Artificial Intelligence in Jurisprudence: Deconstructing the Technology and Analyzing its Prospects
The contemporary stage of technological development is characterized by the pervasive implementation of systems founded on the principles of Artificial Intelligence (AI). Within a dynamically evolving legal landscape, the legal community is actively debating the potential and risks associated with AI's application in professional practice. VERUM Law Firm, consistent with its mission to thoroughly analyze current legal challenges, participated in the FAR EASTERN INTERNATIONAL BANKING LAW FORUM, where the essence and applicability of AI in jurisprudence emerged as a central topic.
Following the Forum, our team conducted in-depth research aimed at deconstructing the concept of "artificial intelligence" and critically analyzing its actual capabilities and limitations within the legal domain. This work seeks to answer the question: "What is under the hood of AI?" and to evaluate the prospects for its integration into legal practice.
1. Artificial Intelligence as a Mathematical Model: Definitions and Essence
The term "artificial intelligence" possesses significant heuristic power and evokes associations with highly intelligent autonomous systems. However, upon detailed examination of the technical and mathematical aspects of its operation, it becomes evident that, at its core, AI is 99.9% a composite of "mathematical models" and "mathematical modeling" processes. Modern neural networks, widely known as "Generative Pre-trained Transformers" (GPT), represent the quintessence of this approach.
In this context, the terms "generative" and "pre-trained" indicate the system's capacity to create novel content based on assimilated data and its prior training on extensive data sets. The word "transformer" reflects the specific architecture of these models, which enables efficient processing of data sequences, such as texts.
Thus, referring to "neural network 'X'" in fact signifies interaction with a highly complex mathematical model – a Generative Pre-trained Transformer. This re-evaluation of terminology is of fundamental importance. While the phrase "I use artificial intelligence" sounds authoritative and progressive, "I use mathematical models/modeling" may be perceived differently by legal practitioners. Nevertheless, it is the latter expression that most accurately reflects the technological underpinnings of AI, emphasizing its instrumental, rather than subjective, nature. In the subsequent exposition, we will consciously employ the expressions "mathematical model/modeling" as synonyms for the term "artificial intelligence" to maintain analytical rigor.
2. Law and Mathematics: Historical Context and Modern Challenges
The history of various scientific disciplines demonstrates a close interrelationship with mathematics. The advent of mathematical methods in economics in the mid-20th century led to its transformation into a more precise and predictive science, enabling the modeling of complex economic processes. Similar processes have been observed in physics, chemistry, medicine, engineering, and even certain humanities such as sociology and psychology, where mathematical models and statistical methods have become an integral part of research.
Law, conversely, remains one of the few fields maximally distanced from mathematics. With the exception of judicial statistics or rare attempts to apply game theory, the arsenal of mathematical tools and methods in jurisprudence is utilized to an extremely limited extent. This phenomenon prompts reflection: is law a unique domain impervious to mathematization, or is it merely a matter of the absence of adequate approaches?
Attempts to mathematize law have been made repeatedly. The idea of expressing legal norms through formalized constructs such as "if – then – else," as well as employing mathematical logic, fuzzy logic, graph theory, set theory, or UML diagrams, appears highly attractive. However, to date, these attempts have not resulted in the creation of universal or widely applicable mathematical models of legal systems. This is attributable to the inherent complexity and ambiguity of legal norms, their dependence on context, interpretation, ethical, and social factors that are difficult or impossible to formalize in numerical terms.
3. Technical and Infrastructure Requirements for AI Implementation
The implementation of systems based on mathematical modeling necessitates significant resources and specific competencies:
Thus, the implementation of AI systems requires not only technological expertise but also significant financial investment and a willingness to manage complex technical and legal risks.
4. Application of AI in Natural Language Processing (NLP) and Text Generation
One of the most discussed areas for AI application in jurisprudence is Natural Language Processing (NLP), which resides at the intersection of mathematics, computer science, and linguistics. Neural networks are capable of analyzing texts, identifying statistical characteristics: the number of characters, words, sentences, and the frequency of individual words and phrases. For example, analysis of the U.S. Constitution using a mathematical model might reveal that the word "shall" appears 191 times and "state" 48 times, with an average word length of 4.7 characters. In the case of the judicial decision Salomon v. Salomon & Co Ltd [1897] AC 22, algorithms quickly identify that the top five words include "company," "share," "appellant," "business," "one," providing a general understanding of the case's subject matter in a matter of minutes. This is undoubtedly useful for rapid orientation within large volumes of documents.
