Ensuring Autonomy of Decision-Making by Artificial Intelligence for the Purposes of Legal Public Relations.

Atabekov Atabek Rustamovich PhD in Economics Associate Professor of the Department of Administrative and Financial Law, Legal Institute, Peoples' Friendship University of Russia 117198, Russia, Moscow, Miklukho-Maklaya str., 6 atabekoff1@mail.ru Other publications by this author

Regulation of legal relations arising from the phenomenon of artificial intelligence and transparency of decision-making for the purposes of public legal relations is strategically important both for Russia and for foreign countries. At the same time, the issue of AI positioning in the sphere of public legal relations occupies a separate role, the issue remains systemic and complex both at the level of theoretical research and practical solutions in the field of clarity and transparency of decisions made by AI.

At the level of the German Ministry of Justice ^[1], the EU Commission ^[2], individual US states ^[3], the issue of the need to resolve the "black box" is already being systematically considered.

Among representatives of foreign and domestic scientific schools, questions also systematically arise both in terms of the technical feasibility of opening this "box" ^[4], as well as the fact that the illusion of AI accountability for humans raises the question of terminological certainty of the concepts of "explainability", "comprehensibility", "meaningfulness", etc.^[5-8].

When considering the technical and legal aspect of the functioning of AI for the purposes of public authority, the first aspect is to ensure the transparency of AI, since the dynamic structure of the processed data is used, which is expressed by the absence of a direct relationship between input data and "output" decisions (outcomes) ^[9]. The reason lies in the fact that AI (especially those based on machine learning) use a large number of ways to analyze and achieve goals with different algorithmic principles and approaches to their interpretation^[10]. For example, the issue of interpretation of a typewritten text, a logarithmic formula for assessing damage (provided by one or another subordinate institution), the evidence base of the participant in the administrative process, which may include oral explanations using non-standard speech turns, etc., and AI data processing.

In addition, the issue of data processing speed and constant data processing forms a conditionally "living organism" having a constant dynamic data structure, potentially updated with each interaction with the user^[11-12].

At the same time, at the level of doctrinal perception of AI for the purposes of public legal relations, it should also be noted that AI and its basic component element - machine learning - work according to the sample of cases that were previously developed by the relevant official.

It should be understood that the general codification of administrative law and process allows, for the most part, to put the activities of employees of relevant ministries and departments on "algorithmic rails"; at the same time, it should be noted that within each case there are both common features (inherent in a particular class of cases provided for by the codes) and distinctive features forming a subset of cases with generalizable properties.

It is necessary to understand that the issue of generalization of administrative cases, as well as their personalization for AI purposes, generates a number of problems, such as problems of the selection effect and formed bias.

Analyzing the problem of the selection effect, we see that an algorithm running on one database may cease to effectively implement its public functionality for the purposes of another department or generate data in the same segment, but with a random sample of the source data ^[13]. The issue of data bias is determined by the extent to which the applicable norm is used in administrative practice and whether there is a representative sample of facts formed, which is also systematically highlighted by both domestic and foreign scientists^[14-16].

The basic solution to these problems can be the following three tools (individually or in combination):

1. Generation of random situations/samples, which allows the most accurate formulation of the practice of AI decision-making, especially if the basis of this is the widest range of cases taken by various officials within the same process ^[17].
2. The use of counterfactual analysis of the case ^[18], which gives the generation of branches to the decisions made, and segments that influence AI decision-making; the information noise created by this sample and branches minimizes the number of unconventional decisions made by a person.
3. Introduction of the AI + human bundle, where AI acts as a means of monitoring random samples that highlight problem areas to an official for verification and evaluation of evidence.

It should be understood that this bundle can also preserve the risks of unethical decision-making, unreliability of the sample when making an AI decision (human factor), as well as the potential impact of AI on the basic rights provided for by the Constitution (Part 2 of Article 19). However, we understand that these actions will later make human biases or mistakes more noticeable.

In addition, it should be borne in mind that AI technology can be used not only for the benefit of the state, but also in opposition to the current measures provided for by legislation.

The whole complex of measures laid down when making an autonomous AI decision can be empirically identified, and, as a result, strategic countermeasures can be formulated provided for by the implementation of the principles of legal certainty ^[19].

Among these, it is possible to highlight the issues of creating information noise of the AI databases used (especially those located behind the contour of the formation of profile authorities), as well as the formation of behavioral adaptation when a controlled subject, who is not at risk of being included in the sample, is less "law-abiding".

These doctrinal studies of the integration of AI into the sphere of public legal relations are reflected in the practical public activities of the authorities of various countries:

The Dutch authorities, as part of the control of fraud in the field of social security, used the SyRi system ^[20], which, in fact, used a large array of data in its work, but was not sufficiently transparent in its data processing and the empirical results provided. As a result, the specified algorithm was banned for use by the authority due to the inconsistency of its principles with article 8 of the ECHR (European Convention on Human Rights) ^[21].

The Ministry of Labor and Social Policy of Poland used a highly specialized AI, which was supposed to categorize potential recipients of unemployment benefits ^[22].  At the same time, the binary processing of the AI solution passed the appropriate verification of the department's employees, however, due to negligence, in almost 100% of cases, the position of the AI was not disputed by the employee. As a result of the analysis of AI actions for transparency, the Constitutional Court of Poland recognized this product as unconstitutional ^[23].

The Unemployment Insurance Agency of the State of Michigan (USA) used a similar SyRi tool for the purpose of detecting fraud cases related to receiving unemployment benefits. At the same time, the specified AI carried out a completely autonomous decision-making regarding the recovery of benefits, without the possibility of appealing the decision within the agency ^[24], which subsequently entailed a trial and appropriate penalties to the AI developer^[25].

Based on the above, in order to form a balanced approach to the use of AI for the purposes of public legal relations, the following is proposed:

1. It is necessary to identify the sectors of AI application that do not entail a violation of the basic constitutional rights of citizens.
2. It is necessary to form test data polygons within the framework of profile FOI, with the possibility of rapid response within the framework of failures, errors, etc. allowed by AI and profile employees of the FOI.
3. It is necessary to use AI on the most representative sample of cases, eliminating the risk of predicted prosecution due to the poor quality of the sample of cases or weak administrative practices within the specified sector of legal relations.
4. It is necessary to introduce counterfactual explanations of AI in the actions of all participants in the process, in order to remove the problems caused by the technological nature of AI.
5. It is necessary to control the risks associated with AI in the context of the formation of countermeasures by intruders, by creating information noise in external databases or the formation of behavioral adaptation in the target category of citizens.
6. When implementing fully autonomous AI, it is important to take into account the prevailing negative international experience in terms of non-transparent and unconstitutional AI decision-making mechanisms, with the introduction of appropriate compensatory mechanisms in the Russian law enforcement field to ensure transparency of decision-making and its verification by a relevant FOIV employee.