Artificial intelligence and justice: a systematic literature review and future research perspectives on Justice 5.0 Open Access

AI innovation is advancing rapidly. Projections suggest that in the next decade, AI technology will become widespread in homes, offices, businesses and communities, permeating almost every facet of our lives. Today, a series of “smart” devices (smartphones, smartwatches, robot vacuum cleaners, self-driving cars and drones) are pervading different aspects of daily life, and this tendence is expected to grow. AI plays a significant role in various sectors, including robotics, technology, healthcare (especially in medical diagnosis and surgery), transportation, military operations, security, government and public administration, finance and marketing (Walters and Novak, 2021). Consequently, also its usage in the legal domain is gradually increasing; in the last decade, the debate on the topic has particularly heated up (Castro and Guimaraes, 2020). Enhancing legal systems with AI technology (LegalAI) is not only a challenge that promises a variety of advantages but also discloses a series of unintended side effects to be considered. Innovation with AI in the public sector and justice system is crucial for enhancing efficiency, transparency and accessibility of legal processes (Mergel et al., 2023; Sedkaoui and Benaichouba, 2024 and Spalević et al., 2024). Its radically transformative impact on the conception and structure of legal processes, along with its ethical and legal implications and the resistance to change from key stakeholders, positions it as a potentially disruptive innovation in the justice sector (Păvăloaia and Necula, 2023). Indeed, automated systems can handle routine tasks such as document review, case management and legal research, allowing legal professionals to focus their time and resources on more complex and strategic aspects of their work (Razmetaeva and Razmetaev, 2021). Additionally, AI algorithms can analyse vast amounts of legal data and precedents to provide insights and predictions that can inform legal strategies and decision-making. This predictive capability can help identify trends, patterns and potential outcomes, leading to more informed and equitable judicial decisions (Quezada-Tavárez, 2021). Moreover, AI technologies have the potential to enhance access to justice by providing affordable and accessible legal services, particularly for underserved communities and individuals with limited resources. Overall, the use of AI in the justice system holds the promise of improving efficiency, accuracy and fairness, ultimately contributing to a more effective and accessible legal system (Rubim Borges Fortes, 2020). The automation brought by big data analytics, machine learning and AI systems could be also effective for improving security, preventing crime and terrorism attacks and reconsidering fundamental questions of criminal justice (Završnik, 2021). On the other hand, a significant concern is the potential for bias and discrimination inherent in AI algorithms. These algorithms are trained on historical data, which may reflect systemic biases and inequalities present in society. As a result, AI systems may perpetuate and exacerbate existing disparities in the justice system, particularly for marginalized communities (Blount, 2022). In addition, there are concerns about the erosion of human judgement and discretion in legal decision-making processes as AI systems become more prevalent (Rubim Borges Fortes, 2020). Furthermore, the rapid advancement of AI technologies may outpace regulatory frameworks and ethical guidelines, leaving legal systems ill-equipped to address emerging challenges and risks (Hoffmann-Riem, 2020).

The challenging combination between AI innovation and justice can be considered as a multidisciplinary problem, it requires expertise from various fields such as technology, law, ethics, sociology and policymaking to address the complex challenges and implications involved (Sousa, 2024).

A theoretical framework is required to summarize the problem by mapping the existing literature, serving as a foundational reference for future research developments. Addressing this gap, this article aims to map the literature on AI and justice across various disciplines and establish key theoretical concepts that delineate the domain, laying a robust foundation for subsequent research initiatives. To achieve this goal, the following research questions are considered as a starting point:

A systematic literature review (Kraus, 2022) of studies published in peer-reviewed journals and indexed in the Scopus database is proposed. The authors analysed the initial set of 287 selected papers manually and also employed various scientometric techniques, including qualitative and quantitative analysis through mapping, visualization, bibliographic coupling and comprehensive examination of the full texts, evidencing: (1) the main issues and the main variables used for the design, management and use of AI innovation technology in justice; (2) the role of technology and (3) the ethical concerns.

The contribution of the paper is then threefold: (1) mapping and summarizing the literature on the topic of AI innovation and justice highlighting the main design and management dimensions as well as the main algorithmic approaches; (2) providing a systematic classification of the application field of AI innovation in justice systems in the big existing discussion panorama and (3) proposing an agenda for future research by confronting the main design variables with the existing research discussions and approaches.

The contribution of our article is useful both for researchers in the field of justice and law and in the field of technology advances, informatics, innovation and AI systems. Finally, it is important for policymakers in courtrooms and government to understand how to face and dominate the digital transformation in the justice systems.

The paper is structured as follows: in Section 2 the methodology for the systematic review of the literature is presented. In Section 3 the results of the bibliometric analysis as well as the results from the analysis of the full texts are described. In Section 4 a discussion of findings and an agenda for future research is provided. The last section is for conclusions.

For the purposes of this review, artificial intelligence (AI) is defined as “the capacity of machines or software systems to perform tasks that normally require human intelligence, such as perception, reasoning, learning, and decision-making” (Russell and Norvig, 2016). This comprehensive definition includes both symbolic AI (e.g. rule-based systems) and data-driven approaches (e.g. machine learning and deep learning), which have been the primary engines of innovation in the field of legal informatics.

This interpretation also aligns with the definition adopted by institutional actors such as the European Commission’s High-Level Expert Group on AI, which describes in their deliverables AI as “systems that display intelligent behavior by analyzing their environment and taking actions—with some degree of autonomy—to achieve specific goals” (European Commission, 2024).

To provide a strong theoretical foundation, this review draws on the multidisciplinary perspective offered by Dwivedi et al. (2021), who explore the evolution of AI alongside its societal impact and policy considerations – dimensions particularly relevant for justice systems navigating the twin imperatives of innovation and institutional legitimacy.

This systematic literature review has been conducted by using the PRISMA methodology, as described by Page et al. (2021), combined with a theory-building review of literature exposed by Swanson and Chermack (2013), Kraus et al. (2022) and Öztürk et al. (2024). To select the set of articles for the systematic literature review, the following search string was introduced in Scopus on January 29th, 2025, looking for relevant keywords in titles, abstracts and keyword sections of each article and obtaining a set of 276 documents:

(TITLE-ABS-KEY (machine AND learning) OR TITLE-ABS-KEY (artificial AND intelligence) AND TITLE-ABS-KEY (criminal AND justice) OR TITLE-ABS-KEY (e-justice) OR TITLE-ABS-KEY (courtroom) OR TITLE-ABS-KEY (e-process)) AND (LIMIT-TO (DOCTYPE, “ar”) OR LIMIT-TO (DOCTYPE, “re”))

The selection of keywords was grounded in a preliminary exploratory analysis of the literature to ensure comprehensive coverage of both technological and legal dimensions of AI applications in justice. Terms such as “machine learning” and “artificial intelligence” were chosen to capture the technological core of AI innovation, while keywords like “criminal justice”, “e-justice”, “courtroom” and “e-process” reflect the procedural and institutional contexts of legal systems undergoing digital transformation. These terms were selected to encapsulate both academic and applied research, covering predictive algorithms, legal informatics and digital legal processes under the Justice 5.0 paradigm. The inclusion of title, abstract and keyword fields was meant to broaden the scope and reduce retrieval bias, enabling a multi-angle review suitable for a multidisciplinary approach.

The PRISMA methodology for new systematic reviews was adopted; the flow diagram is reported in Figure 1 (see Moher et al., 2009).

The methodology comprises four distinct stages: identification, screening, eligibility assessment and inclusion. Initially, an identification process using the specified query is conducted, followed by a thorough review of the abstracts to detect duplicates and assess their relevance to this review’s objectives. Subsequently, four duplicated records and 132 documents that did not align with the review’s scope were excluded, resulting in a post-screening dataset of 151 papers. Further analysis of the full texts led to the retention of 146 eligible papers, while six were deemed inappropriate based on predefined inclusion/exclusion criteria reported in Table 1.

The evaluation of full papers was conducted collaboratively by all authors and their assessments were shared within the team. Additionally, all selected papers underwent analysis using the R-studio software app Biblioshiny (see Aria and Cuccurullo, 2017). The articles included in the systematic literature review were then subjected to analysis using a theory-building approach, as outlined in prior research. Theory building (Paul et al., 2023) is regarded as a pivotal outcome of literature reviews, facilitating the integration of diverse research streams to construct theoretical frameworks.

As an initial phase of the systematic literature review, we utilized the R-studio software in conjunction with the Biblioshiny app to analyse the selected papers. These tools enable the exploration of various facets, such as authors’ productivity, citation and impact metrics, source relevance, topic trends and keywords, the thematic progression of topics over time as well as co-occurrence networks and factorial analysis related to topics and keywords. In the following, only the significative analysis is reported and described. Figure 2 shows a high increase in the number of publications related to AI & justice from 2018.

Figure 3 presents the most relevant sources in terms of citation frequency within the dataset. Notably, the majority of the scientific output is published in computer science journals, highlighting the strong technological drive that characterizes this field. From a multidisciplinary perspective, top-ranked journals such as Artificial Intelligence and Law and Artificial Intelligence and Society stand out, illustrating the increasing convergence between legal scholarship and AI research. This convergence confirms the interdisciplinary nature of the topic and reinforces both the relevance and the scientific authority of the literature selected for the systematic review.

This further confirms post hoc that the dataset was selected in accordance with existing literature and the target readership of prestigious journals.

