In the fourth year of the global boom in generative artificial intelligence (hereinafter — AI), this article could undoubtedly have been written without our involvement. Yet human beings remain a vital element in this creative process.

Despite all the magic we witness when interacting with artificial intelligence, modern language models remain a product of mathematics. They operate on numerical representations of tokens — embeddings. At the same time, within the field of artificial intelligence there are different approaches, methods, and architectures that differ not only in their design principles, but also in the mathematical models and formulas underlying their operation. That is why such systems may process identical queries differently, exhibit distinct strengths and weaknesses, and make errors in their outputs. AI can misinterpret a prompt, rely on incomplete context, make faulty generalizations, or generate plausible-sounding yet inaccurate information.

Therefore, the process of implementing AI technology must be well-considered, guided by a clear plan, a sound methodology, and extensive research at both theoretical and applied levels.

It is precisely the absence of such an AI implementation plan that has given rise to numerous misunderstandings, apprehensions, and reservations among judges and representatives of the judiciary.

The cornerstone of this debate has become the idea of delegating quasi-judicial powers to AI to draft court decisions in certain categories of court cases. In his presentation "AI in National Judicial Proceedings: Evolution of Legal Positions," Yan Bernaziuk, Judge of the Cassation Administrative Court within the Supreme Court, refers to the draft resolution of the Cabinet of Ministers of Ukraine "On Implementing an Experimental Project on the Consideration of Certain Categories of Administrative Offense Cases Using AI," under which the Ministry of Digital Transformation of Ukraine proposes conducting a two-year experiment in which artificial intelligence modules within the systems of the Ministry of Internal Affairs of Ukraine and the Unified Judicial Information and Telecommunication System (UJITS) would carry out automated selection and analysis of administrative offense cases in the field of road traffic safety. According to the speaker, such modules would generate draft court decisions and submit them to a judge for signing or rejection.

Without disputing the idea itself, it should be noted that such an experiment must be implemented in clearly defined stages.

Stages of Technology Implementation

In general, the introduction of AI in the judiciary should be divided into stages determined by a systematic approach to verifying the technology’s capability to perform tasks of varying complexity.

First Stage (Research)

At this stage, data accumulation, verification, and analysis should take place. Based on the findings of this research, a reliable pipeline (data ingestion, processing, analysis, verification, output) will be established.

Specifically, the following studies should be conducted at this sub-stage:

A mandatory sub-stage must be the analysis of data represented in datasets. Before starting to generate decisions, clear mechanisms for data analysis must be developed. High-performance analytical systems should become a valuable byproduct of this work.

Draft decisions may be generated at this stage, but only as "alternative" decisions based on already adjudicated cases, and in order to safeguard the principle of legal certainty, they must be accessible only to a defined circle of persons (researchers, administrators, and authorized stakeholders).

Second Stage (Assistive)

Once the quality of analyzing case materials and court decisions reaches a sufficient level, it becomes possible to proceed to direct implementation as an assistive tool.

First and foremost, the assistive function should help judges draft the introductory and descriptive parts of a decision.

The next step should be preparing a full draft decision, provided that the results of the preceding stage demonstrate minimal divergence between the draft decisions proposed by AI and the decisions ultimately rendered by judges upon consideration of the cases.

Third Stage (Mediative)

Once a substantial body of court decisions has been prepared on the basis of AI-proposed drafts without undergoing major modifications or being overturned (i.e., upheld in force), the system may be applied at the stage of filing a claim. Thus, the plaintiff or defendant will be able to obtain a forecast regarding the resolution of a prospective dispute. This will yield the following beneficial effects:

Stage 1: Analysis of Existing Decisions

Unity of judicial practice is one of the fundamental prerequisites for public trust in justice, legal certainty, and the genuine realization of the rule of law. When courts apply consistent legal approaches in similar legal relations, citizens, businesses, attorneys, and public authorities are able to foresee the legal consequences of their actions and thus rely on the law as a stable and fair regulator of social relations. For this reason, the consistency of case law is not a purely technical aspect of court operations, but serves as one of the key guarantees of equality before the law and the right to a fair trial.

