions to help ChatGPT understand the context and for you to gauge its knowledge.

Drafting with ChatGPT

With ChatGPT properly primed, we move into the drafting stage. Here, we provide snippets from various papers and instruct ChatGPT on how to weave these into the draft, mentioning the topic of the papers, argument structure, and desired word limit. For example, “I will give you these 2–3 articles, and I’d like you to draft a section based on them, adhering to a 500-word limit and including in-line references to the articles.” It is always a good practice to ask ChatGPT to provide in-line references to the articles so we can ensure that it is not hallucinating. Another means to ensure that ChatGPT does not hallucinate is to instruct it explicitly to only rely upon the information provided and to not seek any outside information. Here, because we asked ChatGPT to provide in-line references, we can cross-check if the argument drafted by ChatGPT matches the argument made in the article that it has referenced. If the argument drafted by ChatGPT aligns with, or uses the correct facts from, the article provided, we know that it has not hallucinated. Furthermore, as we have already done our research and read the articles ourselves, we are able to catch any argument that does not align with the provided articles.

Again, we want it to repeat our ask and ensure that we are on the same page. I often provide ChatGPT with the articles in one message. I inspect the output that ChatGPT generates. At times, regenerating a few more iterations generates a better output. Other times, a few cursory edits in the draft are sufficient. There are, of course, times when ChatGPT has totally missed the point. In such cases, you want to have a chat with it to point out what you want from it.

Wait until the very end for your introduction and your conclusion sections.

You need to keep in mind that you are in control. You are steering the car; you are in control of this horseless carriage; if it steers itself into a ditch, you are responsible. You need to ensure that you are directing it on how you want the content drafted. The more effectively you can convey your needs to ChatGPT, the more likely the result generated will be closer to what you intended. It does take some practice to achieve your desired results.

ChatGPT for Final Touches and Proofreading

Once we have performed the third stage for all the sections of our paper, we can start with the introduction and conclusion sections. Here, we will provide ChatGPT with the existing draft of our paper and instruct it to draft the introduction and conclusion sections.

Ensure ChatGPT is aware of your expectations for the sections. For example, “Ensure the introduction section starts with a thesis statement and a roadmap. The goal of this section is to introduce the topic and ensure that the reader knows what to expect in the following sections.” You can also add certain aims like “Aim to engage the reader’s interest.” Such prompts may produce some superfluous language, which can be either edited or specified to be avoided. With regards to a thesis statement, you probably have a statement in mind or know what to say but are finding it hard to put pen to paper. You can ask ChatGPT to generate a thesis statement, telling it how you want the statement structured, and then edit it to your liking. However, my preferred way is to write my own thesis statement and ask ChatGPT to revise it in the context of all the sections that have been drafted so far.

Similarly, for the conclusion section, we can ask ChatGPT to summarize all the arguments made in the paper and leave the reader with closing remarks. As with every section, a word limit can be specified.

Finally, now that we have all the sections of our paper drafted, it is time to turn it into a single, cohesive document. Tell ChatGPT that we are going to give it a draft of the entire paper, and it is supposed to proofread it. We can ask it to make the paper cohesive, check the grammar, and improve the flow of the arguments. Or we can ask it to critique the paper and give us feedback on the paper. I prefer the latter as, in that case, ChatGPT does not redraft the paper, which helps retain footnotes.

Wider Applications For AI In the Legal Profession

The ABA Model Rules state that lawyers must remain current with technological advancements and their implications. Artificial Intelligence is poised to transform our practice fundamentally. I personally believe that if we train ourselves to be adept at effectively using these new tools, we can become a lot more efficient in the service that we provide to our clients. We need to ensure that we are aware of the limitations of the tools that we use. For example, sharing confidential information with ChatGPT is not wise as it stores information and learns from it. Keeping the limitations in mind is a key step in using these tools effectively. In this context, ChatGPT might well be likened to the first horseless carriage of the legal profession, signaling a significant shift in our profession. I personally feel that we are yet to see the Model T equivalent of generative AI. Playing around with these tools now would prevent potential pitfalls in the future, such as sanctions for misuse in practice due to unfamiliarity.

Law schools should attempt to integrate AI tools like, LexisNexis AI, and Google’s Bard into their curriculum to prepare students for modern practice. While the LSAT ensures students can read and analyze logically, writing skills vary and are often influenced by one’s undergraduate education. Here, AI can aid in drafting legal documents, but the onus remains on law students to critically evaluate and apply their research, e.g., case law andlaw reviews, accurately within their work, ensuring their use of technology enhances rather than undermines the drafting of their legal arguments.

P.S. If you are wondering, then, yes, this article was drafted with the help ofChatGPT, but, naturally, a great deal of editing and revision were done outside ChatGPT. Law students must, of course, follow the rules of their law school and individual law professors surrounding the use of AI and ChatGPT.

    The material in all ABA publications is copyrighted and may be reprinted by permission only. Request reprint permission here.

    Harsh Mahajan

    Harsh Mahajan is a law student at Rutgers Law School in Newark, New Jersey, and a clinical research assistant at Rutgers Law School Intellectual Property Law Clinic. He holds a degree in Software Engineering from McMaster University, Canada, and previously worked in supply chain management.