Finally, the semantic analysis of the contents of the articles has initially been conducted through the investigation of authors’ keywords; then, the investigation of the themes and topics emerging from the dataset of the selected papers was conducted. In detail, Figures 4–6 gives the directions for a major classification of interesting topics.

Figure 4 illustrates the evolution of keyword frequency over time, revealing a substantial and consistent increase in publications that reference both “artificial intelligence” and legal system–related terms (such as “justice”, “courtroom” and “legal process”) starting from 2018. This growth signals a burgeoning academic interest and confirms the timeliness and significance of this review.

Figure 5 shows a word cloud that visualizes the frequency of terms used in the selected literature, excluding those used in the search query. Key terms such as “natural language processing”, “big data”, “forecasting”, “decision making” and “ethics” emerge prominently, underscoring the technical, predictive and normative dimensions of the discourse. These terms helped refine the classification of major themes in subsequent analysis, especially in constructing the thematic map (see Figure 6).

Figure 6 displays the thematic map generated through the Biblioshiny tool, based on the co-occurrence and centrality of keywords across the selected literature. The map organizes themes into four quadrants according to their relevance (centrality) and development (density):

1. Motor themes (high centrality and high density): These are well-developed and crucial to the structure of the research field. Notable clusters include ethical technology, human–computer interaction and algorithmics, highlighting the robust interdisciplinary integration of ethics, user interaction and computational design within legal innovation contexts. These themes drive the core discussion around the responsible implementation of AI in judicial systems.

2. Basic themes (high centrality and low density): These themes are usually related to topics largely explored but not directly related to the topic of the survey, because they usually referred to more basically knowledge.

3. Niche themes (low centrality and high density): These are specialized and deeply developed areas with limited external connections. Although this quadrant is sparsely populated in this study, it includes a focused cluster around decision support systems, suggesting a technically mature but somewhat siloed area of research with significant potential for integration into broader AI and justice debates.

4. Emerging or declining themes (low centrality and low density): These themes can be excluded because of their low interest from the scientific community.

This mapping confirms that the field is coalescing around a few well-developed, impactful clusters while also revealing underexplored areas ripe for future research, particularly in the ethical governance of decision support tools and the systemic integration of AI in legal practice.

Our analysis of the papers highlighted three main thematic macro-areas for describing the relations and interactions between AI and justice (see Figure 7), that are related to predictive justice, human–machine combination and the emerging and debated technology of the robot judge. These areas emerged from the convergence of qualitative content analysis and the bibliometric results produced through Biblioshiny, particularly from the motor themes identified in the thematic map (see Figure 6):

1. Predictive justice emerged from the cluster focused on machine learning, big data and legal analytics, encompassing AI techniques that aim to forecast legal outcomes, enhance decision-making and support pre-trial or trial judgements through data-driven insights.

2. Human–machine combination drew from themes such as ethical AI, human–computer interaction and AI explainability. This area focuses on the synergies and tensions in shared decision-making environments where AI tools assist – but do not replace – human legal actors.

3. Robot judge was conceptualized by extrapolating from discussions surrounding algorithmic justice, automated decision-making and AI governance. While it builds on both predictive justice and human–machine combination, this area speculatively explores the potential replacement of human judges by AI systems. As such, it delves into philosophical and ethical concerns closely related to algorithmic opacity, legal personhood and fairness. While the first two perspectives focus primarily on decision-support systems grounded in the paradigm of weak AI – designed to perform specific tasks without consciousness or understanding – this third perspective shifts the attention towards the notion of strong AI, characterized by human-like cognitive abilities, such as autonomous reasoning, learning and awareness. Although still hypothetical and not realized in practice, strong AI raises profound questions about the future of judicial authority and the legitimacy of non-human decision-makers (Bory et al., 2024).

The term “predictive justice” refers to ICT tools utilized by jurists, based on jurisprudential datasets and aided by sorting algorithms and neural networks, enabling them to foresee the likelihood of success statistics in any legal dispute. The machine will emulate judicial reasoning, whether it is inductive, deductive, based on precedent, by analogy or by principles, utilizing quantitative techniques.

Under this perspective, four main topics are introduced for describing the predictive justice major theme: Big Data, language processing algorithms, predictive models’ application in criminal justice and predictive models’ application, more generally, in law. A description of each topic is introduced in the following.

Big Data. The formal definition of Big Data is about a large volume of data that is selected, extracted and processed to create a solid statistical basis on a particular topic to discover and contextualize patterns and correlations (Jagadish, 2015). The rise of Big Data is fostering a shift towards behavioural optimization and the concept of “personalized law”, wherein legal decisions and rules are fine-tuned for optimal outcomes. This approach tailors the law to individual consumers through the analysis of historical data, lead on the evidence that data analytics is a promising tool for automating and innovating existing legal services. Nevertheless, the application of Big Data in the legal realm comes with significant limitations and risks (Devins et al., 2017).

Big Data, in conjunction with algorithms and ML, has emerged as a focal point in the strategies and initiatives of intelligence, security, defence, anti-terrorism, crime policies and law, driving a new perspective generally defined as “algorithmic justice”, to build automated decision-making tools in the justice sector (Završnik, 2021), also widely defined as “Legal Tech”. A recent survey by Park et al. (2021) shows that, in the last five years, the researchers’ interest has remarkably increased, focusing on a great variety of techniques for data analysis in the justice sector, such as data and text mining, classification and clustering, deep and ML and prediction techniques. The authors focused also on the geographical distribution of investments in Legal Tech: the world’s largest economic investment in legal sector automation relates to Asian countries (50%), followed by Europe (31%), America (17%) and Africa (2%), evidencing a sort of scepticism in some continents, towards the use of these systems.

Under this perspective, a long list of limitations and risks associated with such a technology are described in the literature. First, Big Data must necessarily be stored in databases, which together with algorithms represent human artefacts. About the database, the first question concerns how data are collected, cleaned and prepared. Main sources can be related to court records, civil petitions, laws, legal judgements, legislative documents, police or victims reporting. As a result, the quality of data varies. In instances of low-quality data, providing more quantity of data to the data-crunching machine does not improve the situation. In cases of poor data quality, the principle of “garbage in, garbage out” applies (Završnik, 2021), partially invalidating the result of the analysis. About algorithms and statistical techniques, it has been underlined that the statistical modelling applied in justice has its roots in other domains (informatics, engineering, earth science and marketing) where ground-truth data are more reliable and the tolerance for false predictions differs. Therefore, errors in analysis’ results may vary between domains: in predictive justice, for example, treating innocent individuals as criminals poses a serious threat to fundamental liberties. Moreover, the procedure of an algorithm is not neutral but instead reflects human choices about data selection, connections, inferences and interpretations. In the literature, consequently, it has been pointed out that there exists a difficulty of purifying algorithms from non-objective influencing factors like race, ethnic background or sexual preference and actuarial instruments that were used as specific predictors, including a sort of discrimination within the algorithm (Harcourt, 2015).

A variety of other problematic aspects are mentioned in the literature: the topic of data transparency in practical applications is mentioned by Zolea (2023), while open problems about legal document anonymization are investigated by Csányi et al. (2021), the epistemic fragilities of heuristics predictions and the risk of turning into rights violations are described by Lettieri (2021), while a great overview about troubles in avoiding automation of bias in ML techniques is introduced in Varona et al. (2021).

Finally, a series of practical cases enlivens the debate in the field: for example, in Brazil, a web-based visual analytics system was experimented with to support the analysis of legal documents (Resk et al., 2023); while applications of the prevention of criminal offenses and the use in pre-trial and trial investigation of Big Data analytics are described in Demura and Kepla (2021). Moreover, the automatization of data transcription for the Indian Supreme Court’s documents is presented by Vaissnave and Deepalakshmi (2020), while an application for predicting appeal cases in the Indian Supreme Court is introduced by Sivaranjani et al. (2021). For a deeper overview of the field, the reader can refer to the surveys of Završnik (2021), Devins et al. (2017), Park et al. (2021) and Rosili et al. (2021).

Language processing. LegalAI aims to enhance legal systems with AI technology, particularly through natural language processing (NLP). NLP models are increasingly being applied in various aspects of the legal field to process and analyse legal texts, documents and communications; becoming especially useful in building decision support systems, digital archives and supporting tool in the legal context, facing the emerging and changing needs of societies and responding to social, political, economic and technological changes. Given its heavy reliance on written text, law stands to benefit significantly from NLP; for this reason, many researchers in LegalAI are dedicating their efforts to applying NLP to legal tasks. The most diffused applications of NLP in a legal context are:

1. legal document summarization: NLP techniques are used to summarize lengthy legal documents such as contracts, court opinions and statutes (Kanapala et al., 2019; Deepali et al., 2021);

2. legal information retrieval: NLP models can enhance the search and retrieval of legal information by understanding natural language queries and returning relevant legal documents, cases or statutes (Sansone and Sperlì, 2022);

3. contract analysis and review: NLP is used to extract and analyse information from contracts, including clauses, terms, obligations and deadlines; also detecting inconsistency (Antos and Nadhamuni, 2021);

4. legal entity recognition: NLP models can identify and extract entities mentioned in legal texts, such as names of parties, dates, amounts and legal concepts (Vardana et al., 2021);

5. legal question answering, legal chatbots and virtual assistants: NLP-powered question-answering systems can understand and respond to legal queries posed in natural language; while NLP-driven chatbots and virtual assistants are used to automate routine legal tasks, provide legal information or guidance to clients and assist with basic legal inquiries (Jorge, 2023).

