Artificial Intelligence and Context
Guest post by Justin Tung, Reference Librarian at Tarlton Law Library, University of Texas School of Law
Recently, Adam Bent, a fellow Law Librarian and former classmate of mine, published an article in Pace Law Review entitled “Large Language Models: AI’s Legal Revolution.”[1] In it, he lays out a well-researched history of Chatbots and Large Language Models (LLM’s), giving historical perspective to the situation the legal industry is currently facing in terms of the “AI Legal Revolution.” The overarching conclusion is that treating all LLMs as suffering from the same undesirable characteristics, namely issues of confidentiality and hallucinations, is inaccurate, and that legal-oriented LLM such as CoCounsel or Lexis+ AI are designed specifically to not suffer from those issues. The argument is that since such priorities are integral to the design of these models, their use by legal academics, practitioners, and judiciaries can and should use them ethically.
Although I agree with Bent on many points, this perspective on LLMs should be contextualized. Any conversation about LLMs is undeniably going to mention its limitations and shortcomings. Because of this, in addition to excitement and novelty, it’s easy to lose sight of what these tools are offering. This piece aims to incorporate this greater context into views on LLMs and their place in legal practice and research.
Confidentiality
Firstly, Bent quotes Matt Reynalds quoting Jeff Pfeifer (chief product officer for LexisNexis in the United Kingdom and North America) saying that “[Lexis+ AI is a] private model interaction[], where there is no risk of a user interaction being subsumed into a wider model or where any users’ information can flow to another user of our service.”[2] While this sounds promising, it does not necessarily address all concerns about information sharing behavior and capacity.
In 2021, LexisNexis’ Risk Solutions business signed a 16.8 million-dollar contract with United States Immigration and Customs Enforcement. [3] Lexis would provide them with “billions of different records containing personal data aggregated from a wide array of public and private sources, including credit history, bankruptcy records, license plate images, and cellular subscriber information.”[4] On one level, this situation is somewhat comical since LexisNexis had this opportunity because the Federal Government’s contract with CLEAR (a similar risk industry service operated by Thompson Reuters) was expiring. This was the same Hatfields and McCoys all over again.
On another level, this news was shocking and highlights another concern with AI tools. Just because a corporation such as LexisNexis or Casetext (the owners of CoCounsel) can promise that user data will not be leaked to another user, this is not a guarantee of confidentiality in any other regard. Confidentiality, as a term of art in the legal field, is most closely associated with the idea that disclosures between client and her attorney cannot be shared with anyone under any circumstances except in some narrow, well-defined situations. For that reason, perhaps a better term to describe the concern raised by this issue is “privacy”.
Even if this information isn’t monetized by being sold in a deal like the one between LexisNexis and ICE, the information will almost certainly be monetized by using it to train the LLM itself. Part of how LLMs develop over time is through processing of user interactions. Bent does not explicitly note this functionality of legal LLMs, but implicitly notes this when he discusses Jabberwacky, a chatbot developed in 1988 which “learned how to converse with users by recording the inputs of all users across all conversations that Jabberwacky had previously.”[5]
Bent discusses this concern, noting that “For decades, client data via legal research has been shared with online legal research platforms. . . . That is not to say that client confidentiality is unimportant; indeed it is. It is only to say that the privacy concern over LLMs is not new; it is one that online legal research platforms have been overseeing for decades.”[6] I disagree with this reasoning. By analogy, vehicular assault was a concern when the top speed of the Model T ford was 42 mph. This does not mean that concerns about vehicular assault are invalid in discussing modern cars which can easily reach top speeds of 80 or 90 mph.
Similarly, discussions about collections and uses of information is also more pertinent today than ever. When an AI product is produced and commercialized by a corporation, one of the attributes of it which directly and overtly increases its desirability is the “intelligence” it has gathered through a history of interactions with users.[7] This is what made Jabberwacky better. This behavior of gathering, processing, and incorporating of data is a reason to buy it, not a reason against it. While the collection and capitalistic leveraging of this data may have existed before LLMs, this behavior has never been advertised as such an overt asset built into the very nature of the product.
To be fair, two points ought to be made. Firstly, Bent is certainly correct that different LLMs behave differently and can be configured differently. I wholeheartedly echo his point that the legal industry’s impulse to treat all LLM’s as alike in their development of perspectives and policies towards them without differentiation or nuance is detrimental. Determinations about the appropriateness of technologies requires, at minimum, a robust understanding of what those technologies are.
Secondly, concerns about privacy do not equate to concerns about confidentiality. The point raised about privacy is different by its nature, tending more towards moral issues of individual rights and freedoms. However, in the process of adapting LLMs for legal use, these concerns certainly cannot be hand-waved by saying that confidentiality concerns have been technologically resolved.
