Approach and development
There is no major dispute that such professions like drivers, journalists, accountants, tax experts, and stock traders will be replaced by artificial intelligence within the next 10 years. Lawyer is one of the professions where humans will be outperformed.
Our ambition is much stronger than that of existing competitors. We do not want to teach AI to specialize in a particular area of law, like bankruptcy, migration, parking tickets, or bank claims. Our ambition is to teach AI to think (i.e., to simulate intellectual structures of lawyers) in order to make human lawyers useless in complex litigation, in so-called “hard cases”.
Our strongest advantage is the judicial reasoning research of Prof. Dr. Stanislovas TOMAS, PhD (Sorbonne). This research distinguishes 98 intellectual patters divided into 411 structuring reasoning of European courts [https://www.academia.edu/41911802]. Lawyers think by intellectual structures, and those structures allow producing legal arguments regardless of meaning of the legal facts and principles employed. This research is rooted in the idea of American Legal Realism on automatic argument-bite production. Those 98 techniques of automatic argument-bite production structure legal reasoning in a manner understandable for or simulatable by a neural network, and this is the key to success.
Automatic legal argument-bite production allows producing legal reasoning without understanding the meaning of the very words within that reasoning. This is why certain American legal scholars even propose to replace the word “law” or “legal principle” by “boiling kettle”, “sausage of Toulouse”, “tû-tû”, “have good time”, or “love”. The name and meaning of the principle are not pertinent if it is efficiently placed within the intellectual structures making up the judicial reasoning.
Our vision is that it is possible to teach the artificial neural network to distinguish facts from principles, and sub-principles from more general principles or supra-principles. Let’s take “rape”. Many people think that “rape” is a fact, but it is not. It is a concept (principle) consisting of such facts as, for example, presence of sperm of a particular male DNA in the vagina as established by a laboratory, presence of the same male DNA under the nails of the woman, and a statement of the woman that she did not give a consent. Moreover, if there are only two facts “the man says he didn’t rape” and “the woman says he did”, the conclusion will be that he didn’t, and this is an application of another principle the AI will be able to discover independently.
The project of teaching artificial neural network to reason like a human lawyer, and instead of a human lawyer is divided into 2 main stages:
Our ambition is much stronger than that of existing competitors. We do not want to teach AI to specialize in a particular area of law, like bankruptcy, migration, parking tickets, or bank claims. Our ambition is to teach AI to think (i.e., to simulate intellectual structures of lawyers) in order to make human lawyers useless in complex litigation, in so-called “hard cases”.
Our strongest advantage is the judicial reasoning research of Prof. Dr. Stanislovas TOMAS, PhD (Sorbonne). This research distinguishes 98 intellectual patters divided into 411 structuring reasoning of European courts [https://www.academia.edu/41911802]. Lawyers think by intellectual structures, and those structures allow producing legal arguments regardless of meaning of the legal facts and principles employed. This research is rooted in the idea of American Legal Realism on automatic argument-bite production. Those 98 techniques of automatic argument-bite production structure legal reasoning in a manner understandable for or simulatable by a neural network, and this is the key to success.
Automatic legal argument-bite production allows producing legal reasoning without understanding the meaning of the very words within that reasoning. This is why certain American legal scholars even propose to replace the word “law” or “legal principle” by “boiling kettle”, “sausage of Toulouse”, “tû-tû”, “have good time”, or “love”. The name and meaning of the principle are not pertinent if it is efficiently placed within the intellectual structures making up the judicial reasoning.
Our vision is that it is possible to teach the artificial neural network to distinguish facts from principles, and sub-principles from more general principles or supra-principles. Let’s take “rape”. Many people think that “rape” is a fact, but it is not. It is a concept (principle) consisting of such facts as, for example, presence of sperm of a particular male DNA in the vagina as established by a laboratory, presence of the same male DNA under the nails of the woman, and a statement of the woman that she did not give a consent. Moreover, if there are only two facts “the man says he didn’t rape” and “the woman says he did”, the conclusion will be that he didn’t, and this is an application of another principle the AI will be able to discover independently.
The project of teaching artificial neural network to reason like a human lawyer, and instead of a human lawyer is divided into 2 main stages:
1
Creation of 50 000 of examples of division among facts, sub- and supra-principles led by a team of lawyers in a cooperation with a team of AI developers. This is a dataset to teach the AI. The creation of the dataset will take from 6 to 7 months.
Actually, this dataset itself will have a commercial value of about € 20 million, because, first, such dataset in an analyzed form does not exist. Second, the dataset will be created on the basis of international investment arbitration case law, which is the most expensive segment of law on this planet, and which is valid worldwide, since it is a part of international law. Even if the AI were not created, this dataset will be the most advanced analysis of investment case law at the World market, substantially outperforming all leading law firms together. This dataset will bring profit even in the absence of next stages. However, our objective is to implement the project until the end, because we want this AI to be able to deal with all other parts of law, in any country and in any language of this planet (i.e., including family law of the Kingdom of Tuvalu in the Tuvaluan language).
2
Teaching the AI to distinguish facts, sub- and supra-principles led by a team of AI developers, and controlled a team of lawyers. This will take about 12 months.
3
Creation of 50 000 of examples of 98 intellectual structures connecting facts, sub- and supra-principles, and conclusions therefrom led by a team of lawyers in a cooperation with a team of AI developers. This is a dataset to teach the AI. The creation of the dataset will take from 6 to 7 months.
4
Teaching the AI to reason, i.e., to make those 98 connections among facts, sub- and supra-principles in order to produce certain conclusions, led by a team of AI developers, and controlled a team of lawyers. This will take about 12 months.