6. legal language translation: NLP models can translate legal documents and texts between different languages, facilitates cross-border legal transactions.

7. legal data analysis and prediction: NLP techniques are integrated into predictive analytics models; for this aspect the reader can refer to the following paragraphs.

The introduction of deep learning techniques like the pre-trained processing models (PLMs) facilitates the spread out of practical applications in the field. However, utilizing PLMs for legal tasks remains challenging due to the extensive nature of legal documents, which typically consist of thousands of tokens, a quantity far exceeding what traditional PLMs can effectively process (Mumcuoğlu et al., 2021). In the application of NPL in law, different problems must be considered, for example, how to extract features from a legal text (also defined as named entity recognition, NER). NER systems specifically devoted to the legal domain have been studied in the literature under different perspective (Leitner et al., 2019); for example, Elnaggar et al. (2018) demonstrated that training a NER system first on non-legal data, then refining it with additional training on legal data, is more effective in terms of performance than directly training NER only on legal data. Another crucial aspect of extracting features from legal texts involves identifying specific legal attributes at the word or sentence level, such as facts, obligations, prohibitions and principles (Sleimi et al., 2018). Finally, a great challenge is also related to identifying logical relationships between texts, such as deciding whether a given text is based on law as well as if there is an entailment between a given query and a law article (Nguyen et al., 2018). A comprehensive overview of the NLP principles in the legal domain is described by the survey of Chalkidis and Kampas (2018).

In addition to theoretical studies on algorithms to better represent and analyse legal language, numerous studies have focused on the practical application of such systems to legal contexts. For example, Xiao et al. (2021) propose a pre-trained language processing model, called Lawformer, for Chinese legal long documents understanding, and test it on a variety of LegalAI tasks, including judgement prediction, similar case retrieval, reading and answering legal questions.

The study by Ferreira and Ruiz (2022) attempted to develop a model capable of predicting the judicial outcomes of the Court of Justice of São Paulo by analysing, translating and inserting data flows extrapolated from previous jurisprudential cases into the machine. This study can be considered the first to develop a predictive model using Brazilian Portuguese or the Brazilian language. Tagarelli and Simeri (2021) present a deep learning framework named LamBERTa (Law article mining based on BERT architecture), which is specifically designed for civil-law codes and specifically trained on the Italian civil code. In detail, the system is based on a pre-trained BERT (Bidirectional Encoder Representations from Transformers, Devlin et al., 2019) on the Italian civil code or its portions for law article classification and retrieval. Note that BERT is renowned as one of the most efficient frameworks in many NLP tasks. Mumcuoğlu et al. (2021) introduce the first work with an application for the legal system of the Republic of Turkey, codified in Turkish. They predict the rulings of the Turkish Constitutional Court and Courts of Appeal, using only fact descriptions as dataset; comparing the performances of different analytical techniques: decision trees, random forests, support vector machines and deep learning methods (gated recurrent units and long short-term memory networks); demonstrating that outcomes of the Turkish courts can be predicted with high accuracy with deep learning approaches.

Finally, Chen et al. (2022) have worked on classifying legal texts. They tried to identify which is the best ML model for effective classification. The authors also investigated how the domain concept size affects classification performance, finding that the top 5% of domain concepts can generate the most effective and robust random forest classifier. Finally, a framework with guidelines on four factors, including data, performance, computation and interpretation, was proposed for model selection in domain-specific text classification. A literature review about other applications in the field is introduced by Robaldo et al. (2019).

Predictive model applications in criminal justice. Algorithmic justice and LegalAI spread out generally in the field of criminal justice for the introduction of predictive model in legal activities both in international and within domestic’s courts. Among the most frequent criminal application areas in which predictive algorithms can contribute, according to specialized literature, it is possible to cite the crime prevention, the pre-trial restrictions and the crime trial (Demura and Kepla, 2021). Crime prevention consists in the prevention of criminal offenses. This approach is implemented by utilizing various information tools to prevent criminal acts, such as identifying potential locations where they may occur or individuals who might participate in them. They include the so-called “precautionary police control” tools to prevent certain types of offenses with elements of regularity, such as burglary, street violence, vehicle theft/carjacking or unexpected events like terrorisms attacks. Different real projects can be cited as examples: in 2015 the city of Trento (Italy) developed a project on ICT for knowledge-based and predictable urban security, introducing predictive methods analysing different data sources and predicting criminal acts with a success rate of about 60–65%. Similarly, in the UK a pilot project consisting of using an AI-based software projection, called PREDPOL, to prevent possible location of burglary, theft and attacks, showed that the predictions became true in 78% of cases, as compared with 51% of the predictions made with the use of conventional methods. Social network analysis (SNA) is also a branch of AI technology that helps examine relationships between individuals and patterns in social networks, useful for analysing networks between criminals and their individual status in organized crime (Wang, 2020). Finally, the European Ethical Charter on the use of AI in the judicial systems underlines the importance of such useful tools, i.e. the travel ban based on big data analysis, which gathers and analyses data on potential terrorists to prevent the commission of terrorist acts as well as algorithms used to detect fraud or money laundering between the most used and useful tools.

The US, as one of the leading adopters of AI in the justice system, has integrated such technologies into the process of determining pre-trial restrictions. For instance, researchers at Stanford University’s Computational Policy Lab have devised an algorithm to aid judges in deciding whether to detain or grant bail to a defendant as a pre-trial restriction. After examining approximately 100,000 procedural documents pertaining to pre-trial restriction decisions, the developers observed significant variation among judges’ tendencies, with some releasing defendants on bail in 90% of cases while others did so in only 50%. This programme facilitates a fair assessment of risks, resulting in detention being selected for a significantly smaller proportion of individuals. A similar use of this and other models is envisaged in Ukraine, where the choice of a precautionary restriction with the use of AI algorithms significantly improve the quality of the procedural acts and accusation. Indeed, the analysis has shown that, in most cases, the investigator/prosecutor’s request for a precautionary restriction contains only formal law references to the risks foreseen, not supported by factual data. Moreover, it is observed that investigating judges, when considering such requests, put the interests of the criminal case before the interests of the accused and satisfy the requests in 90% of cases. Moreover, Demura and Kepla (2021) contend that employing AI in determining pre-trial restrictions helps circumvent the judicial practice prevalent in Ukraine, where the most stringent restriction during the preliminary investigation stage – detention – is frequently favoured. Therefore, the authors showed that the use of AI helps to avoid the subjective factor and analysed the investigator’s/prosecutor’s request in an impartial manner.

Furthermore, a recent study of Khoei and Sight (2024) introduce the concept of Emotional AI, capable to detect, analyse and understand human emotions. This innovation represents a groundbreaking opportunity for enhancing various aspects of policing and criminology, in particular in crime detection and prediction.

In contrast, the topic of predictive policing has aroused a fair level of concern among scholars and human rights advocates. In general, courts use such systems to assess the likelihood of recidivism or absconding of those awaiting trial or offenders in bail and parole procedures. For example, the well-known Arnold Foundation algorithm, which is implemented in 21 jurisdictions in the US (Dewan, 2015), uses 1.5 m criminal cases to predict the behaviours of defendants in the pre-trial phase. Similarly, Florida uses ML algorithms to set bail amounts. The authors questioned how such predictive tools could discriminate between people belonging to the lowest income strata and the least emancipated sectors of the population (Završnik, 2021).

Finally, Mitchell et al. (2020) have focused their attention on the role of the jury in the criminal trial and on how this role can be improved thanks to the use of AI systems. The central and most challenging task facing juries in a criminal trial must be to reach a verdict based on their evaluation of the evidence. Regardless of which model best represents the jury’s reasoning, the authors sought to discover whether the decision-making process is amenable to support by ML. That misconceptions can be avoided with such an application of ML in the future is an attractive prospect. An obvious difficulty facing any attempt to allow ML to determine a verdict is that the prosecution and defence evidence presented at trial must be mathematically processed.

Indeed, predictive algorithms have been found to produce both false positives and false negatives that have misled decision-makers in sensitive matters such as the deprivation of liberty of an innocent individual. Then it would be inaccurate to convey a decision that affects people when there is even the slightest margin of error in obtaining the result. In this direction, the study of Varona et al. (2021) is related to the identification of current gaps in achieving fairness in the context of AI-based predictive systems in criminal context, to mitigate the risk of application of discriminatory decisions.

Application in law. Predictive models in law utilize data analysis techniques to forecast legal outcomes, assess risks and provide insights for legal decision-making. These models leverage historical case data, statutes, regulations, court opinions and other legal documents to identify patterns, trends and probabilities, also using NLP techniques as a base for analytics. The most common applications of predictive models in law, except for the application in crime justice already discussed, are described in the following:

1. case outcome and sentence prediction: Predictive models can analyse past case data to forecast the likely outcome of ongoing or future cases. This information can help lawyers and clients make informed decisions about litigation strategies (Cui et al., 2023);

2. legal research enhancement and risk assessment: By analysing large volumes of legal texts, predictive models can assist legal researchers in finding relevant cases, statutes and precedents more efficiently. This can save time and improve the quality of legal research; moreover, predictive models can assess the potential risks associated with various legal strategies, such as the likelihood of success in court, the probability of regulatory enforcement actions or the potential financial implications of legal decisions (Schwarcz et al., 2023);

3. resource allocation: Law firms and legal departments can use predictive models to optimize resource allocation by identifying cases or matters that require more attention or resources based on their likelihood of success, complexity or potential impact (Vasconcelos et al., 2023).

