Wittgenstein once wrote that “the limits of my language mean the limits of my world” (1). In fact, the development of applications focused on natural language processing has always been a challenge, as algorithms capable of “understanding” human requests and demands with the sensitivity that only another human can internalize have fascinated researchers around the globe for decades. However, to date, no program sophisticated enough to achieve this goal has ever been developed.
Some refer to the term Artificial General Intelligence (AGI) to describe the hypothetical transformative phenomenon of the technological singularity—in which this goal will be achieved— especially driven by the power of more advanced techniques, such as deep learning, which fosters so-called “generative” systems, since they are based on recurrent neural networks (RNNs) or transformers (large-scale language models used in natural language processing), which makes this reality all the more intriguing, especially when analyzing its consequences—which are many —some exciting and challenging—while its risks pose major obstacles for legal science.
What cannot be denied is that the paradigm shift (2) foreshadowed since the beginning of this new era has an essential foundation: information. Everything changed with Alan Turing and the famous problem (Entscheidungsproblem) (3) that would inspire his subsequent research, through which he sought to investigate a machine’s potential to process information to the point of generating responses in the same way a human would (4). In short, it was expected that the unbeatable processing power of microprocessors would eventually allow machines to “emancipate” themselves and, in fact, emulate human behavior. What Turing did not anticipate was that the precise and hermetic concepts of mathematics would, in the test, prevent machines from deceiving, lying, and dissimulating—which made it easy to distinguish human responses from those of machines.
Nevertheless, contemporary artificial intelligence systems—notably the Generative Pre-Trained Transformer (GPT)—have already proven capable of passing the “Turing Test” (and many other similar tests, such as the Winograd Schema) (5) and even exams administered by humans, such as licensing exams for professions like medicine or law in the United States (6).
While computer science has ushered in a new era of computational power and hardware development, the accumulation of information has been responsible for “fueling” this new equipment and driving the advancement of information structures through software. In this context, the solution to the Entscheidungsproblem would necessarily involve addressing the main objections to the proposition that machines can “think.” More than just a test, it was necessary to make viable what Turing called the “imitation game” (7): a scenario in which a given machine would be capable of deceiving one-third of its interlocutors, leading them to believe that it was a human being (8). Would such a machine be “thinking”?
This single question sparked various theories regarding how to bridge the differences between humans and machines (the human-machine divide) (9) and the dilemmas surrounding the development and evolution of artificial intelligence (10). It can be said that, given the current state of the art, it is not yet feasible to conclude that an algorithm can become “intelligent” (11). In the words of Howard Gardner, “one might conclude that logical-mathematical ability is not as ‘pure’ or ‘autonomous’ a system as others reviewed here, and perhaps it should be counted not as a single intelligence, but as some kind of higher-order or more general intelligence” (12).
The example of ChatGPT—which is nothing more than a commercial tool offered by OpenAI for processing text commands (prompts) that trigger the large Transformer language model —is merely a simple demonstration of the results that can be achieved through the maturation of systems capable of “developing” rapidly, thanks to the evolution driven by machine learning based on colossal amounts of data.
In this scenario, expertise in data science becomes extremely relevant, as it is in this multidisciplinary field that the study of all aspects of data is concentrated—from its generation to its processing to convert it into a valuable source of knowledge (13) and to facilitate predictive analysis, which uses statistical (14) and machine learning techniques to “predict” outcomes (15). This can be useful in various contexts, such as the analysis of historical data, the mapping of past judicial decisions to identify patterns and estimate the probability of success in a case, or even the analysis of legal text. In general, with Data Science, it is possible to extract meaningful information from large volumes of legal texts, such as laws, regulations, contracts, legal opinions, and judicial decisions; and the analysis of this content can facilitate the identification of trends, the analysis and interpretation of specific legal terms, and even the automation of tasks.
In the field of law, both data analysts and data scientists play important roles in the analysis and interpretation of legal data, enabling a deeper understanding of complex issues, the evaluation of results, and support for informed decision-making. In other words, legal professionals operating within the complex information society that has reached its peak in the 21st century do not typically develop such competencies and skills—at least not in the traditional sense (16). Thus, in seeking to identify some skills (17) necessary for the future (18), we consider the need for transparency and, ideally, explainability so that we can assimilate parameters derived from computational heuristics and the causal inferences of language models.
