It has been seen in a former post on artificial intelligence (AI) and business models that AI may foster standardisation.1 The post that you are reading is of interest not only to lawyers but also to anyone who is interested in AI, including data scientists. The concept of artificial intelligence is misleading and I have never seen anyone asking oneself whether a thing was behaving in an intelligent manner although I have heard people complaining about the stupidity of an automated process such as a computer program because it was behaving accordingly to its source code instead of producing the result that the end user was expecting. Let us leave the intelligence of a machine aside to focus on artificiality (1). We will then reconsider data and ethical regulations (2).
1 Artificiality
Artificiality is a vague notion that is easy to understand without machines (1.1). Its understanding by lawyers differs from the one shared among data scientists or software developers (1.2).
1.1 A vague notion
An artificial thing is a thing that does not come exclusively from nature. Data and legal facts are artificial. Has anyone seen data walking across the street? When have you seen a legal fact taking the bus for the last time? It is difficult to agree about the degree of artificiality of a thing in order to decide whether it is rather natural or rather artificial. Let us follow Lewis Carroll who helps us to formulate this issue in a simple way:
"'Classification', or the formation of Classes, is a Mental Process, in which we imagine that we have put together, in a group, certain Things. Such a group is called a Class."2
When it comes to raw data or legal facts, classification is a sign of artificiality. Young children start to show things that they recognise as members of a given class such as dogs or trees before they speak fluently even if they have not seen this dog or tree before. A machine does not know what a tree or a dog is; it does not understand anything. It just treats data that represent what a human classifies into two classes: "dog" or "tree". The machine is able after training on a specific dataset to sort images according to these two classes. The machine seems intelligent because it has modified its software to sort images accordingly to the training results. The fact that a machine does not understand data that it is treating confuses the layman who has to show that he is not a robot. Turing tests that typically involve image recognition are more and more difficult to pass by humans because automated treatments have improved since the early years of network communications. Hence, images have to be more complex to defeat robots. Nevertheless, as noted by authors affiliated to the Pisa KDD Lab, changing the background of a picture may lead to misclassification by a machine even if this change did not seem significant to the data scientist who conceived the dataset.3 The data scientist treats data that have been or will be processed by a machine at some point. This means that data that cannot be processed by a machine are useless. Legal facts are different since they involve a social dimension that makes them artificial, but they do not have to be processed by a machine.
Let us see that data scientists and lawyers may look at information from different perspectives.
1.2 Two different perspectives
Classification is a basic process in AI and law that is not necessarily based on logic.
Reasoning in French private law is based on syllogism. Two premises lead to a logical conclusion by linking two unrelated classes. Lewis Carroll links red apples with ripe apples to draw the conclusion hereafter that can be used to sort apples logically.
"Some red Apples are unwholesome;
No ripe Apples are unwholesome.
Some red Apples are unripe."4
Oliver Wendell Holmes Jr. argues that common law is not based on logic.
"The official theory is that each new decision follows syllogistically from existing precedents. But just as the clavicle in the cat only tells of the existence of some earlier creature to which a collar-bone was useful, precedents survive in the law long after the use they once served is at an end and the reason for them has been forgotten. The result of following them must often be failure and confusion from the merely logical point of view."5
Baldus reminds us that Romans did not aim at perfection.6
Between you and me, it strange to read authors who have been educated mainly in a common law jurisdiction writing on something they dislike, considering that this thing is not necessary at common law. Logic is not useful to decide that one's house is one's castle in order to restrict its entry and its search.7 Efficient AI systems are not necessarily based on logic either.
It has been seen that law as AI systems relied on two main approaches of reality: one that paid attention to logic, the other that disregarded it. None of these approaches are obsolete. Lawyers and AI specialists have to find a practical way to represent reality in an artificial manner. Regulation seems however closer to data science than to law.
2 Regulation
The proximity between data and regulation clearly appears when one looks at evidence (2.1) and ethics (2.2).