Predictive models have already been developed in numerous practical experiences in sentence predictions.

Brazil has brought to light several studies on the implementation of AI in different law areas. The study conducted by Bertalan and Ruiz (2022) resulted in the creation of a labelled dataset comprising court cases involving instances of murder and corruption crimes. This dataset holds potential to facilitate future research endeavours exploring novel predictive methodologies. Moreover, the tool may prove valuable for analysts seeking to investigate whether variables such as the judge’s gender or the geographical municipality of the case could impact the overall labelling of legal documents. Additionally, the study demonstrated that ML-based predictive algorithms consistently achieved accuracy rates exceeding 95% in predicting judicial decisions based on the textual content of court rulings.

Sukanya and Priyadarshini (2023) proposed a modified hierarchical-attention network for pre-processing, feature extraction and judgement prediction in Indian civil processes, trained on 10 types of real-time criminal cases and obtaining promising results. In India, an additional model has been developed by Sivaranjani et al. (2021), capable of classifying and predicting the behaviours of the Supreme Court (in accepting or dismissing an appeal). The model is based on the analysis of based cases from the last 20 years and on various ML algorithms. Another tool is proposed by Zahir (2023) for the Moroccan Court of Cassation. The model collects the greatest number of cases and automatically predict legal case outcomes from written description of the events in Arabic using deep learning. Another study conducted in the US regarding the prediction of cases of violation of online privacy highlighted how the classification and regression tree (CART) technique had the best classification performance and how each factor significantly influences the creation of the model (Park and Chai, 2021).

Furthermore, some applications are experiment in resource allocation. For example, de Oliveira et al. (2022) attempt to predict the duration for the magistrate to issue a ruling using AI. The tool is tested on a regional labor court, where different ML techniques are experimented with and compared: the Adaboost technique is better adapted to the characteristics of the problem, allowing the development of an inference model capable of predicting, with a low error rate and good performance, the times needed to issue a sentence in judicial cases in the region, with 84% accuracy. Vasconcelos et al. (2023) employed jurimetrics analysis to assess the impact of various factors on procedural timelines, encompassing budgetary constraints and judicial management styles. The dataset was categorized into three primary clusters: procedural, administrative and organizational elements. The challenge facing the courts is to optimize case resolution rates to ensure the effective functioning of society. The efficiency analysis of the judicial system is conducted by leveraging ML methodologies, focusing on expenditure and other variables. The objective is to assess if resources are allocated and utilized effectively, encompassing both human and material resources and to investigate potential correlations between productivity and resource allocation strategies.

Finally, other two decision support systems have to be noticed. Firstly, in the Brazilian context, a project named INSIDE (Integration, Analysis, Visualization of Data for Investigation) was developed to create the INSIDE platform, a robust big data platform designed for crime analysis (Silva et al., 2021). This platform enables the development of scalable, real-time and data-driven applications. Its primary aim is to expedite the analysis of criminal evidence, a process known for its time-intensive nature, by effectively processing large volumes of data within the realm of criminal investigations. The architecture’s ability to read, process and classify large quantities of images with agility and precision allows the forensic analyst to have a simpler and faster process in analysing this digital evidence. Secondly, Cohen et al. (2023) try to use AI to resolve a classic legal dispute in the field of labour law, “employee versus independent worker”. The study was conducted between Canada and California and led to the implementation of a direct public and open access legal aid tool to provide a minimum of legal assistance to workers and small businesses.

To facilitate a concise understanding of the analytical outcomes related to predictive justice, Table 2 summarizes the main thematic areas discussed in this subsection. Each entry highlights the key findings and implications of AI-based predictive tools in the legal domain, offering a clear overview of the technological applications and the associated challenges.

The interaction between human and machine within judicial processes is expressed through the introduction, in various fields, of the digitalisation of the procedures (PCT electronic civil process; PAT electronic administrative process). The legal decision is increasingly becoming the result of the human–machine connubia. The machine thus guarantees and ensures that very complex issues are solved simply and in a neutral, effective and efficient manner through the mere entry of data within the system; the human component guarantees the correct functioning of the machine and compliance with the fundamental principles of the sorting. The human component also ensures that a correct risk assessment is carried out in the use of the algorithm. In the following, different topics regarding the human–machine combination in justice systems are introduced and discussed.

Opportunities, challenges and risk assessment. The emphasis placed on ICT tools and the complete integration of telematic processes in upcoming reforms prompt authors to offer concise insights aimed at stimulating deeper reflection. Pajno (2022) greatly introduces the debate about recognizing that ongoing technological advancements impact all facets of human endeavours, including those entwined with the legal sphere. Thus, it becomes imperative to discern both the manner and scope through which emerging technologies will shape legal proceedings, ensuring the preservation of their core principles and values, including constitutional tenets. Additionally, it necessitates an examination of how the legal process can, and perhaps should, adapt to the pervasive application of contemporary technologies. The uses of technology in law can be very different: from an app that searches case law to a platform for intelligent document automation, technology seems to change the future of legal practice (Everton da Silva et al., 2019). With authority shifting from humans to algorithms, we may no longer perceive the world as the playground of autonomous individuals striving to make the right choices. Instead, we might view the entire universe as a stream of data and believe that humanity’s cosmic calling is to create an all-encompassing data-processing system and then merge into it (Burt, 2021). Despite the high potential offered by ICT tools and LegalAI, the debate between opportunities and risks remains highly contentious. These challenges primarily revolve around the compatibility of these new tools with the traditional structure and principles of legal guarantees. This compatibility issue stems from the need to uphold fundamental individual rights, ensure a fair trial and potentially redefine certain aspects of criminal justice within the broader context of societal governance (Cesari, 2019).

One of the major risks is the misalignment between human values (e.g. those of the judge) and the implementation of the algorithm Fischer-Abaigar et al. (2024). As underlined by Winter (2022), given the inherently distinct nature of AI and the potential for novel failure modes, it is imperative for AI systems to attain a level of assurance equal to or surpassing that of human judges to uphold crucial judicial values. Incorporating insights from the literature on security and AI alignment in assurance could significantly enhance the discourse surrounding AI in the judiciary, leveraging advancements that have not yet been fully integrated. Moreover, the use of algorithmic justice includes the problem of lack of transparency. As argued by Contini (2020), systems that offer judges relevant input and suggestions on case decisions require a distinct form of accountability, because judges receiving suggestions from the machine are not able to verify its functionality, resulting in a significant lack of transparency. Moreover, judges are influenced by the system and bear responsibility for the decisions made. Under this perspective, the judicial decision-making system fails in guaranteeing the fundamental rights and the core principles of judicial independence and fairness. The main question is about the acceptability or not of the machine suggestion. The answer is not straightforward, but the question underscores the need for thorough evaluation of the consequences of introducing a system (whether digital, statistical or otherwise) that interacts with the judicial decision-making process. Under this perspective, technologies like electronic forms, e-justice platforms, text-to-speech systems or predictive systems can be introduced into the judicial process if, and only if, adequate accountability mechanisms are in place as well as users can ensure effective control of the system. This aspect is in line with the Ethical Charter of the Council of Europe on the use of AI in judicial systems, principles which establish respect for fundamental rights and maintenance of user control.

In this direction, the results of the research by Kaplina et al. (2023) led to the necessity, first and foremost, to distinguish the use of AI within processes, linked to the execution of auxiliary tasks, from its use on decision-making routines. Auxiliary tasks include automatic preparation of forms of certain procedural acts, generalization and systematization of evidence, selection of relevant case law, prediction of judicial prospects, automated preparation of court transcripts using NLP recognition technologies and so on. Such usage presents minimal risks in terms of disproportionate and uncontrolled interference with human rights and freedoms. The legal profession has a responsibility, by virtue of the privileged positions it occupies in each nation, to react to these challenges in an intelligent and humane way and to harness the enormous power of AI to create a more just and equitable society (Burt, 2021).

Recently, Lettieri et al. (2023) proposed a general human-centred AI paradigm in justice crime administration, providing a model of human–machine collaboration aimed at equipping judges with the benefits of AI and computational heuristics while maintaining control and comprehension of the machines’ role. The system is based on two components: an online learning model crafted to aid judges and investigators in assessing the criminal dangerousness of individuals and groups (learning from the user feedback) and a user-friendly tool that utilizes visual metaphors to simplify judges' interaction with data, AI models and other heuristics algorithms to reduce the risk of opacity, aberrations and injustices.

An example in this direction is given by Colombian courtrooms (Perona et al., 2024), where an experimentation of the use of ChatGPT in the decision-making process is conducted, producing a proof of the necessity to reinforce legal practitioners’ ability to critically evaluate and interpret AI-generated outcomes.

Robotics (robot lawyers). Robotics has brought to the fore a formidable dimension of AI by creating an automated artificial agent. The robot regime is significant and demands careful consideration to prevent any dangers stemming from its widespread utilization. The fusion of human and machine capabilities is a recurring theme in the history of AI (Lettieri et al., 2023).