These are relevant concepts for the law, as they help to understand and justify decisions in a more transparent and well-reasoned manner (19). Transparency, because it makes explicit the use of a specific dataset or a specific processing technique. And, in general, explainability refers to the ability to “understand” and explain how a decision-making model or algorithm reaches certain conclusions (20).
In law, explainability is essential, as legal systems are based on the principle of justification. Thus, when applying data science techniques, it is important that the results and decisions generated be comprehensible and able to be explained in a clear and accessible manner, both to legal professionals and to the citizens involved in the process.
It should be noted, in this regard, that transparency and explainability (Art. 3, VI) are two of the several principles defined in the text of the substitute bill drafted by the Commission of Jurists responsible for providing input on the drafting of a substitute bill on AI (CJSUBIA) (21), which held several meetings and public hearings during the drafting of the substitute bill—presented in December 2022—and which inspired the final text of Bill No. 2.338/2023 (22).
If algorithms are not capable of “thinking,” but are sufficiently advanced to make decisions “resulting from a combination of non-original programming inputs” (23), at least a broad interpretation of the concept of “responsibility” could guide the resolution of problems such as “decision-making opacity, the lack of explanation regarding the criteria used, and the inheritance of flawed, biased, and prejudiced inputs, which culminates in the production of unjustified discrimination” (24).
In short, there are no definitive answers to all the challenges posed by widespread access to transformers (and other large-scale language models, such as ChatGPT), and the topic will continue to raise many questions, but progress toward establishing the precautionary function of civil liability seems truly necessary to reconcile technological development—which is rampant and unstoppable—with the necessary protection of fundamental rights (and the maximum prevention of algorithmic biases), especially through the proliferation of a culture of accountability.
(1) WITTGENSTEIN, Ludwig. Tractatus Logico-Philosophicus. 2. ed. Londres: Routledge Classics, 2001, p. 68, tradução livre. No original: “The limits of my language mean the limits of my world”.
(2) Thomas Kuhn used the term “paradigm” in the following sense: “a paradigm is what members of a community share, and, conversely, a scientific community consists of people who share a paradigm.” In this study, the term “paradigm” takes on a broader connotation, referring not only to a scientific community or a specific historical period but also to the various shifts in Western thought. KUHN, Thomas. A estrutura das revoluções científicas. São Paulo: Perspectiva, 2000, p. 221.
(3) Friedrich Kittler, referring to the “Turing Test” and the power of information, highlights the following: “Only in Turing’s paper On Computable Numbers with an Application to the Entscheidungsproblem there existed a machine with unbounded resources in space and time, with infinite supply of raw paper and no constraints on computation speed. All physically feasible machines, in contrast, are limited by these parameters in their very code. The inability of Microsoft DOS to tell more than the first eight letters of a file name such as WordPerfect gives just a trivial or obsolete illustration of a problem that has provoked not only the ever-growing incompatibilities between the different generations of eight-bit, sixteen-bit and thirty-two-bit microprocessors, but also a near impossibility of digitizing the body of real numbers formerly known as nature”. KITTLER, Friedrich. There is no software. CTHEORY.net. 18 out. 1995. Disponível em: http://www.ctheory.net/articles.aspx?id=74. Acesso em: 23 jul. 2023.
(4) TURING, Alan M. Computing machinery and intelligence. Mind, n. 236, 433-460, out. 1950. Disponível em: https://dx.doi.org/10.1093/mind/LIX.236.433. Acesso em: 23 jul. 2023.
(5) FALEIROS JÚNIOR, José Luiz de Moura. Breves reflexões sobre os impactos jurídicos do algoritmo GPT-3. In: BARBOSA, Mafalda Miranda et al (coord.). Direito digital e inteligência artificial: diálogos entre Brasil e Europa. Indaiatuba: Foco, 2021, p. 521-524.
(6) VARANASI, Lakshmi. AI models like ChatGPT and GPT-4 are acing everything from the bar exam to AP Biology. Here's a list of difficult exams both AI versions have passed. Business Insider, 25 jun. 2023. Disponível em: https://www.businessinsider.com/list-here-are-the-exams-chatgpt-has-passed-so-far-2023-1 Acesso em: 23 jul. 2023.
(7) TURING, Alan M. Computing machinery and intelligence. Mind, Oxford, n. 236, p. 433-460, out. 1950, p. 25. Disponível em: https://dx.doi.org/10.1093/mind/LIX.236.433 Acesso em: 23 jul. 2023.
