This talk investigates an intelligent assistance (IA) approach to utilizing Large Language Models (LLMs) in the legal domain by addressing the risks associated with unchecked artificial intelligence (AI) applications. We emphasize the importance of understanding the distinctions between AI and IA, with the latter involving human-in-the-loop decision-making processes, which can help mitigate risks and ensure responsible use of this rapidly developing technology. Using ChatGPT and GPT-4 as a prime example, we demonstrate its dual role as both an AI and IA application, showcasing its versatility in a variety of legal tasks. We look at recently reported explorations in particular in using very LLMs in addressing tasks such as multiple-choice question answering, legal reasoning, case outcome prediction, and summarization. We argue that to fully achieve "augmented intelligence," a reasoning and knowledge base component is required, allowing IA systems to efectively support human users in decision-making processes.
this demonstration, we seek to show how these models have been efectively deployed for a variety of tasks This extended abstract presents a comprehensive critique within the legal domain, including multiple-choice quesand analysis of the application of Large Language Models tion answering, legal reasoning, case outcome prediction, (LLMs), with a focus on GPT-4 and ChatGPT, within the and summarization. legal sector. The talk unpacks the concept of Intelligent Nevertheless, the crux of this presentation argues that Assistance (IA), diferentiating it from artificial intelli- the full realization of “augmented intelligence” necessigence (AI), and underscores its value within the context tates not just the computational prowess of LLMs, but of human-in-the-loop decision-making, particularly in also a reasoning and knowledge base component. We the legal domain. By doing so, we delve into the signif- posit that IA systems should be designed to augment icant benefits and potential pitfalls associated with the rather than replace human users in decision-making prounchecked use of these technological innovations in the cesses. This means that while LLMs can process and legal sector. generate human-like text based on vast amounts of data,
The presentation initially emphasizes the critical dis- they should also be built to collaborate with human users, tinctions between AI and IA. While both possess their enhance their decision-making capacities, and make their strengths and unique features, IA is proposed as a more work more eficient and efective. ethically responsible and practical solution in the legal Via this talk, we aim to stimulate further discourse on sector due to its requirement for human involvement in the responsible and beneficial integration of LLMs in the the decision-making process. The key argument lies in legal sector, reinforcing the need for more sophisticated the fact that IA, in contrast to pure AI, has a better po- IA systems that can efectively balance the benefits of tential to mitigate the risks associated with unsupervised advanced AI technology with the invaluable expertise of technological applications, enhancing overall responsible legal professionals. use.
The talk uses the OpenAI-developed language models, ChatGPT and GPT-4, to highlight the dual capabilities of 2. AI vs. IA these models as both AI and IA applications. Through
In: Proceedings of the Third International Workshop on Artificial Intel- sions how the system would interact or be autonomous ligence and Intelligent Assistance for Legal Professionals in the Digital from human activity: Workplace (LegalAIIA 2023), held in conjunction with ICAIL 2023, June 19, 2023, Braga, Portugal. $ frank.schilder@thomsonreuters.com (F. Schilder)
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org)
based on algorithms and data it has been trained
on. An example of this would be AlphaGo, an AI
developed by Google DeepMind, that defeated the
South Korean professional Go player Lee Sedol,
one of the best players at Go. It did this without
any human input during the game, just based on
its previous training data [
Mind’s AlphaFold predict protein structures with
remarkable accuracy, a task that has remained
unsolved for decades. They do this by analyzing vast
amounts of data without any human intervention
[
users in their decision-making processes but does
not make the final decision itself. For instance,
IBM’s Watson for Oncology helps doctors in
diagnosing cancer and suggesting treatment plans, These examples elucidate the key diference between AI
but the final decision is always made by the hu- and IA: AI operates with relative autonomy, whereas IA
man doctor [
predictive text features in email clients (such as
Google’s Smart Compose) assist users in writing
emails by suggesting phrases but do not compose
entire emails autonomously [
IA On the other hand, Intelligent Assistance systems
like Grammarly, a language-correction tool, aid
humans in spotting and correcting errors,
facilitating a learning process that is heavily reliant
on human cognition [
erable potential when used in AI applications, it is important to acknowledge and address associated risks to ensure responsible and ethical use. These risks can range from reliability issues to ethical and legal concerns.
LLMs, including an AI system like GPT-4, are trained on vast amounts of data, and while they can generate human-like text, they do not understand the content in the same way humans do. This can lead to potential errors or misinformation. For instance, if a legal AI system misinterprets a statute or case law, it could provide inaccurate advice or predictions [10].