Given the breadth of challenges presented by AI and robotics, law schools, which form the foundation of legal education, are introducing courses on legal education to support efforts for a more informed vision of this progressive technology in today’s setup. The spread out of such technologies favourites the birth of innovative decision support systems defined as robot lawyers. “Robot lawyers” is a term used to describe AI-based technologies applied in the legal profession to perform tasks traditionally carried out by human lawyers. These technologies can range from simple chatbots providing basic legal information to sophisticated algorithms analysing complex legal documents and predicting case outcomes. For example, Bhatt et al. (2022) exemplify how a robot lawyer will increase the speed with which the “wheels of justice” will move. The authors argue also that the proposal to increase the AI role in the legal discipline cannot be ignored because the legal profession is burdened by the enormous weight of voluminous paper documentation. With the increasing intervention of AI, the legal profession will undergo a complete renewal, giving way to speedy justice and more accurate legal advice. The advent of lawyers guided by artificial intelligence is already a reality: robot lawyers have the potential to streamline various aspects of legal practice (legal research, contract analysis, document review and even some aspects of litigation). They can improve efficiency, accuracy and access to legal services, particularly in areas where legal assistance is limited or costly. For example, the first robot lawyer, “DoNotPay”, showed off its lawyering skills in 2015. “DoNotPay” is a chatbot-based legal service that helps users with various legal issues, such as parking ticket appeals, small claims court filings and landlord disputes, using automated legal assistance. Recently, chatbots have been redeployed for asylum seekers seeking financial support from the UK government. In the US and Canada, chatbots aid refugees in completing immigration applications. Beyond the surface of technological advocacy, chatbots possess various abilities that yield a profound social impact (Ng et al., 2022). A variety of chatbots with different roles have been developed in the last years: Kira Systems (https://kirasystems.com/), an AI-powered contract analysis tool that helps lawyers review and analyse contracts; CaseText, an AI-driven legal research platform that uses ML to analyse and summarize case law, statutes and regulations as well as Lex Machina, an AI platform that analyses legal data to provide insights for litigation strategy, including case outcomes, judge behaviour and opposing counsel profiles (https://lexmachina.com/).

It is worth observing that robotics can potentially be pushed in the consultancy institutions: these usually provides legal advice/assistance for consumers, immigration, civil cases and employing early-stage career lawyers. Indeed, in comparison to traditional legal consultancy institutions, the swift expansion of AI robots has the potential to offer substantial benefits and enhance efficiency and gives the possibility to the employees (without a great experience) to improve the research of similar instances as well as to provide the best solutions to customers. For consultancy institutions, understanding the strategies for deploying AI legal bots and the criteria guiding user interaction with such bots is crucial (Ho et al., 2020).

However, the use of robot lawyers also raises ethical, regulatory and societal concerns. Questions arise about the accountability and reliability of AI-driven legal advice, potential biases in algorithms, data privacy and security issues and the impact on employment in the legal profession. Moreover, furthermore, not all countries worldwide have the same level of access to information technology infrastructure or the capacity to invest in it. This aspect could lead to significant disparities in law enforcement among different countries. As introduced by Chourasia et al. (2023), India represents a symptomatic example under this perspective. The Indian legal system still operates in its traditional style, requiring a lot of resources and hard work: the Indian jurisprudential system is very slow and disorganized, a great number of cases are still pending on a small number of judges and lengthy court proceedings (Srivastava, 2023). There is a need for technology development in legal information so that a huge change can be brought to the legal area of underdeveloped countries.

Under the ethical perspective, it was emphasized that establishing an AI-driven legal service system necessitates leveraging machine expertise. In the study by Chen (2022), the hypothesis of law was juxtaposed with that of an information system. The findings indicated that intelligent robotic systems, derived from extensive empirical data, exhibited greater accuracy and predictive capability regarding the content of empirical laws. Additionally, drawing from other studies, such as one examining “which laws apply if a self-driving car kills a pedestrian”, the necessity of establishing an AI legal services system to address issues of legal liability was examined. The author queries whether intelligent robotic systems are also applicable for exploring criminal responsibility. Despite sentencing decisions being influenced by over 200 factors, including laws and sentencing practices, deep learning has been employed for automatic decision-making. This testing of intelligent sentencing algorithms is now regarded as a supplementary tool to existing sentencing practices. The findings suggest that the widespread integration of AI-based computer-aided judgement should be considered for adoption in people’s lives (Xu and Wang, 2021).

Technological and regulation standardization of AI. The conventional normative approach of legal studies towards law tends to be doctrinal and theoretical rather than empirical. However, even the more peripheral tradition of empirical legal research alone may not fully elucidate the intricate dynamics through which algorithms influence fairness, or more tangibly, the pursuit of fairness in accessing algorithmic justice. Simultaneously, legal research has the potential to contribute additional insights to other disciplines. This is because the examination of legal phenomena can be approached from both an internal-law perspective and an external-law perspective. While the external approach to law may utilize similar methodological tools as other social or scientific fields, when employed by a legal scholar, it retains a distinct comprehension of law as a symbolic and normative system. Hence, delving into the algorithm, particularly within the context of law, needs the potential to enhance comprehension and shape future governance and regulation of the technology (Kontiainen et al., 2022). Under this perspective, Mobilio (2020) argues that a law modelled solely on the hard law paradigm ends up being inadequate for regulating a specific entity such as AI. This is because it is confined to a purely “reactive” approach, capable only of responding to the undesirable consequences resulting from the proliferation of AI, thus finding itself unprepared. While not entirely abandoning its regulatory principles, the law should strive to progress further by adopting a “proactive” approach, capable of foreseeing trends, risks and pre-emptively resolving issues that may arise soon. Achieving this objective requires a thorough understanding of the competition among regulatory systems and identifying the most beneficial forms of interaction and integration of their respective regulatory mechanisms, leveraging strengths and addressing weaknesses. This approach would not only refrain from impeding the evolution of AI but also actively guide technological advancement in a dynamic and responsive manner.

Indeed, the absence of a law on the use of AI triggers a crisis of democracy that contributes to the deterioration of democratic resilience due to the uncontrolled use of AI technology (Rahman et al., 2022). Within this panorama, the big concern not only in Europe (Fabri, 2024) but also globally can be found in the ethical conflicts that can arise at different times with the widespread use of AI, which is why the Assessment Group was created in 2015, whose function is to support the European legislative body. This group has carried out a study entitled “Ethical aspects of cyber-physical systems”, which exposes the various ethical conflicts that are linked to technology projected towards 2050 (Zabala and Zuluaga, 2021).

Successively, the EU proposal of the Artificial Intelligence Act (AIA) in April 2021 represents a pivotal initial stride towards comprehensive regulation of AI through a legally binding instrument. The proposed AIA adopts a balanced and risk-focused regulatory framework for AI, categorizing AI into unacceptable risks, high risks and low or minimal risks. Unacceptable risks encompass practices that violate EU values, hence designated as “prohibited AI practices”. These prohibited practices include AI systems employing subliminal techniques; AI practices exploiting vulnerabilities; social scoring systems; “real-time” remote biometric identification systems (Neuwirth, 2023). By borrowing the reasoning of the jurisprudence of merit and applying its findings also outside the trial venue, it should be possible to exclude the new technologies used to make decisions that affect the freedom rights of the individual, if the subject himself is not placed in the condition of maintaining control and self-determination of his person (De Simone, 2023). Finally, Tzimas (2020) introduces an ontological debate about the coexistence of humans and AI that necessitates regulation to minimize the risk of AI becoming hostile towards humans. The potential rise of non-biological intelligence, matching or surpassing human capabilities, poses a distinct challenge to our societies, existing legal frameworks and the rule of law. Hence, it becomes imperative to invoke human rights to establish checks and balances on the development and deployment of AI, ensuring the preservation of human-centric values within our legal systems.

Legal personality of the machine? In a provocative vein, some scholars have speculated that machines might even be capable of recognizing their own subjectivity. Reyes (2021), in his work, introduces the concept of the autonomous corporate personhood spectrum (combinations of rights, capacities and obligations), providing legislators and legal reformers with a framework to delineate the nature and extent of artificial personality for autonomous entities. In detail, this concept provides a graduated scale for understanding the degree of autonomy and legal personality that can be attributed to autonomous corporate entities. This spectrum can assist legislators and legal reformers in defining the boundaries and characteristics of artificial personality for autonomous businesses.

Furthermore, with the advancement of AI systems and their increasing societal influence, there are at least two clear rationales for considering them as legal persons. Firstly, attributing legal personhood addresses accountability gaps arising from their speed, autonomy and opacity by establishing someone to hold responsible in case of errors. Secondly, recognizing personality ensures that there is someone eligible for rewards when outcomes are positive (Chesterman, 2020).

Pagallo (2018) underlines as bizarre events can happens, without a clear regulation. For example, the AI application, named Sophia, received citizenship of Saudi Arabia in 2017. The question is if citizen Sophia is really considered conscious. This underscores the normative considerations by which one can assess whether the attribution of legal personality is reasonable or merely a matter of arbitrary chance and political uncertainty. When it comes to legal entities like corporations, political determinations revolve around concerns of efficiency, financial transparency and accountability. Conversely, regarding human beings, the focus is on aspects such as dignity, conscience and intrinsic value. However, the limit is tight, we cannot prevent, based on these considerations, the eccentric decisions of legislators to “grant citizenship to robots or appoint horses as senators”. Establishing electronic legal personality solely based on automated agents’ capabilities to perform tasks does not seem to be a feasible route for regulating AI. This strategy could only be considered viable if AI and robots were endowed with goods or physical autonomy, enabling the imposition of civil or criminal liability or if punitive measures were tailored to suit AI and robots’ functions and methods (Lima et al., 2021).