(8) Cf. HARNAD, Stevan. The Annotation Game: on Turing (1950), on computing, machinery and intelligence. In: EPSTEIN, Robert; PETERS, Grace (ed.). Parsing the Turing Test: philosophical and methodological issues in the quest for the thinking computer. Cham: Springer, 2008.
(9) WARWICK, Kevin. The disappearing human-machine divide. In: ROMPORTL, Jan; ZACKOVA, Eva; KELEMEN, Jozef (ed.). Beyond Artificial Intelligence: the disappearing human-machine divide. Cham: Springer, 2015, p. 9.
(10) FLASIŃSKI, Mariusz. Introduction to Artificial Intelligence. Cham: Springer, 2016, p. 3-13.
(11) KAPLAN, Jerry. Humans need not apply: a guide to wealth and work in the Age of Artificial Intelligence. New Haven: Yale University Press, 2015, p. 3-16.
(12) GARDNER, Howard. Frames of mind: the theory of multiple intelligences. Nova York: Basic Books, 2011, p. 168, tradução livre. No original: “(...) one could conclude that logical-mathematical ability is not as “pure” or “autonomous” a system as others reviewed here, and perhaps should count not as a single intelligence but as some kind of supra- or more general intelligence.”
(13) OHM, Paul; DOGAN, Stacey; BESTAVROS, Azer; SELLARS, Andy. Bridging the Computer Science-Law Divide. Boston: Boston University Press, 2022, p. 24.
(14) CLEVELAND, William S. Data Science: An Action Plan for Expanding the Technical Areas of the Field of Statistics. International Statistical Review, Oxford, v. 69, n. 1, p. 21-26, 2001, p. 22.
(15) DOMINGOS, Pedro. The master algorithm: how the quest for the ultimate learning machine will remake our world. Nova York: Basic Books, 2015, p. 6.
(16) SUSSKIND, Richard. Transforming the law: essays on technology, justice and the legal marketplace. Oxford: Oxford University Press, 2000, p. 170.
(17) LEHR, David; OHM, Paul. Playing with the Data: What Legal Scholars Should Learn about Machine Learning. U.C. Davis Law Review, Davis, v. 51, p. 653-717, 2017.
(18) SUSSKIND, Richard; SUSSKIND, Daniel. The future of professions: how technology will transform the work of human experts. Oxford: Oxford University Press, 2015, p. 29. They note: “(...) the most efficient future lies with machines and human beings working together. Human beings will always have value to add as collaborators with machines”
(19) KAMINSKI, Margot E. Understanding Transparency in Algorithmic Accountability. In: BARFIELD, Woodrow (ed.). The Cambridge Handbook of the Law of Algorithms. Cambridge: Cambridge University Press, 2021, p. 121.
(20) PARENTONI, Leonardo. What should we reasonably expect from artificial intelligence? Il Diritto degli Affari, Florença, v. XII, n. 2, p. maio/ago., 2022, p. 195-196. The author explains the conceptual distinction but elaborates on the closeness between the two terms: “Although the notions of transparency and explainability are technically different, this section addresses them altogether since they are intimately connected. Roughly speaking, they mean that a human user is able to understand why an AI system generated a certain output and explain it to an ordinary user of that system. Transparency is undoubtedly a fundamental value provided in numerous legal standards, for both the public and the private sectors, worldwide. It is of paramount importance and should be respected according to the provisions of each legal system. Therefore, the bigger the transparency, the better”.
(21) BRASIL. Senado Federal. Atividade Legislativa. Comissão de Juristas responsável por subsidiar elaboração de substitutivo sobre IA (CJSUBIA). Disponível em: https://legis.senado.leg.br/comissoes/comissao?codcol=2504 Acesso em: 23 jul. 2023.
(22) BRASIL. Senado Federal. Projeto de Lei nº 2.338, de 2023. Dispõe sobre o uso da Inteligência Artificial. Disponível em: https://www25.senado.leg.br/web/atividade/materias/-/materia/157233 Acesso em: 23 jul. 2023.
(23) FERREIRA, Ana Elisabete. Responsabilidade civil extracontratual por danos causados por robôs autônomos: breves reflexões. Revista Portuguesa do Dano Corporal, Coimbra: Imprensa da Universidade de Coimbra, n. 27, p. 39-63, dez, 2016, p. 44.
(24) MEDON, Filipe. Inteligência Artificial e responsabilidade civil: autonomia, riscos e solidariedade. Salvador: Juspodivm, 2020, p. 408.
Article originally written on July 23, 2023