ChatGPT Took U of M Law School Exams. It Got a C+
ChatGPT has generated equal parts fear, skepticism, and hope since it was first launched in late November. The new chatbot is an advanced artificial intelligence program that can produce fluent – and sometimes uncanny – long-form responses to a range of questions.
The technology has the potential to drastically change the way we work and acquire information, but perhaps especially in academia. That’s part of the reason University of Minnesota Law School professor Jonathan Choi decided to take a closer look. This week, he and three other U of M law professors released a new white paper studying ChatGPT’s performance on four actual law school exams.
The paper, titled “ChatGPT Goes to Law School,” compared the chatbot’s performance on law exams to actual students’ scores. The exams included both essay questions and multiple-choice questions. Instructors who graded the exams weren’t told which responses were written by the chatbot and which were written by students. The upshot? ChatGPT received a C+, “achieving a low but passing grade in all four courses,” according to the paper.
“Overall, ChatGPT’s performance on law school exams, while currently uneven at best, suggests considerable promise and peril,” the paper’s four authors wrote. “We expect such language models to be important tools for practicing lawyers going forward; we also expect them to be very helpful to students using them (licitly or illicitly) on law school exams.”
Notably, ChatGPT performed less well on the multiple-choice section than it did on essays, compared to other students. Choi wasn’t surprised by that particular finding. “If you think about the tasks that ChatGPT is trained on, it makes sense that long-form text would be where it’s most comfortable,” he told TCB in a Wednesday interview. “Language models like these are trained, essentially, to imitate text that you would find in the real world – blog posts, essays. Then they are fine-tuned on top of that to be good at this kind of dialogue, which is natural language. They aren’t fine-tuned for multiple-choice.”
And overall, the bot’s performance was far from perfect. “When ChatGPT’s essay questions were incorrect, they were dramatically incorrect, often garnering the worst scores in the class,” the paper’s authors wrote. “Perhaps not surprisingly, this outcome was particularly likely when essay questions required students to assess or draw upon specific cases, theories, or doctrines that were covered in class.”
🚨🚨 new white paper! We used ChatGPT to write exams and graded them blind alongside real law students at Minnesota. ChatGPT got a C+ average and a low but passing grade in Con Law, Employee Benefits, Tax, Torts. 1/5 https://t.co/sN8uW4qtnR
— Jon Choi (@JonathanHChoi) January 24, 2023
Still, ChatGPT, in theory, has the ability to make its way through law school, if its performance remained consistent, the paper’s authors wrote. That may be a terrifying prospect for some law professors, but Choi takes a more pragmatic, proactive view. In his view, ChatGPT’s use in academic settings is inevitable. He pointed to an informal poll conducted by the student newspaper at Stanford University showing that about 17% of students surveyed had already used ChatGPT for “brainstorming and outlining.” Just about 5% reported “having submitted written material directly from ChatGPT with little to no edits, according to the poll,” The Stanford Daily reported earlier this week.
“The people who don’t see the potential of this technology I think are just wrong here, because students are already using it,” Choi said.
As Choi sees it, ChatGPT likely won’t spell the end of the legal profession or legal education. But it certainly has the capacity to “radically reshape” both, he said. And that extends beyond the law school; the technology will likely mean shifts within “any part of the university.”
And perhaps it’s not all bad. Some academics don’t necessarily find an issue with ChatGPT as a brainstorming aid. Vanessa Nyarko, a graduate instructor and PhD student in the U of M’s communications studies department, said she has specifically prohibited plagiarizing responses from ChatGPT for papers. “I put it in my syllabus,” she said. But she still plans to allow use of the technology for idea generation for papers. “I think it’s a very helpful tool for brainstorming,” she said. “It’s giving you things to pull from the internet.”
The U of M, meanwhile, has not yet specifically addressed ChatGPT in its code of conduct. But, in July, before the technology made its debut, the university updated its definition of “scholastic dishonesty” to include the “unauthorized use of online learning support and testing platforms.”
Since it was released this week, Choi’s white paper has certainly generated a lot of discussion and curiosity from law professors and instructors in other departments. “I only put this up on Monday night, and this is the most downloaded paper I’ve ever had,” he said. “Other professors of other law schools are asking whether we want to set up a conference to discuss these issues. I’ve gotten lots of emails.”
The paper does lay out a few positive use cases for working lawyers, too: “Based on our results, we also expect that language models will become helpful tools for practicing lawyers. A lawyer could have ChatGPT prepare the initial draft of a memo and then tweak that draft as needed; she could use ChatGPT to draft her way out of writer’s block; she could use ChatGPT to produce an initial batch of arguments and then winnow them down to the most effective; she could use ChatGPT to adapt past examples of legal documents to make her work more efficient.”
But the prospect of robo-lawyers taking over for human ones? There’s been at least one attempt in that direction, but on the whole of it, that’s probably far less likely. For now. “While ChatGPT and similar tools might help a lawyer work more efficiently, they cannot (yet) replace the need for a lawyer to locate, understand, and reason from relevant sources of law,” the authors wrote in the paper.
“In my view, there’s a lot more promise than peril,” Choi said, playing on the wording of the paper. “This really stands to change the way law is practiced, and by extension, the way that law is taught, if we’re mindful of the way we need to educate our students for the real world. I think it’s an opportunity for students, law professors, and lawyers. The peril is if we refuse to adapt.”