However, such an approach operates exclusively with numerical characteristics of the text, without penetrating its semantic and legal meaning. For a mathematical model, it is irrelevant what "shall" or "state" signify; only their frequency and position in the sequence are important.
As for text generation, it is also executed through mathematical methods, where text is described numerically, and subsequent operations are performed on these numerical representations. The outcome can be text stylized to a specific source (e.g., Leo Tolstoy's novels or judicial acts) but devoid of logical coherence and legal meaning. Skeptics rightly refer to such generative transformers as "stochastic parrots" or "advanced T9 versions," as they "predict" the next word in a sequence based on statistical patterns, rather than on an understanding of content. This process bears a surprising resemblance to the prolonged training method of Edelweiss Zakharovich Mashkin's heuristic machine from the Strugatsky brothers' "Tale of the Troika," where the apparatus learned to type by analyzing Brehm's "Life of Animals" without any comprehension of its meaning.
5. Predictive and Recommendation Models in Jurisprudence
Mathematical models are actively employed to solve classification and regression problems. Classification addresses "yes/no" questions (e.g., will it rain tomorrow, is a tumor malignant). In law, this could pertain to a question of breach of delivery deadlines or exceeding the speed limit. Models build predictions based on statistical data, yielding results in the form of probabilities (e.g., "with 82.3% probability, the delivery deadline has been breached"). However, legal practice rarely operates with probabilities as a final verdict. Moreover, constructing such models requires a significant volume of numerical data, which presents a substantial barrier for most legal tasks where qualitative characteristics predominate. Converting words into numbers without regard for their meaning, as generative transformers do, appears extremely perilous in a legal context.
Regression allows for the prediction of quantitative indicators (e.g., how much precipitation will fall, what amount of damages will be recovered). Similar to classification, these models rely on past data and generate probabilistic distributions of possible outcomes (e.g., "with 10.3% probability, damages of 100 rubles will be recovered; with 34.5% probability, 200 rubles"). The utility of such predictions for a legal professional requiring an unequivocal legal opinion raises serious doubts, and the challenge of collecting adequate numerical data remains unresolved.
Recommendation models, analogous to those used in online cinemas or contextual advertising, can be applied to find "similar cases" or "judicial precedents." They function by creating a preference matrix and calculating "distances" between elements (e.g., cases) based on mathematical similarity. However, as demonstrated by the example of films The Matrix and Fight Club, which may be deemed "similar" by mathematical algorithms despite significant semantic differences, such systems can produce results that are mathematically rigorous but legally incorrect. Mathematical similarity does not always correlate with legal relevance or doctrinal congruence.
6. AI in Highly Specialized Applications: Computer Vision and Speech Recognition
Alongside the aforementioned directions, there are areas of AI application demonstrating high effectiveness. Computer vision and speech recognition are already operational technologies. Examples include the automated recording of traffic rule violations (fines for speeding or crossing a solid line) and real-time speech-to-subtitle translation. The success of these systems is attributable to the fact that in these domains, the relevant norms (e.g., traffic regulations, administrative offenses codes) have been unequivocally and formally translated into a mathematical form amenable to algorithmic processing. This is a key distinction from most legal tasks requiring complex interpretation.
7. Prospects for the Displacement of Lawyers by Artificial Intelligence
Based on the conducted analysis, it can be concluded that "under the hood" of AI, one primarily discovers improved versions of familiar tools – internet search engines and legal reference systems. Mathematical models, currently termed "artificial intelligence," are capable of enhancing the speed and accuracy of information retrieval, enabling the formulation of more complex queries (e.g., "retrieve all judicial practice under Article 10 of the RF Civil Code and export it to a PDF file"). This is undoubtedly convenient and increases a lawyer's efficiency but does not replace them.
The operational principle of a generative transformer, which generates text without semantic comprehension, necessitates approaching this tool with particular caution. Legal activity demands not only textual manipulation but also a deep understanding of its meaning, context, doctrinal foundations, as well as the capacity for critical thinking, ethical evaluation, and the formulation of strategic decisions.