Table 3 provides a summary of the principal findings related to the integration of human and machines in justice systems. The table highlights the opportunities, challenges and conceptual models emerging from the literature, particularly those focused on maintaining human oversight in AI-supported decision-making processes.

The concept of a “robot judge” could vary depending on the context. It could refer to an artificial intelligence system designed to assist human judges in legal decisions, using algorithms to analyse cases and suggest decisions. However, it’s important to note that even in these cases, the ultimate responsibility for legal decisions remains in the hands of human judges, with the AI system providing only decision support. The systems described so far (predictive algorithms, chatbots and robot lawyers) still operate within the bounds of decision support systems (also defined as AI assistants), under human supervision. At the same time, it could also refer to a hypothetical futuristic concept in which judges are entirely replaced by AI in the courtroom. This raises a series of ethical, practical and legal questions about how the legal process would be managed and how to ensure transparency, fairness and accountability. Currently, in many judicial systems, the human element is considered irreplaceable due to the complexity of legal decisions and the need for empathy, discretion and understanding of the human context. For a deep analysis about the classification, the reader can refer to Ulenaers (2020). However, considering a possible evolution of the LegalAI systems in this direction, a series of prerogatives must be respected precisely to avoid risks of discrimination and of unjustified evaluations always due to the characteristic opacity and ethical inability of AI. The issues of complexity and ethicality of the algorithm therefore take on relevance. Note that, in the following discussion, the concept of a robot judge is considered as substitute of the human judge, not simple as an AI assistant, a concept already analysed in the previous sections.

Due process and algorithmic fairness. We are moving towards a future where decisions are likely to be made not by humans, but by machines. The increasing presence of AI in decision-making poses a significant challenge to the fundamental principles of democracy and the fairness of the algorithms. To rebuild trust in AI for the betterment of humanity, it is crucial to implement design systems that incorporate principles of judicial oversight as a core component of AI-driven architecture (Katyal, 2022). This aspect is fundamental for considering a future where robot judges really exist. Generally, problems related to robot judges’ partiality and discrimination are in two areas, one in the scientific and technological field and the second in the field of law and organizational governance (Loza de Siles, 2022). Under the technological perspective, a series of explicit and implicit risks has to be considered (Razmetaeva and Razmetaev, 2021). Explicit risks primarily include threats to security and privacy, as well as direct violations of human rights. Explicit risks entail security lackness and resulting from device and cloud data storage hacking (cyberattacks) and installation of malicious software. For this reason, the data stored on court by process participants cannot remain secure; moreover, direct and explicit security risks are amplified in legal systems lacking due diligence in digital adoption. Finally, they escalate in the presence of numerous legislative conflicts, conflicting administrative and judicial practices and a significant corruption element within public law activities. Implicit risks are related both to the technology and to the organizational rules. The fairness of algorithms can be, indeed, influenced by a variety of factors. In their work, for example, Lazar and Stone (2023) introduce an interesting debate on the moral criticism in building artificial decisions. Authors argue that the predictive injustice on disadvantaged groups in AI-based models depends on different factors: the extent of background structural injustice that divides the advantaged from the disadvantaged population; the extent of performance discrepancy between the two groups; the reasons for the model’s comparatively lower performance and the social influence of the endorsing agent. They proposed a theory as guideline for designing such systems, identified as the Prioritarian Performance Principle (PPP), according to which “A model is predictively just only if its performance for systematically disadvantaged groups cannot be improved without a disproportionate decline in its performance for systematically advantaged groups”. Watamura et al. (2023) underlines this concept, introducing the idea of “emphatetic robot-judge”, whose perceptions of empathy enhance people’s trust in robot judges and influence both their evaluation of judgements made by these judges and their acceptance of them in the courtroom. Moreover, Razmetaeva and Razmetaev (2021) explain also that another hidden risk associated with the manipulation of public opinion via widely used and inadequately regulated social media platforms can impact the autonomy of court judgements or robot judge decisions. Hidden risks to justice may arise from the public sector’s reliance on private entities responsible for creating, modifying, adjusting and maintaining technological tools and solutions. For instance, companies providing algorithms for data processing might withhold the source code, citing trade secrets, thereby preventing users, including government organizations and institutions, from thoroughly examining potential vulnerabilities and technical flaws in the algorithm. Additionally, the issue of accountability for the developers and vendors of these digital tools remains.

The work of Morison and Harkens (2019) introduces a great analysis of two tool used in real context: COMPASS (Correctional Offender Management Profiling for Alternative Sanctions) in the US and HART (Harm Assessment Risk Tool) in the UK. These examples illustrate the growing use of algorithmic tools in justice decisions. However, they also highlight differences in analysis and implementation methods, demonstrating that algorithmic risk assessment can yield varied results based on procedural design. HART relies on official record data for relevant defendants, while COMPAS employs a more intricate approach, combining record data with statistical evaluations from offender interviews and self-reports to analyse socioeconomic and psychological circumstances. Additionally, HART is used immediately after arrest to determine custody decisions, whereas COMPAS is utilized throughout the offender processing continuum, influencing pre-trial, sentencing and post-conviction decisions aimed at behavioural change and risk reduction. Thanks to these examples, authors introduced the concept of tools with human-in-the-loop and human-on-the-loop decisions – meaning those where humans are required to select and guide inputs and those where the tool generally works in an automated fashion and where humans are only needed for final execution or intervention, respectively. Please note that COMPAS functions as a human-in-the-loop tool, involving human participation in data collection through self-report questionnaires, with the algorithm constructing a profile through ML techniques. Similarly, HART operates as a human-on-the-loop tool, where demographic information and official record data are combined to produce a risk score using multiple decision tree algorithms. Human supervision occurs only at the final decision-making stage, determining whether to follow the tool’s recommendation on the individual’s risk level (semi-automated decision). Note that the COMPASS analysis is more related to the role of robot judge, while the HART analysis is quite like simple tool predictions. COMPASS tools faced a series of problems. The work of Rubim Borges Fortes (2020) introduces an important precedent in the field: a clear example of algorithmic opacity and discrimination can be seen in the case State v. Loomis, in which the use of the COMPAS algorithm for sentencing was contested. Another point was related to the debate surrounding the bias of racially discriminatory predictions made by COMPAS. In Florida, black defendants were twice more likely to be incorrectly classified as high-risk by the tool, with developers justifying this discrepancy with the higher base rate of offenses among black prisoners, as indicated by institutional data. The most significant risks of dehumanizing individuals arise when they are solely judged by an automated system, without any human intervention, even upon request for reconsideration by a human judge. In such a scenario, individuals are reduced to mere data inputs processed by an intelligent algorithm. Consequently, their dignity and identity as human beings and members of the political community are undermined, violating constitutional principles of intangible protection (Pizzetti, 2019). Barsotti and Koçer (2022) describe scenarios where algorithmic decision-making impacts credit risk applications, such as determining whether to grant a loan, it is crucial to consider sensitive characteristics like gender, ethnicity and disability. These attributes, beyond individuals’ control, should be protected to prevent algorithms from inadvertently perpetuating biases. While there are legal requirements to safeguard against such discrimination, there is a risk that protected attributes may still influence decisions through other features, leading to invisible biases rather than their mitigation. Nevertheless, the debate conducts to reflections for a path to greatly improves the robot judges’ development. Burk (2021) argues that algorithmic parameters are socially toxic, but, just like biologically toxic substances, they could be useful or even curative if applied in limited and judicious circumstances. A crucial aspect of ensuring the effectiveness of the proposed system lies in emphasizing the principle of “role reversibility” within the decision-making process. This principle suggests that those in positions of judgement should be willing to acknowledge their own vulnerability to the outcomes of their decisions. Essentially, it implies that the roles of individuals involved in the process could potentially be swapped; under slightly different circumstances, the decision-maker could find themselves being judged rather than acting as the judge. Therefore, what truly matters is not merely the existence of humanity, but rather the ability of decision-makers to empathize with the perspectives of all involved parties. By fostering this empathetic understanding, decision-makers and the broader moral community can view judgement as a democratic process, wherein all participants are equally valued and considered (Brennan-Marquez and Henderson, 2019). Researchers are not able to understand, for now, how much empathetic can be a robot judge, under this perspective only semi-automated tool are completely accepted. For example, in Arabia, researchers have developed a model named TaSbeeb, which functions as a computer programme simulating the work of a human legal expert and judge. Created by Almuzaini and Azmi (2023), this model accurately classifies cases and incorporates judicial reasoning to assist decision-making. This development reflects a trend towards hybridization between humans and machines, blurring the lines between natural and AI and fostering a bidirectional exchange. Machines now mimic human attributes such as subjectivity, thought processes and even decision-making capabilities. To achieve a true integration between algorithms and decision-making systems, addressing the legal complexities explored by Ruhl and Katz (2015) is essential. Upon examination, it becomes apparent that the judge and the algorithm embody two distinct facets of justice. The algorithm represents a formal, calculable form of justice expressed through predictive algorithms and automated decisions, while the judge embodies a substantive, human form of justice that incorporates emotions and values, transcending the mathematical precision of machines. The aspiration for a computable legal system struggles to reconcile with the inherent uncertainty of human judgement, which is necessary to uphold jurisdictional safeguards. Nevertheless, the efficiency of artificial judges remains an appealing prospect, potentially offering unparalleled impartiality and remedying human fallibility, thereby embodying the ideal of the constitutionally guaranteed natural judge.