This approach is consistently upheld by international institutions. The Venice Commission regards legal certainty as an integral component of the rule of law, while the European Court of Human Rights emphasizes in its case law that "profound and long-standing" divergences in domestic jurisprudence may call into question compliance with the right to a fair trial. The ECtHR proceeds from the premise that certain divergences in judicial decisions may be a natural consequence of a multi-tiered judicial system; however, the State must maintain effective mechanisms to overcome such divergences and ensure coherent application of the law.

This is articulated with particular clarity in Opinion No. 20 (2017) of the Consultative Council of European Judges (CCJE) on the role of courts with respect to the uniform application of the law. The CCJE stresses that citizens reasonably expect equal treatment in similar cases and are entitled to rely on prior judicial decisions in order to foresee the legal consequences of their acts or omissions. Repeated delivery of contradictory decisions, by contrast, creates a state of legal uncertainty, erodes public trust in the judiciary, and weakens societal perceptions of justice.

For Ukraine, this issue is of not merely theoretical, but direct practical significance. At the level of the national judiciary, the Supreme Court is expressly vested with the function of ensuring the consistency and unity of judicial practice, and the High Council of Justice has repeatedly emphasized that this function is foundational to predictable, fair, and authoritative justice. In this context, unity of judicial practice appears not as formal uniformity of outcomes, but as an institutional precondition for trust in the court, equality of litigants before the law, and public confidence that similar cases are not decided in fundamentally different ways without a valid legal reason.

Consequently, unity of judicial practice should be viewed as one of the core guarantees of human rights, a pillar of legal certainty, and an indispensable element of the rule of law. It strengthens trust in the judiciary, ensures predictability of law application, enhances the quality of justice, and at the same time does not preclude the development of the law, provided such development is reasoned, consistent, and transparently justified in the judicial decision.

Modern AI tools make it possible to analyze judicial decisions with respect to:

One of the popular tools that can be used to analyze court decisions is Google’s NotebookLM service. The service is designed for deep research across up to 300 information sources. Such sources may include direct links to court decisions in the Unified State Register of Court Decisions (USRCD), as well as compilations of such decisions of up to 1,000 pages each. Assuming an average length of up to 10 pages per decision, any researcher can analyze up to 30,000 decisions — roughly equivalent to the total number of decisions rendered by a judge over an entire judicial career.

The complexity of such analysis lies in the preparatory process, which requires consolidating these decisions into combined documents. A professional approach must therefore rely not on off-the-shelf services for ad hoc studies, but on building a full-fledged analytical toolkit capable of aggregating USRCD data by specific criteria, retrieving court decisions, and conducting judicial practice research using predefined instructions (prompts).

Analyzing large volumes of data requires a balanced approach not only regarding the object of study, but also regarding its financial dimension. For example, the cost of input tokens for the gpt-5.4 model is USD 5 per million tokens, and output tokens cost USD 22.5. Analyzing decisions entails significant expenditure primarily on input tokens, as large volumes of material must be passed to the model for examination. Output token consumption, meanwhile, is substantially lower because it represents the aggregated result.

For comparison, a typical decision in an administrative offense case under Article 130 of the Code of Ukraine on Administrative Offenses (2.5 A4 pages, 12-pt font, 1,030 words) amounts to 2,644 tokens ( https://platform.openai.com/tokenizer , https://reyestr.court.gov.ua/Review/97737993 ). On average, about 350 small decisions can be analyzed for USD 3.

Analyzing more complex decisions, as well as appellate court rulings, will cost somewhat more — yet still far from prohibitive when the quality of public services provided by the judiciary is at stake. For instance, analyzing a standard judgment in a civil case (36 A4 pages, 12-pt font, 15,394 words) amounts to 38,974 tokens. Roughly speaking, USD 3 covers the analysis of about 25 such decisions.

It should also be borne in mind that before passing decisions for analysis, each decision must be enriched with additional metadata: names of the parties, date the case reached the court, date the full text of the decision was drafted, versioned representations of the legal provisions applied in the decision, and so on. These metadata make it possible to determine more accurately whether a genuine departure from practice occurred, or whether the divergence resulted from legislative amendments or specific litigants.

To optimize financial costs, it is advisable to employ a preliminary filtering approach (brute-force search) to pre-select the most relevant decisions. Another effective technique is stripping noise from decisions (removing the header, introductory boilerplate, standard procedural notices on appeal rights, court fee payment details, etc.).