To the question "will AI replace lawyers?", the following answer can be offered: yes, it will, but not before the moment when legal norms can be described not in the language of mathematical statistics, but in some other, yet fully formalized, form. In this hypothetical scenario, a lawyer, confronted with a legal task, instead of studying texts of laws, commentaries, and judicial practice, would launch Excel or Python to construct a mathematical model for the solution. However, such a prospect raises doubts.
The analogy with autonomous vehicles, which, while navigating roads, comply with traffic regulations, demonstrates that technically complex tasks (recognizing signs, distances, pedestrians) can be resolved. But precisely how the vehicle "complies" with rules—that is, how it comprehends their essence rather than merely reacting algorithmically—remains unclear. Law, unlike traffic regulations, is imbued with concepts of justice, reasonableness, and good faith, which are not amenable to simple algorithmization.
Conclusion
The research indicates that artificial intelligence, in its current manifestation, represents a powerful toolkit founded on mathematical modeling and statistical methods. Its application can significantly enhance the efficiency of routine legal operations, such as searching and analyzing large volumes of textual data. However, the fundamental limitations of AI, related to its lack of semantic understanding, dependence on formalized data, and inability for comprehensive legal reasoning, interpretation, and ethical evaluation, render it incapable of fully replacing a legal professional. The future interaction between law and AI will likely entail a symbiosis: AI as a potent assistant, liberating the lawyer from routine tasks, and the lawyer as an indispensable subject, ensuring the depth of legal analysis, strategic counseling, and decision-making based on a unique combination of knowledge, experience, intuition, and ethical principles.
The contemporary stage of technological development is characterized by the pervasive implementation of systems founded on the principles of Artificial Intelligence (AI). Within a dynamically evolving legal landscape, the legal community is actively debating the potential and risks associated with AI's application in professional practice. VERUM Law Firm, consistent with its mission to thoroughly analyze current legal challenges, participated in the FAR EASTERN INTERNATIONAL BANKING LAW FORUM, where the essence and applicability of AI in jurisprudence emerged as a central topic.
Following the Forum, our team conducted in-depth research aimed at deconstructing the concept of "artificial intelligence" and critically analyzing its actual capabilities and limitations within the legal domain. This work seeks to answer the question: "What is under the hood of AI?" and to evaluate the prospects for its integration into legal practice.
1. Artificial Intelligence as a Mathematical Model: Definitions and Essence
The term "artificial intelligence" possesses significant heuristic power and evokes associations with highly intelligent autonomous systems. However, upon detailed examination of the technical and mathematical aspects of its operation, it becomes evident that, at its core, AI is 99.9% a composite of "mathematical models" and "mathematical modeling" processes. Modern neural networks, widely known as "Generative Pre-trained Transformers" (GPT), represent the quintessence of this approach.
In this context, the terms "generative" and "pre-trained" indicate the system's capacity to create novel content based on assimilated data and its prior training on extensive data sets. The word "transformer" reflects the specific architecture of these models, which enables efficient processing of data sequences, such as texts.
Thus, referring to "neural network 'X'" in fact signifies interaction with a highly complex mathematical model – a Generative Pre-trained Transformer. This re-evaluation of terminology is of fundamental importance. While the phrase "I use artificial intelligence" sounds authoritative and progressive, "I use mathematical models/modeling" may be perceived differently by legal practitioners. Nevertheless, it is the latter expression that most accurately reflects the technological underpinnings of AI, emphasizing its instrumental, rather than subjective, nature. In the subsequent exposition, we will consciously employ the expressions "mathematical model/modeling" as synonyms for the term "artificial intelligence" to maintain analytical rigor.
2. Law and Mathematics: Historical Context and Modern Challenges
The history of various scientific disciplines demonstrates a close interrelationship with mathematics. The advent of mathematical methods in economics in the mid-20th century led to its transformation into a more precise and predictive science, enabling the modeling of complex economic processes. Similar processes have been observed in physics, chemistry, medicine, engineering, and even certain humanities such as sociology and psychology, where mathematical models and statistical methods have become an integral part of research.
Law, conversely, remains one of the few fields maximally distanced from mathematics. With the exception of judicial statistics or rare attempts to apply game theory, the arsenal of mathematical tools and methods in jurisprudence is utilized to an extremely limited extent. This phenomenon prompts reflection: is law a unique domain impervious to mathematization, or is it merely a matter of the absence of adequate approaches?
