Artificial intelligence poses some distinct dangers to lawyers as evidenced by the cockup of the lawfirm of Morgan & Morgan in submitting a brief with bogus citations. It is hard to imagine a more serious dereliction of professional duty than in submitting a brief with citations to cases that do not exist, and which no lawyer could have ever have read and selected as appropriate to cite.[1] Apparently, a user of A.I. cannot simply direct the computer to stop making stuff up, any more than we can expect Felonious Trump to stop making stuff up.
Other dangers lurk in artificial intelligence. Back in 2020, Thomson Reuters, the owner of the Westlaw legal database sued an A.I. company, ROSS Intelligence for copyright infringement. As all lawyers know, Thomas Reuters publishes legal decisions in the Westlaw database. The content of federal decisions are in the public domain, but Westlaw adds headnotes that categorize parts of opinions by legal concepts, with a précis of the subject matter addressed in the opinion. In attempting to create an A.I.-powered rival to Westlaw, ROSS Intelligence trained its system with Westlaw content, including the headnote content written by Thomson Reuters employees.
ROSS apparently failed, however, to check whether it’s A.I. system lifted headnotes wholesale from Thomson Reuters. Thomson Reuters alleged that over 21,000 headnotes infringed its copyright. Some of the headnotes used by ROSS were not clearly infringing, but an expert witness for ROSS acknowledged that 2,830 headnotes were copied. The trial judge reviewed these 2,830 headnotes and determined that 2,243 of the headnotes were so clearly infringing that partial summary judgment was warranted in favor of Thomson Reuters.[2] The trial court found that ROSS’s assertion of a “fair use” defense was improvidently deployed and unavailing to avoid liability.[3] Many of the pleadings in the case can be found here.
In a previous post, I reported some of the frustrations I experienced in working with A.I. large-language models for obtaining an accurate summary of medical research.[4] Since then, I have had a chance to review additional output directed in the hopes of providing a comprehensive summary of various problems in the communication of scientific ideas, both in the scholarly journals, main-stream news media, and social media. Although I would not close the door on attempts to harnass A.I. for creating sophisticated, meaningful content, I remain deeply skeptical of the current state of the art. As the ROSS case shows, there are serious problems with copyright violations because of wholesale copying of texts that are part of the domain of the A.I. engine. From the point of view of the human author, there is also the very real danger of unethical plagiarism.
There are other dangers lurking in the use of A.I. for writing tasks. From what I have seen to date, some of the problematic output includes:
- glib generalizations without support,
- incorrect assertions without support,
- incorrect assertions with erroneous support,
- haphazard citations,
- citations to known unreliable sources, and
- recycling of old claims without any new insight or analysis.
And then there is also the distinct danger of “hallucinations” in the form of non-existent citations, a danger that put the Morgan & Morgan lawyers into shame and disrepute. Maybe we humans should simply learn how to read, research, and write ourselves!
自由西藏
[1] “Hallucinations in Law and in Government,” Tortini (Feb. 19, 2025).
[2] Thomson Reuters Enterprise Centre GmbH & West Publishing Corp. v. ROSS Intelligence Inc., 1:20-cv-00613-SB, ECF 770 (D. Del. Feb. 11, 2025).
[3] Jonathan Bailey, “Judge Rules Against Fair Use Defense for AI Company,” Plagiarism Today (Feb. 11, 2025); Ashley Brooks, “Thomson Reuters v. Ross Intelligence: Use Of Copyrighted Material To Train AI Systems Is Not Fair Use,” Roth Jackson (Feb. 19, 2025); Richard M. Assmus, Brian W. Nolan, and Megan P. Fitzgerald, “ROSS AI Decision Gives Early Indication of Strengths and Weaknesses of Fair Use Defense, Mayer Brown (Feb. 12, 2025).
[4] “Debating a Computer about Genetic Causes of Mesothelioma – Part Two,” Tortini (Feb. 18, 2025).
