AI in the criminal justice system: From ‘AI ethics’ to ‘hermeneutic risk’

Artificial intelligence (AI) technologies have become omnipresent, rapidly proliferating across developmental boundaries to embed themselves in the fabric of everyday working life. Claiming transformative potentials and professing optimal efficiencies, these technologies now permeate all domains of research and practice, not least the Criminal Justice System (CJS). Consequently, CJS professionals are now expected to assimilate a diverse array of AI applications that claim to offer support across a multitude of core competencies, including expert decision-making, ¹ data analysis, ² pattern recognition ³ and workflow optimisation. ⁴ In addition to contending with the abrupt encroachment of such applications, CJS professionals are also expected to absorb an attendant deluge of analytical research, reports, and regulatory guidance, all seeking to compass the nature, dimensions, and risk-laden implications of these opaque new technologies.

Thus, research journals abound with articles that seek to introduce novel concepts and frameworks, from algorithmic injustice to AI ethics, from explainable AI to algorithmic opacity. ⁵ In an effort to address this rapidly expanding body of applications and literature, and to offer some clarity to CJS professionals – particularly those working in the forensic sciences – this editorial attempts to distil a cluster of useful concepts and principles. In so doing (and at risk of adding to the already burgeoning argot of AI concepts) the authors propose the introduction of an umbrella term that is attuned to the needs and concerns of CJS professionals: hermeneutic risk. ‘Hermeneutic risk’ represents the potential for epistemic and evaluative disjunction, particularly when utilising opaque machine learning systems: systems, whose modes of reasoning and decision-making may deviate from the inferential logic and rationality that underpins both legal and forensic analysis, confounding explainability. ⁶ Explainability is not just a technical hurdle. As recent analyses have shown, technical explanations often fail to make sense to non-specialists. And without clear, accessible explanations, AI risks becoming another unaccountable layer in an already opaque system. ⁷ Transparency is another pressing concern. Many AI systems used by public authorities are proprietary, imposing limits on external scrutiny. Without access to training data, design choices, or decision logic, neither affected individuals nor legal professionals can meaningfully contest the system outputs. Recent policy reviews have therefore emphasised the importance of explainability and transparency, not as a mere technical standard but as a democratic imperative. ⁸ Indeed, citizens subject to automated decisions must be able to understand and challenge them. By way of example, Pavlidis ⁷ points out that although the EU's Artificial Intelligence Act ⁹ requires systems to be explainable, there is little guidance on what this should look like in real terms.

The need to focus on hermeneutic risk is thus both timely and necessary, given that the UK forensic science sector is already regarded as being subject to a number of epistemic pressures. Successive Parliamentary inquiries have highlighted serious structural issues with forensic science provision in the UK. ¹⁰ Similarly, the latest report from the Westminster Commission on Forensic Science ¹¹ highlights the critical state of provision in England and Wales. The report also draws attention to the urgent need for sectoral reform, adequate funding, and scientific independence, to prevent miscarriages of justice whilst restoring public trust in the criminal justice system. The report further notes that the turn to digital evidence has exacerbated these structural problems, due to a significant expansion in the volume and complexity of digital submissions, which has generated backlogs, strained limited resources, and rendered the triage and storage of data more onerous, with attendant risks to evidential integrity. Whilst the report does not specifically propose the use of AI as a solution to these issues, the scale of challenges identified in handling digital evidence does lend itself to suggestions that AI could play a central role in tackling ongoing problems through improving efficiency.