2.1 Evidence or data
When one alleges that something exists or has happened to support a legal claim, one has to prove it. A lawyer has to prove a fact that matters.
"The reason why a lawyer does not mention that his client wore a white hat when he made a contract, while Mrs. Quickly would be sure to dwell upon it along with the parcel gilt goblet and the sea-coal fire, is that he forsees that the public force will act in the same way whatever his client had upon his head."8
Holmes explains that to determine what fact matters, a jurist does not only look at the law but also at the social context to guess what to pay attention to. The social context exists prior to the dispute and another party will represent the same situation in a different manner. Mrs. Quicky's hat is a hat. A lawyer may well wish to argue that it is a car and may well find a way to prove it. He is however likely to experience frustration after his brilliant demonstration since the conclusion of his reasoning will seem far-fetched and thus unconvincing. A good point has to seem as natural and simple as possible. One sometimes has to read and to think for some time to find a simpler way to represent an issue. Simple is rarely easy. People know what a hat is and what it is for. Shin helps to understand that the layman does not have a similar comprehension of AI.9 Legal facts do not only rely on legal definitions but also on the common understanding that a thing is a thing. Legal evidence is needed to prove that an object exists, not to show how an object is a source of data that can be processed in a given way. Showing that an object has been correctly processed does not require any legal knowledge; it is a matter of compliance and regulation. When one focuses on data, one is attentive to process not to nature. Environment makes an object what it is, be it a person, a thing or an obligation. When one enters into a contractual relationship with someone else, does one consider the other party, the document, or what has to be done according to the contract? Probably, each of the three objects, i.e., a person, a thing, and an obligation. Law mainly deals with nature even though environment is social and thus artificial while data mainly are a matter of process since machines can only process data. AI has to be regulated, and therefore, any activity that may be affected by it has to be considered from a data perspective for the regulation to be efficient. The issue is that if IA is a thing that is as shapeless as energy or electricity it may well affect anything. In a world where every tangible or intangible thing is a source of data, persons, things, and obligations will still exist but data regulation will prevail over law and custom that have shaped common understanding over time.
2.2 Ethics or constraint
Law is related to social consensus expressed in a legislative process that in many countries is designed to protect freedom through democratic elections. These considerations seem to have nothing to do with a private dispute. Tocqueville reminds us what many people seem to have forgotten.
"It is difficult to draw a man out of his own circle to interest him in the destiny of the State, because he does not clearly understand what influence the destiny of the State can have upon his own lot. But if it be proposed to make a road cross the end of his estate, he will see at a glance that there is a connexion between this small public affair and his greatest private affairs; and he will discover, without its being shown to him, the close tie which unites private to general interest. Thus, far more may be done by entrusting to the citizens the administration of minor affairs than by surrendering to them the control of important ones, toward interesting them in the public welfare, and convincing them that they constantly stand in need one of the other in order to provide for it."10
The legislative process favours a civilised debate that should help anyone to reflect upon the link between private interests and common good. Finding a concrete solution to a political issue does not only mean finding an electoral majority. Law is useful because it helps to structure thinking in order to represent a concrete situation by combining practical details while taking non measurable constants into account. The most important of these constants is freedom. One cannot measure it, and yet it has to be taken into account during parliamentary debates. Machines can only process data and therefore cannot produce a piece a regulation that preserves freedom. They can measure to what extend the regulatory constraint should be increased or decreased to achieve the desired outcome. Any ethical or aesthetic issue can be reduced to a measurable increase or decrease. Nevertheless, I am not sure that measurable ethics allow one to understand that something is good or evil because one has to exercise his free will to understand it. One cannot be absolutely sure that an action is good. Complying to an ethical regulation does not suppress doubt. Freedom allows to do wrong, but it gives the person who does not know whether something is good a chance to do better by overcoming this doubt. Leaving the possibility to do evil, preserves the opportunity to do better by exercising free will. The painful experience of freedom enables us to overcome the disastrous Fall from Eden.11 A simple rule (Do not eat this fruit!) has been breached; immediate action has been taken to sanction the misconduct. Nevertheless, the effect of this sanction does not follow logically from the sanction itself: there is no logical explanation to the potentially lethal proximity between Eve and the snake after the Fall. As seen thanks to Baldus, Romans were not looking for perfection when thinking about law. They were right: law is not about achieving a predictable outcome that would make anyone safe. It is about exercising freedom to take risk in order to try to find a better solution to a practical issue.