The team of lawyers will be composed of 10 lawyers. The team of AI developers will be composed of 5 persons. The work will be supervised and controlled by Prof. Dr. Tomas and Mr. Oļegs Roščins (Oleg Roshchin). Our central computer with Graphics Processing Unit and Tensor Processing Unit will be located in the Silicon Valley.
In order to teach an artificial neural network to figure out facts and principles, then summing and biasing them in order to predict the pattern of covering those facts by those principles.
The Y output of this facts-principles distinction perceptron will become an input for the next perceptron producing legal construction.
A binary activation model used by the neural network at the facts-principles distinction perceptron, will be replaced by a 98-type activation model in the perceptron producing legal construction. By the way, currently we use the number 98, because it is based on already existing database of Professor Tomas, but this number might be simplified in certain respects and increased in other aspects depending on the empirical data we gather during the development.
From the moment when the AI is able to preview application of principles on the basis of facts, it means that this AI is able to distinguish facts from principles, as well as multiple levels of principles, and their interaction. This is the most challenging task. When this is completed, the next task of replacing human lawyers by the AI is solved.
The artificial neural network that is able to distinguish facts from principles and sub-principles from supra-principles, will be able to structure those pieces of information into those 98 intellectual patterns in order to produce new legal constructions. Since the AI has much better memory and operational connection speed than humans, the human race will leave competition. It will happen in the same manner as in chess and in Go.
At the initial stage, the weights for input of information will be defined and supervised by our human lawyers’ team from case law of international investment arbitration. However, within 1 year we expect back propagation to start adjusting weights automatically. This is possible, because the correct answers on functioning of patterns are always in the very text of case law. Error T< Error T-1 will lead us to success by the end of the 3rd year.
By a patient empirically adjusted combination of big and small learning rates we will achieve the desired local cost minima, i.e., the highest performance of the perceptrons.
Due to the complexity of the legal and factual dataset, we give priority to the stochastic gradient descent computing single samples, and converging faster. Online learning is preferable for the reason of weight and threshold adjustments’ performance after each sample. At an advanced stage, the AI will start a continued unsupervised deep learning, and establishing creative connections of legal facts and principles.
The fact that lawyers reason with identifiable intellectual structures, provides un a possibility to describe this activity so precisely that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.
We are perfectly aware about the problem of contradictory judgments, contradictory reasonings of the same judge, incoherence of case law, and law in general. However, this is not a problem for a lawyer representing a particular interest, since this private lawyer or our AI will simply apply all known intellectual structures and patterns. It is a problem for a judge who must be governed by the principle of social responsibility and by a policy of a certain coherence. In other words, our AI is a private and partial chess player, but not an authority dealing with the question what is the best and the most beautiful chess play acceptable for both contradicting parties.
Connecting the AI to the ordinary human language, to a chatbot, and extending it to other languages, segments of law, and jurisdiction will be much less challenging.
In order to teach an artificial neural network to figure out facts and principles, then summing and biasing them in order to predict the pattern of covering those facts by those principles.
The Y output of this facts-principles distinction perceptron will become an input for the next perceptron producing legal construction.
A binary activation model used by the neural network at the facts-principles distinction perceptron, will be replaced by a 98-type activation model in the perceptron producing legal construction. By the way, currently we use the number 98, because it is based on already existing database of Professor Tomas, but this number might be simplified in certain respects and increased in other aspects depending on the empirical data we gather during the development.
From the moment when the AI is able to preview application of principles on the basis of facts, it means that this AI is able to distinguish facts from principles, as well as multiple levels of principles, and their interaction. This is the most challenging task. When this is completed, the next task of replacing human lawyers by the AI is solved.
The artificial neural network that is able to distinguish facts from principles and sub-principles from supra-principles, will be able to structure those pieces of information into those 98 intellectual patterns in order to produce new legal constructions. Since the AI has much better memory and operational connection speed than humans, the human race will leave competition. It will happen in the same manner as in chess and in Go.
At the initial stage, the weights for input of information will be defined and supervised by our human lawyers’ team from case law of international investment arbitration. However, within 1 year we expect back propagation to start adjusting weights automatically. This is possible, because the correct answers on functioning of patterns are always in the very text of case law. Error T< Error T-1 will lead us to success by the end of the 3rd year.
By a patient empirically adjusted combination of big and small learning rates we will achieve the desired local cost minima, i.e., the highest performance of the perceptrons.
Due to the complexity of the legal and factual dataset, we give priority to the stochastic gradient descent computing single samples, and converging faster. Online learning is preferable for the reason of weight and threshold adjustments’ performance after each sample. At an advanced stage, the AI will start a continued unsupervised deep learning, and establishing creative connections of legal facts and principles.
The fact that lawyers reason with identifiable intellectual structures, provides un a possibility to describe this activity so precisely that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.
We are perfectly aware about the problem of contradictory judgments, contradictory reasonings of the same judge, incoherence of case law, and law in general. However, this is not a problem for a lawyer representing a particular interest, since this private lawyer or our AI will simply apply all known intellectual structures and patterns. It is a problem for a judge who must be governed by the principle of social responsibility and by a policy of a certain coherence. In other words, our AI is a private and partial chess player, but not an authority dealing with the question what is the best and the most beautiful chess play acceptable for both contradicting parties.
Connecting the AI to the ordinary human language, to a chatbot, and extending it to other languages, segments of law, and jurisdiction will be much less challenging.
Prof. Dr. Stanislovas TOMAS, PhD (Sorbonne) research distinguishes 98 intellectual patters divided into 411 structuring reasoning of European courts