Principles of the Harmonised Use of Artificial Intelligence in the Judicial System

PREAMBLE

Conscious of the inevitability of the further joint development of justice and technology, believing that technological innovation should join in protecting the authority of justice, and seeking to make judicial procedures more human-centred, the High Council of Justice approves these Principles of the Harmonised Use of Artificial Intelligence Technology in the Judicial System (the “Principles”).

The use of artificial intelligence technologies must be compatible with the right to a fair trial, judicial independence, the equality of participants in proceedings and other fundamental rights and guarantees. Artificial intelligence is a tool in the service of justice, not a subject of it.

The introduction of artificial intelligence must not undermine public trust in the courts through opacity, algorithmic bias or technical error. Artificial intelligence technologies are to serve the efficiency and quality of justice without damaging its authority.

These Principles are intended to bridge the stability of legal procedures in the judicial system and the dynamism of modern technology. The Principles do not establish separate prohibitions; they set out the direction in which the use of artificial intelligence technologies should develop. The Principles do not replace the rules on managing security risks in the use of specific software products, or the guidance of manufacturers and individual institutions on their use.

NECESSITY

The final decision on using artificial intelligence technology when performing tasks is for each person to make independently, unless such use has been made mandatory by a superior, is dictated by the business process, or is carried out automatically by the automated information systems that support the work.

Explanation. Technical and supporting tasks (search, systematisation, translation, formatting) allow the technology broader autonomy. Tasks that directly form the basis for taking decisions require enhanced human verification.

COMPETENCE

Explanation. The effectiveness and safety of using the technology is determined above all by the competence of the user. Understanding the nature of the models (the probabilistic character of answers, dependence on the quality of the prompt, the possibility of hallucinations) is a precondition for applying all the other principles.

AWARENESS

Explanation. Artificial intelligence is merely an interface operating on a prompt-and-answer basis. The operator sets the direction of the answer in the prompt and can thereby influence the result obtained. The quality of the answer depends on the quality of the prompt, its nature, form and other factors that depend directly on the person asking. In essence this is a form of programming, in which the judge (or other court official) is the programmer setting the commands.

In the doctrine of intellectual property law, the results of such computer programming are protected by a right of a special kind (sui generis).

Explanation. Responsibility follows authority: whoever is empowered to produce the result answers for it, whatever tools were used to achieve it.

Explanation. This rule is needed as a counterweight to, and elaboration of, the mistaken approach embedded in Article 16 of the Code of Judicial Ethics, which provides that artificial intelligence cannot replace human assessment of evidence. Artificial intelligence allows algorithms to compare images quickly, extract relevant information and group data, giving the judge a convenient basis for taking a decision without infringing their monopoly on the administration of justice.

At the same time, the process of identifying data and presenting them to the judge implies that those data undergo some processing (alteration). In any event, the judge’s decision will depend on whether (1) the full volume of information is obtained and (2) that information has been altered.

The model in which the judge examines the evidence personally works in exactly the same way. There, the decision depends on whether (1) the judge examines all the evidence and (2) does so with due care.

Thus only one question remains open for debate: “Who is liable to make more mistakes when analysing a large volume of information: AI or a judge?”.

IMPARTIALITY

Explanation. The first law of robotics, formulated by the science-fiction writer Isaac Asimov, states: “A robot may not injure a human being or, through inaction, allow a human being to come to harm.”

Artificial intelligence is not a classic robot, since its assistive nature lies in carrying out the instructions of a human being — a human being with their good and, possibly, bad intentions. The foundation and at the same time the essence of justice is impartiality. That is precisely why, from the very outset of interaction with artificial intelligence, it must not be given any ethical instructions as to who acted well and who acted badly in the situation under examination, or which circumstances are negative and which positive.

Explanation. A model’s bias is not always obvious and may manifest itself in the selection of arguments, in emphasis or in the tone of the answer. Critical human appraisal of such results is the principal safeguard against transferring algorithmic bias into the application of the law.

Turning to another model in analogous circumstances may be done only in order to test its independent reaction to the same prompt, and must not aim at overcoming the restrictions revealed by the previous model.

Explanation. In individual cases a model may refuse to give an answer, invoking its own rules. Those rules may be relevant restrictions established by legislation, or they may be shaped solely by the developer company’s policies. Further attempts to obtain an answer from the model may produce an irrelevant result if the operator seeks to distort the input data or conceal the real purpose of the research.

RELIABILITY

Explanation. Where the materials to be examined for the purpose of taking a decision already contain legal conclusions of the same or another law-applying authority, drawn on the basis of those same materials, the earlier conclusions must be separated from the premises.