Table 4 offers a concise synthesis of the key insights regarding the concept of the robot judge. It outlines the theoretical underpinnings, ethical and procedural concerns and real-world implementations discussed in the literature. This summary enables readers to quickly grasp the complexity and critical issues surrounding the potential automation of judicial roles.

The literature review has identified three key thematic areas concerning the integration of smart technologies into the justice system, which are poised to significantly modify the current panorama: (1) the implementation of predictive justice tools; (2) the synergy between human–machine interaction and (3) the emergence of the robot judge. To effectively summarize the findings of our analysis, it is essential to broaden the discussion to encompass the larger societal evolution catalysed by intelligent technologies. Understanding the full impact of disruptive technologies on the justice system necessitates stepping back and tracing the evolution of society, culminating in the concept of Society 5.0, introduced by the Japanese Government in January 2016. Society 5.0 envisions a future society driven by scientific and technological innovation, striving to create a human-centred, highly intelligent and efficient society. Its goal is to ensure that all individuals can lead lives of quality, comfort and vitality by seamlessly integrating cyberspace and physical space to deliver essential goods and services promptly and at the desired level (Narvaez Rojas et al., 2021; Mouazen et al., 2025). Human society has undergone four distinct transformations: from the hunter-gatherer society (Society 1.0) to the agricultural society (Society 2.0), then to the industrial society (Society 3.0) and later to the information society (Society 4.0). Currently, it is in transition towards the super-smart and human-centric Society 5.0. Concurrently, the Industrial Revolution, originating from Society 3.0, has evolved through four phases driven by advanced technologies. Presently, humanity is on the brink of another industrial revolution, known as Industry 5.0 (Huang et al., 2022).

It is noteworthy that the last two phases of the industrial revolution, marked by significant technological advancements, have influenced various sectors of society in a similar direction. This has given rise to paradigms such as Tourism 4.0 (Stankov and Gretzel, 2020), Agriculture 4.0 (Zhai et al., 2020) and Healthcare 4.0 (Tortorella et al., 2020), which have also impacted the realm of justice. This impact is evident in the introduction of e-justice or the Justice 4.0 paradigm (Sari et al., 2021), as represented in Figure 8.

Figure 8 illustrates how traditional justice, largely reliant on paper documents, remained unchanged until the widespread digitalization of society (Society 4.0) and the introduction of automation systems in the industrial sector (Industry 3.0). Consequently, the initial impetus towards the automation of legal processes was provided by the Justice 2.0 paradigm, less than 20 years ago (Roznai et al., 2015).

In this case, the most significant changes in the judicial sector have involved the digitization of processes and archives and the emergence of telematic tools to manage courts and to speed up slow procedures prone to material errors, making judicial activities more accessible and transparent. The next step in the digital transition of justice has primarily occurred in the last five years, thanks to the proliferation of intelligent technologies, opening new perspectives for efficiency within the field (Justice 4.0). Finally, nowadays, the first signs of technological evolution that places again humans at the centre (a new “anthropocentric” approach in Justice 5.0) of the digital systems are relevant, as already happened in the industrial sector in past years. For these reasons, the synergistic vision of justice within a society based on industrialization and digital evolution is imperative.

In detail, the analysis conducted in this work falls between the Justice 4.0 and Justice 5.0 paradigms, the smarter phases of the evolution. The introduction of 4.0-AI techniques within legal processes, for example, has led to several innovations within predictive justice or smart decision support systems (such as chatbots or robot layers), as previously described. However, the time is ripe for broader reflections within the judicial domain. As anticipated in Section 3.2.3, there is now a question about the possibility of replacing human decision-making with artificial decision-making, thanks to the use of robotic judges and about the methods of constructing human-centred systems (5.0 perspective) that can always prioritize the importance of humans. It is worth noting that the process of change within the justice system is inevitable; it has already been triggered and cannot be altered, as it is part of a radical societal change process. However, it must be appropriately guided. This process is driven on one hand by the 5.0th industrial evolution, which has matured the technologies related to AI, and on the other hand, it fits into a broader debate on how the Society of the future (Society 5.0) should not only be in terms of smartness but also in terms of fairness, attention to the individual and their needs. From this dual perspective, the debate on LegalAI and Justice 5.0 emerges as multidisciplinary, as previously mentioned, and unfolds into a series of future research developments concerning three fundamental spheres: technological, legislative and ethical, as deeply described in the following.

As aforementioned, there are many problems that the researchers will still have to deal with. The phenomenon of AI in the legal field is still in its infancy, dealing with a potential that is not only infinite but also with challenges that must be addressed from an interdisciplinary perspective (technological, legislative and ethical). The interdisciplinary comparison on the topic of AI – its applications in everyday life and its interrelations with legal, ethical and sociological profiles –has been open for a long time (Perlingieri, 2021).

Technological perspective. Regarding the prospects and risks of algorithmic decisions, the debate has been open for years, as it is dominated by discordant opinions, there are those who believe that such decisions can achieve a greater degree of precision and therefore be more precise than human ones (Luciani, 2019). On the other hand, there are those who argue that the algorithm can be discriminatory, the so-called disparate treatment, since it could base its decisions on characteristics specific to the individual, such as race, ethnicity or gender. To determine what the social impact of algorithmic decisions is, it is necessary to evaluate the phenomenon not only by parameterizing it to human decisions but also by trying to analyse it on a broad basis (Lepore, 2021). On the one hand, it is essential to analyse the accuracy and fairness of the decision, but on the other hand; however, one cannot help but also highlight criteria such as its pervasiveness and cost-effectiveness. From the first point of view, it is necessary to allow the recipients of the decision to understand the decision-making process through the backward reconstruction of the same (so-called reverse engineering) to verify its compatibility with the constitutional guarantees of due process (Perlingieri, 2006). Moreover, the idea of ​​making predictive justice more and more effective and the presence of robots supporting justice systems require further studies to make AI algorithms increasingly precise, reliable, transparent and versatile. The judicial system represents a niche from an informatics standpoint because it is characterized by a particular language, laws that vary from state to state and bureaucratically detailed procedures. Therefore, customization of algorithms is strongly necessary and computationally complex. Finally, it is necessary to increasingly experiment with intelligent applications in practice to assess their shortcomings in a process of continuous improvement. On the other hand, application potentials which concern all those procedural activities which do not require interpretative moments but which can be resolved in mere documentary requirements are welcomed. This is the real contribution provided by AI to the process which, from the perspective of optimization, allows formal obligations and procedures to be carried out with an efficiency that was unthinkable until a few years ago.

Legislative perspective. From a regulatory perspective, it appears essential to include ex ante the interpretative criteria that will be used by the interpreter and, in the case of AI algorithms, by the machine. Just look at the EU report which warns of the importance of timely EU action to set clear standards based on EU values and to prevent these from being defined elsewhere. Technology, ethics and law are essential components in this debate. It is not easy to combine each of these dimensions of AI, but only a serious and continuous commitment to mutual understanding between them would allow us to respond to these challenges (Carrillo, 2020). It is necessary to encourage the existing debate to expedite regulatory procedures. AI tools have enormous potential if used appropriately but can also lead to catastrophic consequences if mishandled. It is important for governmental bodies to delineate the proper and improper use of such technologies through ad hoc laws, clearly defining the civil and criminal liability of the machine and those who design or use it.

Ethical perspective. Placing emphasis on the ethical question is a duty which the interpreter cannot and must not escape in any way. Kontiainen et al. (2022) analysed the topic from a dual perspective: first, to promote critical research on algorithmic fairness in the context of the administration of justice and, second, to suggest a methodological approach for the interdisciplinary study of the same. The legal profession must recognize that it has an ethical obligation to care for the tools of its trade as much as the conduct of its members. It is necessary to consider predictive coding as a powerful and potentially beneficial tool. It must do so, however, with a critical eye and a firm commitment to the use of new technologies to serve the objectives and values of the legal profession and the judicial system (Remus, 2014). The importance of ethical principles in AI is generally recognized in the institutional framework, in the scientific community and in society in general. AI has led to a broad and intense debate on its ethical aspects. Ethics plays an essential role, but ethical concepts and principles vary over time, space and between the different subjects involved. For this reason, the ethical debate must be open and inclusive, never exclusive or selective. Furthermore, the role of ethics is different from that of law. The law is mandatory and has legal and jurisdictional mechanisms to ensure its application (Carrillo, 2020). Therefore, a question arises spontaneously. Is it precisely in the emotional dimension that the quid pluris of human judgement compared to automated judgement can be identified? Upon closer inspection, this is a dimension that is structurally precluded from AI and, although there is no shortage of proposals to recognize a certain empathic capacity in machines, from this point of view the gap appears unbridgeable (Arduini, 2021). One of the themes is undoubtedly that of the automation of justice which, as we have seen, is a process that certainly involves critical issues – such as the cancellation of the interpretative moment – but also infinite potential (Quattrocolo, 2020). It can be hypothesized that the machine uses, almost by default, for its very architecture, the rule of stare decisis, but the machine would be able to deviate – in the presence of suitable conditions, also deriving from the evolution of the socioeconomic context, etc., from previous guidelines already widely consolidated over time? In other words, one wonders whether the machine could be a precursor of a new jurisprudential orientation that would derive from the change in the social context. By answering affirmatively, in a purely hypothetical key, further critical expressions would lead to the question of whether the machine should be considered as a single entity or whether each machine, in each district, district, court and state, is totally autonomous compared to the others. If this were the case, it would take little for a real jurisprudential conflict to arise between two artificial intelligences, with results that are, a priori, difficult to predict.