Overall, the preprocessing pipeline looks as follows:

  1. Cleaning the text of technical noise.
  2. Normalizing the document structure.
  3. Extracting metadata separately from the main text.
  4. Condensing passages that have no material significance for the analysis.
  5. Isolating the relevant substantive sections for analysis instead of passing the entire text (for substantial cost optimization).
  6. Submitting the prepared data for in-depth analysis.

A large share of these operations can be performed on-premise — using the judiciary’s own computing infrastructure — which substantially reduces analysis costs. This is typically achieved through data normalization (conversion to Markdown), entity processing (regex/rule-based preprocessing with NLP rules), coarse filtering of relevant decisions (vector search for semantic similarity), and similar methods.

However, the final in-depth analysis should be entrusted to state-of-the-art frontier models, even though they are commercial products of third-party developers. The quality of legal analysis is not a stage where corners can be cut.

Once the object of study is prepared, a prompt must be designed specifying the markers on the basis of which consistency of practice will be evaluated. A judicial practice research prompt can be divided into three main blocks:

Technically, the output is structured as a JSON array, making it possible both to enrich the dataset for each individual decision and to build corresponding infographics.

As a result of running these analytical tools, we obtain not merely answers to a researcher’s questions, reports, or infographics, but a comprehensive cross-section and identification of systemic problems — which is the first step toward resolving them and initiating legislative amendments.

Implementation Path for the Judiciary

Every day, the information landscape is enriched with news of new artificial intelligence tools and new releases of flagship models.

Leading market players such as OpenAI, Anthropic, and Google have pulled far ahead in this race. Attempts by state institutions to build their own sovereign foundational models from scratch are unlikely to succeed. As leading global media report, OpenAI has secured agreements covering roughly 40% of global DRAM output.

In October 2025, reports emerged that OpenAI’s "Stargate" project signed agreements with Samsung and SK hynix for up to 900,000 DRAM wafers per month — nearly 40% of global production. Because memory manufacturers (Samsung, SK hynix, Micron) reallocated capacity toward HBM for AI accelerators, a shortage of conventional DRAM ensued, driving prices up by 50% or more ( https://www.tomshardware.com/pc-components/storage/perfect-storm-of-demand-and-supply-driving-up-storage-costs , https://www.reuters.com/world/china/ai-frenzy-is-driving-new-global-supply-chain-crisis-2025-12-03 , https://www.bloomberg.com/news/articles/2026-02-15/rampant-ai-demand-for-memory-is-fueling-a-growing-chip-crisis ).

Resource advantages pale in comparison with the technological lead of the AI market flagships. Even the mechanisms traditionally afforded by copyright and patent law are losing relevance amid the pace at which new technologies emerge. At this velocity, technology protection relies on trade secrecy, while classical patenting has largely lost its purpose.

Leading companies confidently release their earlier-generation models into open access because they know no competitor can catch up with their frontier development. For instance, Google released open-weight models such as Gemma with 27 billion parameters, and OpenAI released gpt-oss-120b with 120 billion parameters ( https://huggingface.co ).

Historically, companies occasionally disclosed the parameter counts of their commercial models. For example, an earlier generation, GPT-3 (whereas the current version is GPT-5.4), had 175 billion parameters. Today, companies generally do not disclose full architectural configurations, yet their capabilities are orders of magnitude higher than those of older models. By comparison, the model announced by the Ministry of Digital Transformation as the baseline for certain domestic projects is Gemma 3 with 12 billion parameters.

Under these circumstances, implementing AI in the judicial system requires a well-considered decision. By choosing the path of training domestic models from scratch, the judiciary risks losing precious time and falling far behind users who adopt state-of-the-art solutions. Reinventing the wheel while others are preparing for a hyper-jump is hardly the best strategy.

The security paradigm — currently one of the main factors holding back the adoption of modern tools — must be revisited in favor of legally binding safeguards and guarantees from service providers. Neither the legacy Comprehensive Information Protection System (KSZI) nor its successors in the form of security profiles and the Risk Management Framework (RMF) prohibit the use of commercial AI systems, though they do require a professional approach to information security architecture.

Ukrainian justice deserves state-of-the-art technology and has no time for dubious experiments without solid guarantees. The judiciary’s interaction with artificial intelligence should be guided by the principle articulated back in 2011 by Nobel Laureate in Physics Saul Perlmutter: "Treat it as a tool that supports thinking, rather than doing the thinking for you."