The basis on which artificial intelligence forms its results must be clean source data, without documents containing an assessment of the circumstances made by another authority. When using artificial intelligence, a two-stage analysis must be observed:

Stage 1

Materials must be supplied that do not contain decisions of a law-applying authority, unless those decisions are the subject of an independent challenge irrespective of such materials. For example, the artificial intelligence must draw its own conclusions on the basis of:

• the parties’ submissions (claim, response, objections and so on) without court decisions;
• the disciplinary complaint without the disciplinary inspector’s conclusion;
• the materials concerning the candidate without the rapporteur’s conclusion.

Stage 2

All materials must be supplied, including the conclusions of a law-applying authority of the same institutional affiliation.

Examples:

Tax disputes. At stage 1 the taxpayer’s materials that were the subject of examination are supplied, together with the statement of claim, the audit report and the tax assessment notice. At this stage the decisions of the tax authorities in the administrative appeal procedure are NOT supplied, nor are the decisions of lower-instance courts or those made in earlier examinations of the case (if any, or if the case was remitted for fresh consideration).

Once the results of that analysis are obtained, they are passed on together with all the materials for a second assessment, with the task of identifying errors in the earlier conclusions.

Appeal and cassation review, fresh consideration. At stage 1 only the case materials are supplied, without the decisions of the first-instance and appellate courts (if any) or of courts that have already examined the case. At stage 2 all materials and decisions are supplied.

Disciplinary proceedings against a judge or prosecutor. At stage 1 the materials of the disciplinary proceedings are supplied without the conclusion of the disciplinary inspector or of the Qualification and Disciplinary Commission of Prosecutors. At stage 2 the materials are supplied with all conclusions and decisions.

Building the judiciary. At stage 1 all the materials obtained concerning the candidate are supplied. Where the procedure is conducted by the High Council of Justice, at stage 1 the decisions of the High Qualification Commission of Judges of Ukraine containing its own conclusions on the candidate’s integrity are not supplied. At stage 2 all materials are submitted.

This approach will yield clean baseline research results unencumbered by the conclusions of other law-applying authorities. At the second stage it will be possible to test the fallibility of the clean conclusions against the qualified opinion of an authority that is making, or has made, the same analysis.

Explanation. Models are capable of generating plausible but inaccurate or non-existent information (so-called hallucinations), in particular invented citations of court decisions or non-existent versions of legal provisions. Verification against the primary source is the minimum reliability standard for any legally significant result.

Explanation

In certain categories of complex dispute, or in tasks requiring the preparation of reports, it is worth ensuring a mechanism for cross-checking results with different models. For example, if it is necessary to analyse case materials on the calculation of servicemen’s monetary allowances for different periods, identical tasks are sent for analysis to different models. Ideally a quorum of 3 models (ChatGPT, Claude, Gemini) should be secured. Once identical results are obtained, they are taken into use. If there are discrepancies, all three results are passed to one of the models for verification. If two results were similar and one differed, the matter is passed to the model whose result coincides with that of another.

Despite the higher cost, this technique will substantially reduce risk and may prove cheaper and faster than involving a person.

Explanation

When using artificial intelligence, the recommended prompts and SKILLs should be used where such prompts and SKILLs have been created in advance and recommended for use within the body or sector. Uniformity of methodology ensures that results are comparable and predictable across homogeneous tasks.

To that end, it must be possible to obtain from the person who created or supplied such documents a confirmation as to whether or not artificial intelligence was used in creating or altering them.

Where documents were created or altered using artificial intelligence, this should be communicated to the model used for their further examination.

TRANSPARENCY

Explanation. Transparency does not mean that every instance of use must be labelled, but it does presuppose the ability to reconstruct the logic of the work at any moment. It is precisely that ability which is the condition for preserving public trust should doubts arise as to the quality of the result.

CONTROLLABILITY

Explanation . Artificial intelligence cannot be entrusted with taking decisions in any procedure where such a decision is final and will affect rights and obligations with no possibility of correction or appeal. This concerns not instances of analysis, but instances where decisions are prepared without detailed human analysis.

CONFIDENTIALITY

For work with case materials or other sensitive information, only local or specially licensed solutions may be used, whose providers have disclosed the terms of data protection and model training.

ADAPTABILITY

Explanation. The dynamism of technology declared in the preamble requires that the guidelines themselves remain a living document. Review of the Principles must be based on analysis of the practice of their application, in particular the risks and achievements identified.