Tesla’s Autopilot feature, which can control the Since LLMs are trained on real-world data, they may
car’s steering, acceleration, and braking within perpetuate existing biases present in the training data [11,
its lane without human input, albeit under super- 12]. If unchecked, these biases can influence the advice
vision for safety reasons [
facturing line, like those developed by Universal
Robots, works alongside humans, assisting them
in tasks that require heavy lifting or precision,
but always under the control and supervision of
human operators [
As LLMs become more complex, the reasoning behind their outputs can become opaque, leading to a "black box" problem [13]. This lack of transparency makes it challenging to ascertain accountability if the AI system leads to incorrect or harmful decisions, especially in highstakes legal settings.
sensitive data. Ensuring the security and privacy of this data is crucial to protect client confidentiality and comply with legal requirements such as GDPR [14]. The misuse or breach of this data represents a significant risk.
complacency and diminished critical thinking abilities among users. Legal professionals must continue to apply their expertise and judgment in conjunction with AI tools [15].
Addressing these risks requires a combination of technical solutions (like refining training techniques and improving transparency of AI decision-making processes), regulatory measures, and fostering user awareness about the strengths and limitations of LLM-based AI systems.
ing, comes with an inherent limitation: these models lack an understanding of the semantic content they process.
Despite their ability to generate human-like text, they do not grasp the implications or nuances of the content they generate or analyze in the same way a human user would [10].
In the context of the legal sector, where precision, understanding, and complex reasoning are paramount, this limitation is critical. Legal professionals need to reason about the law, apply it to specific cases, understand complex interdependencies, and navigate ambiguities. These tasks are not just about processing language, but about reasoning and understanding the underlying principles and consequences [16, 17].
To fully realize the potential of LLMs as IA in the legal sector, there is a need for a reasoning and knowledge base component that goes beyond mere language processing.
Such a component would enable the IA system to support human users efectively in complex decision-making processes.
Existing LLMs, like GPT-4, are based on transformer models that excel in solving standardized tests but lack explicit reasoning capabilities. To fill this gap, we propose integrating these LLMs with knowledge graphs or similar structures that provide a contextual understanding of the data [18]. This way, the LLM could not only process text but also reason about it in a manner more aligned with human cognition. [10] E. M. Bender, T. Gebru, A. McMillan-Major,
S. Shmitchell, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, in: Proceedings Frank Schilder is a Senior Research Director at Thomson of the 2021 ACM conference on fairness, account- Reuters with TR Labs, leading a team of researchers to ability, and transparency, 2021, pp. 610–623. explore new machine learning and artificial intelligence [11] S. Matthews, J. Hudzina, D. Sepehr, Gender and techniques in order to create smart products for legal racial stereotype detection in legal opinion word NLP problems. His research interests include summarizaembeddings, in: Proceedings of the AAAI Confer- tion, question answering and information extraction, and ence on Artificial Intelligence, volume 36, 2022, pp. natural language generation. Frank received the master’s 12026–12033. degree in computer science (Diplom-Informatik) from [12] T. Bolukbasi, K.-W. Chang, J. Y. Zou, V. Saligrama, the University of Hamburg and the Ph.D. degree in cogniA. T. Kalai, Man is to computer programmer as tive science from the University of Edinburgh, Scotland. woman is to homemaker? debiasing word embed- Before joining Thomson Reuters, he was an Assistant dings, Advances in neural information processing Professor at the Department for Informatics, University systems 29 (2016). of Hamburg, Germany. [13] D. Castelvecchi, Can we open the black box of AI?,
Nature 538 (2016) 20–23. URL: https://doi.org/10.
1038%2F538020a. doi:10.1038/538020a. [14] S. E. McGregor, H. Zylberberg, Understanding the General Data Protection Regulation: A Primer for Global Publishers (2018). URL: https://academiccommons.columbia.edu/doi/10.
7916/D8K08GVB. doi:10.7916/D8K08GVB. [15] R. E. Susskind, D. Susskind, The future of the professions: How technology will transform the work of human experts, Oxford University Press, USA, 2015. [16] K. D. Ashley, Artificial intelligence and legal analytics: new tools for law practice in the digital age,
Cambridge University Press, 2017. [17] D. Zhang, F. Schilder, J. G. Conrad, M. Makrehchi,
D. von Rickenbach, I. Moulinier, Making a computational attorney, in: Proceedings of the 2023 SIAM International Conference on Data Mining, SDM 2023, Minneapolis, MN, USA, April 27-29, 2023, SIAM, 2023. [18] M. F. A. R. D. T. (FAIR)†, A. Bakhtin, N. Brown,
E. Dinan, G. Farina, C. Flaherty, D. Fried, A. Gof, J. Gray, H. Hu, A. P. Jacob, M. Komeili, K. Konath, M. Kwon, A. Lerer, M. Lewis, A. H. Miller, S. Mitts, A. Renduchintala, S. Roller, D. Rowe, W. Shi, J. Spisak, A. Wei, D. Wu, H. Zhang, M. Zijlstra, Human-level play in the game of diplomacy by combining language models with strategic reasoning, Science 378 (2022) 1067–1074. URL: https://www.science.org/doi/abs/10.1126/science. ade9097. doi:10.1126/science.ade9097. arXiv:https://www.science.org/doi/pdf/ 10.1126/science.ade9097.