Attempts to mathematize law have been made repeatedly. The idea of expressing legal norms through formalized constructs such as "if – then – else," as well as employing mathematical logic, fuzzy logic, graph theory, set theory, or UML diagrams, appears highly attractive. However, to date, these attempts have not resulted in the creation of universal or widely applicable mathematical models of legal systems. This is attributable to the inherent complexity and ambiguity of legal norms, their dependence on context, interpretation, ethical, and social factors that are difficult or impossible to formalize in numerical terms.
3. Technical and Infrastructure Requirements for AI Implementation
The implementation of systems based on mathematical modeling necessitates significant resources and specific competencies:
- Human Resources:Specialists capable of developing, implementing, and maintaining AI systems must possess profound knowledge in advanced mathematics (linear algebra, differential and integral calculus, mathematical statistics, analytical geometry, higher algebra, set theory, probability theory). Additionally, proficiency in various development tools and programming languages (Python with its extensive libraries, C++, Fortran, as well as tools such as Matlab, Wolfram Alpha, Maple) is required. Knowledge of databases and SQL is also mandatory. The question of the availability of a sufficient number of legal professionals meeting these requirements remains rhetorical.
- Data:Mathematical modeling relies on processing exceedingly large volumes of data and executing an exceedingly large number of calculations. For small law firms, the volume of proprietary data may be insufficient to yield statistically significant and adequate results. Larger entities, possessing extensive data sets, confront the challenge of processing such information.
- Infrastructure:Data processing demands substantial computational power. Standard office computers and laptops are typically inadequate. Two primary approaches are possible:
- Local Computing Resources: The acquisition and maintenance of powerful computing systems, comparable in performance to cryptocurrency mining equipment, which entails significant capital expenditures for the equipment itself, its support, maintenance, and updates, as well as substantial operational costs for electricity.
- Cloud Services: The leasing of computational capacities in the "cloud." This approach involves the necessity of transmitting confidential, personal, and other sensitive data to third-party servers, thereby generating risks of data leakage and requiring a meticulous assessment of legal and reputational implications. Furthermore, the utilization of cloud resources incurs charges, the magnitude of which can be substantial.
Thus, the implementation of AI systems requires not only technological expertise but also significant financial investment and a willingness to manage complex technical and legal risks.
4. Application of AI in Natural Language Processing (NLP) and Text Generation
One of the most discussed areas for AI application in jurisprudence is Natural Language Processing (NLP), which resides at the intersection of mathematics, computer science, and linguistics. Neural networks are capable of analyzing texts, identifying statistical characteristics: the number of characters, words, sentences, and the frequency of individual words and phrases. For example, analysis of the U.S. Constitution using a mathematical model might reveal that the word "shall" appears 191 times and "state" 48 times, with an average word length of 4.7 characters. In the case of the judicial decision Salomon v. Salomon & Co Ltd [1897] AC 22, algorithms quickly identify that the top five words include "company," "share," "appellant," "business," "one," providing a general understanding of the case's subject matter in a matter of minutes. This is undoubtedly useful for rapid orientation within large volumes of documents.
However, such an approach operates exclusively with numerical characteristics of the text, without penetrating its semantic and legal meaning. For a mathematical model, it is irrelevant what "shall" or "state" signify; only their frequency and position in the sequence are important.
As for text generation, it is also executed through mathematical methods, where text is described numerically, and subsequent operations are performed on these numerical representations. The outcome can be text stylized to a specific source (e.g., Leo Tolstoy's novels or judicial acts) but devoid of logical coherence and legal meaning. Skeptics rightly refer to such generative transformers as "stochastic parrots" or "advanced T9 versions," as they "predict" the next word in a sequence based on statistical patterns, rather than on an understanding of content. This process bears a surprising resemblance to the prolonged training method of Edelweiss Zakharovich Mashkin's heuristic machine from the Strugatsky brothers' "Tale of the Troika," where the apparatus learned to type by analyzing Brehm's "Life of Animals" without any comprehension of its meaning.
5. Predictive and Recommendation Models in Jurisprudence
Mathematical models are actively employed to solve classification and regression problems. Classification addresses "yes/no" questions (e.g., will it rain tomorrow, is a tumor malignant). In law, this could pertain to a question of breach of delivery deadlines or exceeding the speed limit. Models build predictions based on statistical data, yielding results in the form of probabilities (e.g., "with 82.3% probability, the delivery deadline has been breached"). However, legal practice rarely operates with probabilities as a final verdict. Moreover, constructing such models requires a significant volume of numerical data, which presents a substantial barrier for most legal tasks where qualitative characteristics predominate. Converting words into numbers without regard for their meaning, as generative transformers do, appears extremely perilous in a legal context.