However, the temptation to negate inefficiencies by embracing AI technologies may usher in a paradigmatic distortion to the rational analytic core of forensic and legal analysis. Indeed, the nature of this potential distortion has been highlighted in the latest Discussion Paper from the Law Commission for England and Wales, which focuses on AI and the Law. ¹² At their heart, AI systems comprise a cluster of complex mathematical functions which can perform vast statistical analyses in ways that exceed basic human cognition. However, to the extent that these systems ‘reason’, they do so in ways that diverge markedly from human inference, particularly due to the inability of an AI system to situate tasks within a conceptual model of either the world, or of the sphere of activity in which it is deployed. Thus, even the designers of AI systems will experience difficulties when attempting to discern how the system derived a particular output from a given input. And the attendant danger is that those utilising these hyper-efficient systems may interpret machine logic in ways that reflect expert human reasoning. This brings us back to the concept of ‘hermeneutic risk’: an AI system, operating within the CJS, performs its functions with no comprehension of the high-stakes nature of its predictions and decisions. Nor can it situate its tasks in terms of the role of forensic analysis, or the justice system's overarching principles, aims, and objectives. For these reasons, it has been argued that technical accuracy is only part of the picture, for even a perfectly calibrated AI system may not be ethically fit for purpose. This raises further pressing questions. Should machines ever influence decisions about proof and punishment? If so, under what conditions, and with what safeguards? These are not computer engineering questions: they are fundamental ethical and legal conundrums, and they demand careful thought. ¹³

Indeed, the introduction of AI systems to a justice system that is already experiencing a number of economic and organisational tensions may serve to compound problems: particularly when these systems may have been developed without adequate research into robustness and foundational validity, ¹⁴ without suitable programs of education and training for practitioners, and hampered by a dearth of funding within the sector to address practical issues of quality management and raw computing power. ¹¹ In response, the authors adopt the core recommendation of the Westminster Report: namely, that the underlying structural issues which afflict the forensic sector must be urgently addressed, prior to the addition of another layer of opaque technology capable of generating vast quantities of additional data. In short, there are no quick technical fixes for the justice system and the introduction of systems freighted with hermeneutic risk may distort the objectives of the CJS in quite unexpected ways.

Thus, it is understandable that the EU's Artificial Intelligence Act classifies criminal justice applications as “high risk,” requiring strict oversight and transparency measures. ¹⁵ In the UK, similar protections are needed, particularly around explainability, auditability, and human control.^8,16 Given the above, what practical steps should CJS professionals take to negotiate the implementation of AI systems within the forensic sector or to challenge the deployment of AI by criminal justice agencies? The UK Forensic Science Regulator (FSR) has remained largely silent on the issue: having acknowledged that forensic work requires AI regulation, ¹⁷ no specific guidance has thus far been issued. Instead, AI applications fall under the existing requirements for method validation, error-rate reporting, quality management, and practitioner competence as set out in the FSR's Statutory Code of Practice. ¹⁸ These requirements are shaped by a broader ‘principles-based’ approach to government regulation, which emphasises transparency, fairness, accountability, and safety. Thus, CJS professionals using AI tools must continue to meet established forensic standards while aligning with the overarching principles, in anticipation of the formal oversight mechanisms that are likely to follow. Meanwhile, criminal justice agencies, in congruence with all public sector bodies, are required to comply with transparency principles by uploading information on their use of algorithmic tools to the Government Digital Service's (GDS) Algorithmic Transparency Recording Standard Hub. ¹⁹ At time of writing these requirements apply only to central government departments. However, the GDS plans to extend the requirements to those arms length bodies which provide frontline services. This would include private sector forensic science providers who contract with public bodies.

As regards the requirements of fairness, accountability, and safety, these land in differing ways depending on the nature of AI application. The authors submit that, whilst workflow optimisation may score low in terms of hermeneutic risk, AI-driven data analysis and expert decision-making carry significant hermeneutic risks. Meanwhile, AI pattern-matching technologies offer enormous potentials for development provided that these can comply with requirements for robust validation and error-rate reporting. Given that a diverse array of such AI systems are currently being developed for use across the CJS^20,21 the authors conclude with a brief summary of both prior and proposed AI implementations across the UK justice system, in order to scan the horizon for possible risks.