Hallucinations
The other blanket concern about LLMs discussed by Bent is the issue of hallucination. He defines these as “fabricating information “[8] such as “quotes or citations.”[9] From a technical perspective, this phenomenon poses an issue that many computer scientists and artificial intelligence engineers undoubtedly have a lot of insights and thoughts on. However, from the standpoint of legal use, this phenomenon is nothing new- it is the same problem as plain-old bad research.
Corporations building and marketing LLMs have made representations about these products, promising the comprehensiveness of a database search with the power of computers to prioritize the resources so your time and energy can be focused only towards the most useful and important ones. However, the simple fact is that an LLM does not do database search.
Broadly and simplistically, when running a search on a database platform through a search bar, the search terms are processed as a discriminating criterion. Every result which matches that criteria from the pool of total possible results is retrieved. This is a technical, deductive, and comprehensive process. Note that all search engines use some ranking algorithm on top of this, but this is the basic query mechanism.
Conversely, LLMs work by using a variety of technological processes to analyze huge volumes of data, generating “a response by predicting the most likely next word or sequence of words based on the training data”[10] when prompted with a query. Note the crucial difference, that the response the LLM gives is not a retrieval of the data it has gathered. Instead, it is a generated result based on what the LLM believes “would” be the answer to the question as the LLM processed it. This process is more analogous to an iPhone’s auto-complete suggestions than a deductive database search.
The need to provide research at an appropriate level of comprehensiveness and efficacy will always be integral to the legal research process, and there are very popular discourses which claim that LLMs can do that. both inside and outside of the legal field. From a user standpoint, this seems to be a tool which can comprehend and synthesize vast volumes of information and cohere it into a cogent result, with no more fuss than a text message. From an economic standpoint, corporations know that the best way to sell a product is to represent to a buyer that it meets a need they have. These two dynamics- the representations of those selling LLMs, and the expectations of those using or purchasing it, creates a strong hesitation to evaluate the result of the LLM for what it is, as one would evaluate the output of a bad coworker or poor student submission.
However, this is often the case— It doesn’t take long interacting with the legal LLMs on the market before prompts generate responses with outdated sources, incorrect jurisdictions, lacking important categories of content, misrepresentation of authorities, and misinterpretation of a premise, all in an overconfident tone which lacks substantive detail and is unresponsive to the question. None of these issues is new to the field of legal research. It is simply a bad work product.
While hallucinations would certainly exacerbate these issues, the elimination of hallucinations would in no way resolve them. All these issues likely have technical solutions, but bad LLM generated research ought not to be excused because it manages to mitigate an issue that its own AI nature created. The fact that an LLM does not hallucinate is not a sufficient reason for its adoption in legal contexts. It is merely necessary.
Do not allow expectations, excitement, or salespeople to move the goalposts. Adopt an LLM in a legal context is that it actually meets the needs and use cases. Until that issue is remedied, the LLM simply isn’t very usable, regardless of whether it hallucinates.
Conclusion
Fundamentally, I believe that Bent is right. LLMs have proliferated in many sectors of industry, and are likely here to stay, either openly or at sufferance, given the lack of reliable tools to detect them. His identification of a regulatory need is sound and well-made. However, some paths and attitudes towards adoption are responsible, and others are not.
Adoption of LLM technology ought to be driven by the actual positive impact they have on legal research, legal practice, and society, not the representation of the corporations selling them, the expectation of the buyers, or the excitement of the new and cutting-edge. Law, as an industry, can neither afford to fall behind society’s advances (technological or otherwise), nor can it afford to rush headfirst into integrating tools which are not appropriate or ready.
The telos of this technology proposed by some would be for it to entirely or substantially replace legal work. Automated discovery and automated research would feed into an algorithm which would determine the most judicious outcome and render that as a final and binding judgement. Whether or not this future is a good one is a large and complex issue. But while pondering this, I can confidently say that even if issues with confidentiality and hallucinations are resolved, we are still very far from that future.
[1] Adam Allen Bent, Large Language Models: AI’s Legal Revolution, 44 Pace L. Rev. 91 (2023).
[2] Id. at 130.
[3] Sam Biddle, ICE’s Surveillance Database Exposed by LexisNexis, The Intercept (Apr. 2, 2021), https://theintercept.com/2021/04/02/ice-database-surveillance-lexisnexis/.
[4] Id.
[5] Supra note 1 at 107
[6] Id. at 130-131
[7] LexisNexis’ Privacy policy says, in pertinent part, “. . . we use your personal information to: . . . enhance and improve the Service and our other products . . .” LexisNexis, https://www.lexisnexis.com/en-us/terms/privacy-policy.page (last visited Mar. 6, 2024).
[8] Supra note 1 at 126
[9]Id. at 125
[10] Id. at 117