Let us now imagine that instead of this simple rule, we design a sophisticated ethical regulation that would be applied by an AI. We describe the fruit as a measurable item, conceive a dataset based on these measures, and declare that when something matches the description, it has to be avoided. I am afraid that some people may think that the AI-driven process is easy and "fair" in the sense that they have no will to achieve anything really suitable. It however is less simple than the ancient verdict and leaves no room for freedom nor forgiveness. Furthermore, as you know from an earlier post12, AI may help the legal practitioner who deals with standardised data. It does not seem to me that dealing with standardised data is what I do best. The main benefit of legal education is that it gives the ability to solve practical issues that have an effect on more than one person by relying on human nature that is unpredictable. If everything has to become measurable to become easy and risk-free, then studying law has to become useless. I can imagine that in a near future some practitioners would have to measure the degree of constraint that they have to put on someone to change this person's behaviour according to an ethical regulation. I do not think that ethical compliance has anything to do with law.
It has been seen that a regulation that relied on data to determine the appropriate level of constraint to achieve a result had nothing to do with law.
In brief, beware of anything that can be measured!
-
See Artificial intelligence and business models at 2. ↩
-
Carroll, Lewis. Symbolic Logic. London, New York, Macmillan, 1896, p. 3. Internet Archive, http://archive.org/details/symboliclogic00carr. ↩
-
Guidotti, Riccardo, et al. « A Survey of Methods for Explaining Black Box Models ». ACM Computing Surveys, vol. 51, nᵒ 5, août 2018, p. 93:1-93:42. September 2019, p. 4 https://doi.org/10.1145/3236009. ↩
-
Carroll, Lewis. The Game of Logic. London, New York : Macmillan and Co., 1887, p. 25. Internet Archive, http://archive.org/details/gameoflogic00carrrich. ↩
-
Holmes, Oliver W. The Common Law. Boston : Little, Brown, 1909, p. 35. Internet Archive, http://archive.org/details/commonlaw00holmgoog. ↩
-
Baldus, Christian. « Desarrollar un derecho de diferencias. La aportación del derecho romano a la orientación del jurista europeo ». Espacios de particulares, espacios de juristas, Madrid : Marcial Pons, 2017, nº 11. ↩
-
See the famous Semayne's case (January 1, 1604) 5 Co. Rep. 91. ↩
-
Holmes, Oliver W., « The Path of the Law », 10 Harv. L. Rev. 457 (1897), p. 458. Wikisource, https://en.wikisource.org/wiki/Harvard_Law_Review/Volume_10/The_Path_of_the_Law. ↩
-
See Shin, Donghee. « The Effects of Explainability and Causability on Perception, Trust, and Acceptance: Implications for Explainable AI ». International Journal of Human-Computer Studies, vol. 146, février 2021, p. 102551, §7.1. https://doi.org/10.1016/j.ijhcs.2020.102551. ↩
-
De Tocqueville, Alexis. Democracy in America, Part II, John C. Spencer, translated by Henry Reeve, 8th ed., Pratt, Woodford & Co., New York, 1848, p. 111. Wikisource, https://en.wikisource.org/wiki/Democracy_in_America_(Reeve). Part II has been published in French for the first time in 1840. The French original contains État that is the State with a capital S. The capital has been inserted by the author of this post. ↩
-
See Bible, Genesis, 2,17 and 3,15. ↩