In conclusion, it cannot be excluded that the process of control of law by algorithms has already led to a replacement of law with algorithms. If law were replaced by algorithmic management and control, for the integration of ethical rules into social management there would be nothing left but their programming (Koos, 2022). For now, the most sensible path to follow is that highlighted by the Justice 5.0 paradigm; that try to overcome a series of ethical and technological concerns through the development of systems based on human–machine collaboration (Lettieri et al., 2023). On one hand, these systems can ensure algorithmic efficiency, and on the other hand, they are supervised by the humans, which are irreplaceable and capable of ensuring the empathetic dimension and preventing judgement distortions due to any form of discrimination. Finally, on the horizon, new professions and, above all, new technologies appear – as analysed in the paper – which will be able to redesign the professional system, organizational relationships, and production systems overall. Far from outlining futuristic scenarios that are not very concrete, we can try to hypothesize the composition of the professional “justice” system, starting from the current professions and defining them according to the competency model. Likewise, some professions can be identified that already enrich the professional panorama of justice and are contributing to supporting the digital transition. In other words, we are faced with an epochal mutation given by the evolution of old professions, the rise of new professions and the definition of new organizational relationships and service models strongly governed by digital tech.

The integration of AI into justice systems is not merely a matter of technological advancement but represents a profound transformation of legal reasoning, professional roles and institutional frameworks. The findings of this study yield a set of implications that are both theoretical – contributing to academic debate and model development – and practical – informing design, regulation and implementation strategies.

From a theoretical standpoint, this review contributes to several domains:

1. Multidisciplinary conceptualization: By mapping the literature across the three macro-areas – predictive justice, human–machine combination and robot judge – this study introduces a structured approach to understanding the complex ecosystem where technology intersects with law, ethics and governance. Each macro-area serves as a theoretical lens for interpreting emerging phenomena and framing future research questions.

2. Framing LegalAI within Justice 5.0: The work refines the concept of “LegalAI” within the broader societal shift towards Justice 5.0, characterized by a human-centric, intelligent and inclusive justice model. This conceptual alignment with Society 5.0 and Industry 5.0 provides a systems-thinking approach to evaluating innovation in public institutions.

3. Contribution to algorithmic jurisprudence: By analysing AI-driven legal reasoning tools, especially those that predict case outcomes or support judges, this study contributes to the emerging field of algorithmic jurisprudence. It raises fundamental questions about the delegation of discretion, the replicability of judicial logic and the construction of fairness in automated systems.

4. Foundation for typological and normative frameworks: The classification of AI applications and their risks offers a base for the development of new typologies and ethical-normative frameworks in computational law, enabling scholars to differentiate between assistive, semi-automated and fully autonomous legal technologies.

Beyond academic theory, the findings have direct relevance to the stakeholders involved in the deployment and governance of AI in justice systems.

1. For policymakers and legislators: This study highlights the need for anticipatory governance and a proactive legislative stance. It underscores the urgency of establishing clear regulatory frameworks that define the acceptable boundaries of AI deployment in justice while ensuring protection of fundamental rights. The discussion on transparency, accountability and human oversight aligns with ongoing debates surrounding the EU AI Act and similar national regulations. Furthermore, the study reveals that existing and emerging AI use cases – particularly in criminal justice, risk assessment and sentencing – demand a structured, risk-based regulatory approach. Insights from the reviewed literature can support the classification of AI tools by risk category, in line with the EU’s proposed regulatory architecture. Beyond compliance, regulatory clarity is essential to facilitate public trust, ensure cross-border legal interoperability and stimulate economically sustainable innovation in the legal-tech sector. By reinforcing democratic values such as due process, equality before the law and judicial independence, regulation becomes not merely a constraint but a necessary infrastructure for innovation with legitimacy. This position is regulation as a bridge between technological advancement and social justice, reflecting the Justice 5.0 vision where human dignity remains at the centre of digital transformation.

2. For legal practitioners and judicial authorities: The work provides critical insight into how AI tools can enhance efficiency and access to justice, particularly in repetitive and heavy document tasks. However, it also cautions against over-reliance on opaque systems, urging the integration of human-machine collaboration frameworks that prioritize explainability and judicial autonomy.

3. For AI developers and legal technologists: The analysis draws attention to the importance of context-aware system design, where tools must be tailored to legal language, procedural rules and local institutional cultures. Developers are encouraged to adopt participatory design approaches that involve judges, lawyers and clerks from the outset to ensure usability, reliability and alignment with legal standards.

4. For educational institutions and researchers: The study suggests curricular innovations in legal education, promoting interdisciplinary training that equips future legal professionals with AI literacy while instilling ethical awareness. It also identifies under-researched areas – particularly empirical validations and comparative studies across legal systems – as promising paths for future investigation.

One of the core contributions of this review is its ability to bridge the gap between high-level theoretical discourse and ground-level operational challenges. By organizing the literature through a thematic map and grounding it in the real-world implications of LegalAI systems, the study serves as both a conceptual guide and a policy-shaping tool.

It invites legal scholars, technologists, ethicists and institutional actors to move beyond abstract debates and towards collaborative design and governance of AI in justice – thus ensuring that innovation supports, rather than undermines, the foundational values of law and society.

The investigation carried out highlighted: potential, critical issues, gaps and outlets of AI innovative systems in the legal field through a PRISMA analysis of the existing literature and a classification of the major contributions. The work evidence how, under the 5th Industrial Revolution perspective, the AI innovation is pervasive within the entire society and how its presence is destined to grow also within the justice system. This analysis introduced a series of open points to be further investigated in the future.

First, the only certainty to date is that intelligent systems require immediate regulation given their absolute interference in the daily lives of all individuals. The challenge of the jurist, regardless of his specialist sector, is to be able to exploit the machine without being dominated by it, making it useful for the intended objective, thus counteracting the possible danger of technocracy. In fact, it is important to learn how to predict, control and manage the inevitable change of a society that tends towards conservatism for fear of the unknown (Johnson, 1999).

Secondly, the jurist was warned in the use of algorithmic decisions by computer science experts because he was investigating the foundations of the algorithm, without prejudice however – having the opportunity to observe – to its ability to store and process a number in a neutral manner, an almost infinite amount of data, it is possible to understand how dependent this is on the data entered into it by the human operator (Cath et al., 2018). Algorithms allow a rapid and at the same time accurate evaluation of data and input variables, as long as they are standardized and inserted as input, more than what a human being could do i.e. they allow the data to be indexed and criteria or general rules to be abstracted by the repetitiveness of cases (Vespignani, 2019). In general, the use of AI algorithms is desirable in repetitive, simple – mostly documentary – and modest cases, in which, often, the mere allegation of the fact is sufficient and opposition from the person served is rare (Carleo, 2017).

The perspective that is probably most suitable for fully realizing the fusion between AI and the process is therefore the one that places the emphasis on the figure of the judge, formally supported by the machine in a serving and ritualized function, avoiding all the lucubration’s that would lead to the hypothesis of the birth of the figure of the robot judge who would end up undermining most of the procedural guarantees consolidated over time (Ruffolo, 2021).

What emerges from the analysis carried out is the need to place the human being at the centre of the entire design and functioning cycle of AI innovative systems, that is, the need for an “anthropocentric” approach. The guarantee against the risks envisaged is human control, based on what was also stated by the European Parliament, it is necessary to address the issue from the perspective of “supervised autonomy”. This perspective would also lead to responding to the safety needs that are becoming increasingly concrete among users, in fact, by doing so, it would be possible to maintain the equitable moment in which the operator intervenes to guarantee the fundamental principles (Civitarese Matteucci, 2019).

Despite its comprehensive approach, this study presents some limitations, which should be acknowledged to contextualize its scope and encourage future research.

The analysis relies primarily on theoretical and bibliometric methods, without incorporating empirical validation through interviews, case studies or field-based impact assessments. This methodological limitation reduces the ability to assess how AI tools function within actual judicial environments and legal workflows.

Moreover, the literature review draws exclusively from the Scopus database. This source base may have excluded other contributions indexed in grey literature. As a result, the evidentiary foundation of the review may be incomplete. A language limitation further constrains the scope, as only publications in English, Italian and Spanish were considered – potentially omitting other perspectives from jurisdictions where key legal scholarship is published in French, German, Chinese or other widely used legal languages.

While the review aspires to a multidisciplinary approach, the predominance of sources from technical and computer science domains may have marginalized in-depth legal, philosophical or ethical analyses. This disciplinary imbalance could affect the depth of normative insights into the adoption and implications of AI in justice systems.

Finally, although the paper references a range of international examples, it lacks a structured comparative framework across legal systems. A more systematic comparative analysis could have enriched the findings by illuminating jurisdictional divergences in legal cultures, institutional readiness and normative responses to AI adoption.

These limitations do not undermine the contributions of this study, but they do provide a framework for refining and expanding future research on the intersection between AI innovation and the evolution of justice systems.