Dated: 2025-06-07. Added: 2025-06-08. Alternate title: “Inside OpenAI’s Plan to Embed ChatGPT Into College Students’ Lives”.
But even when trained on specific course materials, A.I. can make mistakes. In a new study — “Can A.I. Hold Office Hours?” — law school professors uploaded a patent law casebook into A.I. models from OpenAI, Google and Anthropic. Then they asked dozens of patent law questions based on the casebook and found that all three A.I. chatbots made “significant” legal errors that could be “harmful for learning.”
“This is a good way to lead students astray,” said Jonathan S. Masur, a professor at the University of Chicago Law School and a co-author of the study. “So I think that everyone needs to take a little bit of a deep breath and slow down.”
OpenAI said the 250,000-word casebook used for the study was more than twice the length of text that its GPT-4o model can process at once. Anthropic said the study had limited usefulness because it did not compare the A.I. with human performance. Google said its model accuracy had improved since the study was conducted.