Regression allows for the prediction of quantitative indicators (e.g., how much precipitation will fall, what amount of damages will be recovered). Similar to classification, these models rely on past data and generate probabilistic distributions of possible outcomes (e.g., "with 10.3% probability, damages of 100 rubles will be recovered; with 34.5% probability, 200 rubles"). The utility of such predictions for a legal professional requiring an unequivocal legal opinion raises serious doubts, and the challenge of collecting adequate numerical data remains unresolved.
Recommendation models, analogous to those used in online cinemas or contextual advertising, can be applied to find "similar cases" or "judicial precedents." They function by creating a preference matrix and calculating "distances" between elements (e.g., cases) based on mathematical similarity. However, as demonstrated by the example of films The Matrix and Fight Club, which may be deemed "similar" by mathematical algorithms despite significant semantic differences, such systems can produce results that are mathematically rigorous but legally incorrect. Mathematical similarity does not always correlate with legal relevance or doctrinal congruence.
6. AI in Highly Specialized Applications: Computer Vision and Speech Recognition
Alongside the aforementioned directions, there are areas of AI application demonstrating high effectiveness. Computer vision and speech recognition are already operational technologies. Examples include the automated recording of traffic rule violations (fines for speeding or crossing a solid line) and real-time speech-to-subtitle translation. The success of these systems is attributable to the fact that in these domains, the relevant norms (e.g., traffic regulations, administrative offenses codes) have been unequivocally and formally translated into a mathematical form amenable to algorithmic processing. This is a key distinction from most legal tasks requiring complex interpretation.
7. Prospects for the Displacement of Lawyers by Artificial Intelligence
Based on the conducted analysis, it can be concluded that "under the hood" of AI, one primarily discovers improved versions of familiar tools – internet search engines and legal reference systems. Mathematical models, currently termed "artificial intelligence," are capable of enhancing the speed and accuracy of information retrieval, enabling the formulation of more complex queries (e.g., "retrieve all judicial practice under Article 10 of the RF Civil Code and export it to a PDF file"). This is undoubtedly convenient and increases a lawyer's efficiency but does not replace them.
The operational principle of a generative transformer, which generates text without semantic comprehension, necessitates approaching this tool with particular caution. Legal activity demands not only textual manipulation but also a deep understanding of its meaning, context, doctrinal foundations, as well as the capacity for critical thinking, ethical evaluation, and the formulation of strategic decisions.
To the question "will AI replace lawyers?", the following answer can be offered: yes, it will, but not before the moment when legal norms can be described not in the language of mathematical statistics, but in some other, yet fully formalized, form. In this hypothetical scenario, a lawyer, confronted with a legal task, instead of studying texts of laws, commentaries, and judicial practice, would launch Excel or Python to construct a mathematical model for the solution. However, such a prospect raises doubts.
The analogy with autonomous vehicles, which, while navigating roads, comply with traffic regulations, demonstrates that technically complex tasks (recognizing signs, distances, pedestrians) can be resolved. But precisely how the vehicle "complies" with rules—that is, how it comprehends their essence rather than merely reacting algorithmically—remains unclear. Law, unlike traffic regulations, is imbued with concepts of justice, reasonableness, and good faith, which are not amenable to simple algorithmization.
Conclusion
The research indicates that artificial intelligence, in its current manifestation, represents a powerful toolkit founded on mathematical modeling and statistical methods. Its application can significantly enhance the efficiency of routine legal operations, such as searching and analyzing large volumes of textual data. However, the fundamental limitations of AI, related to its lack of semantic understanding, dependence on formalized data, and inability for comprehensive legal reasoning, interpretation, and ethical evaluation, render it incapable of fully replacing a legal professional. The future interaction between law and AI will likely entail a symbiosis: AI as a potent assistant, liberating the lawyer from routine tasks, and the lawyer as an indispensable subject, ensuring the depth of legal analysis, strategic counseling, and decision-making based on a unique combination of knowledge, experience, intuition, and ethical principles.