Legal decision-making

AI tools are increasingly used to manage legal workloads, primarily through the use of natural language processing (NLP). NLP systems can scan and summarise vast quantities of text, assisting legal professionals in identifying relevant precedents, reviewing case materials, and drafting documents.^22,23 Whilst NLP systems can accelerate the research process, accuracy remains a concern. A recent empirical study ²⁴ of leading legal AI tools (Lexis+ AI and Westlaw AI-Assisted Research) found that outputs were inaccurate or misleading in up to one-third of cases. These so-called hallucinations – a term used to describe the tendency for AI systems to confidently assert false information – are not mere technical errors. They are a fundamental side-effect of the ways in which large language models (LLMs) operate. These systems do not compile text in a systematic and rational manner. Instead, they perform iterative statistical calculations to choose the most suitable word and the word following that. This leads them to produce text – including legal precedents, statutes and citations – which appear plausible but have no basis in reality. Hallucinations can significantly undermine the integrity of legal submissions or CJS processes. Although Retrieval-Augmented Generation methods may reduce this risk, they do not eliminate it. ²⁴ For this reason, judicial guidance in England and Wales has stressed that legal professionals must remain ultimately responsible for verifying all AI-generated content. ²⁵

The stakes are further raised when predictive AI systems inform judicial decisions. The Harm Assessment Risk Tool (HART), developed by Durham Constabulary and Cambridge University, was one of the UK's earliest attempts to predict recidivism using machine learning. However, the nature of the input variables – including age, gender, offence history, and postcode – whilst statistically useful, raised concerns about replicating entrenched social inequalities.^8,26,27 Although HART was decommissioned in 2020, it provides a salient example of the ongoing requirement that CJS agencies deploy AI in ways that do not exacerbate bias.^16,28

Correctional practices

AI is also being introduced into the prison environment. In the UK, the Ministry of Justice has recently set out plans to use AI across prisons and the Probation Service to calculate the risk of recidivism. Meanwhile, a separate AI tool has been suggested for use in scanning the contents of mobile phones, to rapidly flag messages containing potential intelligence, including coded language. This application is intended to help prison staff to identify threats of violence and the smuggling of contraband, enhancing security, and enabling earlier intervention. ²⁹ Concurrently, within the UK Probation Service, a new pilot will require offenders to remotely check-in via mobile devices, recording brief videos which will be scanned by AI in order to confirm identity. The system will be deployed alongside existing licence conditions such as GPS-tags and in-person appointments. ³⁰

In light of these high-stakes innovations, what practical steps can practitioners take to navigate the introduction of AI into the forensic sector, in the face of a lack of clear guidance. To reiterate, CJS professionals are simply expected to demonstrate compliance with existing requirements for methods validation, error-rating reporting, and quality reporting. Such obligations form part of the broader principles that prioritise transparency, fairness, and accountability. CJS professionals employing AI tools must therefore continue to uphold these established forensic standards, in light of the overarching principles and in anticipation that specific oversight that will follow.

In this context, we are recommending three critical actions for CJS professionals:

Understand before using: If AI is applied in forensic or criminal workflows, CJS professionals must ensure that they understand how the system or tool operates, including the underlying assumptions and limitations. Where such understanding is not obtainable, the AI system should not be employed.

Until sector-specific guidelines and regulations, and clear oversight mechanisms, are in place, the responsibility for navigating AI inevitably falls on CJS professionals, who are already operating under intense pressure. ¹¹ It is acknowledged that expecting professionals to absorb additional technical knowledge, and comply with further regulatory demands, despite an absence of structural support, risks compounding an already overstretched system. Education and awareness may be essential, but these cannot be left to the individual alone. What is required is sustained investment, coordinated training, and institutional commitment to ensure that AI enhances, rather than undermines, professional practice, augmenting rather than compromising both forensic integrity and the credibility of the criminal justice system.

Despite these system-wide challenges, CJS professionals are not without agency. By engaging critically with emerging AI tools, maintaining established forensic standards, and pushing for transparency and accountability, we can together shape how these technologies are integrated. And, with adequate investment bolstered by institutional and government support, AI may achieve its much-vaunted potential: to ease workloads and strengthen practices, and to ultimately support, rather than erode, our criminal justice system.