Transparency in AI-Assisted Criminal Justice Administration
In this series I will explore the potential limitations of AI’s integration with criminal justice. I will analyse the commentary of experts in both AI and law, as well as offering my own insights arising from the emerging research.
AI’s Transparency Problem
Transparency is an integral part of criminal justice systems internationally. Public criminal trials have been established in England and Wales for centuries and are well protected by the common law[1] and the ECHR[2]. Further, publicness, explicability and open reasoning are central to the inherently democratic nature of the law[3], which allows for “contestation, resistance, and change”[4]. Essentially, understanding the reasoning and decision-making processes used by our law makers and arbiters is critical to ensuring the proper execution of justice.
Therein lies one of the primary criticisms of ‘black box’ AI tools (such as ChatGPT, Gemini, and Copilot) and their use within the criminal process. The term ‘black box’ in this context is used to describe systems whose internal workings are opaque to the user. Problem in; answer out. It is obvious to see how this differs from the current system: detailed judgements outline both the ratio decidendi and obiter dicta, experienced HMCTS staff are accountable for their decisions in listing, and lawyers are accountable to their professional regulators for working in their client’s best interests.
In my most recent article[5], I suggested that AI could be used to help tackle the worsening backlog of criminal cases in the courts of England and Wales. One way in which I suggested this may be possible is by reducing the significant number of ‘ineffective’ trials (a trial that must be rescheduled) because of overlisting, which was the primary factor rendering trials ineffective in Q1 of 2026[6]. In the interest of continuity, I will use this and other administrative examples for assessing the transparency related issues facing AI’s use in the criminal process.
Many readers will note that the judicial function of AI in assisting, say, sentencing, as has been adopted in China’s ‘smart court’ structure[7], poses more contentious issues. Whilst I fully agree, this falls outside the more immediate question facing the courts of England and Wales: how might AI relieve administrative burdens without undermining transparency? Rest assured, the use of AI in a judicial capacity and specifically it’s adoption in the Chinese courts will be discussed in detail in the near future.
The use of an opaque ‘black box’ AI system to relieve the administrative burdens on the criminal justice system risks the following:
- An inability to identify and resolve errors; for example, if an AI listing tool repeatedly gives inaccurate predictions regarding the probability of trials cracking, what information did it use to make its predictions? Is it using a formula? If so, what about the formula requires adjustment? Without this information there is no way of improving outcomes and therefore reducing issues of overlisting.
- Accountability – if there is no knowledge of the reasoning behind an AI generated output, how can a human verify its decision? An experienced professional may be able to verify the result by using their existing process, but in that event the workload is effectively doubled and efficiency reduced. Further, without human verification, who is held responsible in the event of a serious failure, such as miscalculating custody time limits or omitting exculpatory evidence?
- Large-scale and compounding mistakes – whereas a single human error may have an impact of a handful of cases, a single flaw in an automated system can reproduce the same mistake many hundreds or thousands of times before anyone can see what is happening, such as in Johnson and ors[8]. The risks are greater still where algorithms create feedback loops (the erroneous outputs they generate directly influence their future decisions). Such concerns were raised over the Home Office’s use of a streaming tool for visa applications. It was suggested that nationality-based risk classification may have contributed to discriminatory patterns[9].
The Transparent Solution (My View)
Not all AI models are ‘black box’. There are also a number of ‘glass box’ systems where the logic from input to output can be observed. Whilst some experts argue that such models may never be capable of the conceptual inference and contextual understanding required for judicial functions, administrative tasks, such as listing, are heavily informed by probabilistic and historic data and therefore are more suitable for interpretable AI solutions.
Given the numerous ‘black box’ issues we have discussed, it seems obvious that a ‘glass box’ solution is more appropriate. This allows for human verification, the identification and resolution of algorithmic errors, and therefore the avoidance of the repeated and compounding mistakes. As for accountability, if each AI-led decision is to be verified by an experienced professional based on the model’s transparent reasoning, that person should sign off the decision and be accountable for accepting and acting upon the output.
Such a system also then leaves an auditable trail of events. It can be established: who or what triggered a function (input), what output was produced, whether it was verified by a human and who that person was, and what subsequent action was taken. It will also be necessary, for the avoidance of doubt or dispute, for all documents, calculations, and processes created or assisted by AI to be digitally stamped as such, accompanied by the unique signature of the approver.
In the event that the approver believes an error has been made, the output must be easily contestable, with a simple mechanism to flag and amend the output. With a ‘glass box’ solution, the reasoning can then be examined and the flaw identified. Where the issue is with the underlying data, this can be resolved internally. Where the issue is with the model itself, this can be raised with the system developer for adjustment.
Sir Brian Leveson has argued that “It is not sufficient for AI merely to accelerate or automate processes that are already inefficient”[10]. Whilst I am very much in agreement that the criminal process can be redesigned for enormous gains in end-to-end efficiency, there is no reason AI cannot also be used to navigate existing inefficiencies more effectively while the longer-term work of systemic reform takes place.
[1] Scott v Scott [1913] AC 417 (HL) 477
[2] European Convention on Human Rights (1950) art 6(1)
[3] John Morison and Tomás McInerney, ‘When Should a Computer Decide? Judicial Decision-Making in the Age of Automation, Algorithms and Generative Artificial Intelligence’ in Sophie Turenne and Mohammad Moussa (eds), Research Handbook on Judging and the Judiciary (Edward Elgar 2025)
[4] Adam Harkens and John Morison, ‘Generative AI and Courts in the UK: A Candle in a Hurricane’ in Monika Zalnieriute and Agne Limante (eds), The Cambridge Handbook of AI and Technologies in Courts (Cambridge University Press 2026), 476
[5] Will Ferguson, ‘The “Numeric Crisis” of the Courts [1]: Can AI Help Tackle the Criminal Court Backlog in England and Wales?’ (Artificial Justice, 10 August 2026) https://artificialjustice.co.uk/2026/08/10/the-numeric-crisis-of-the-courts-1/ accessed 12 August 2026.
[6] Ministry of Justice, Criminal Court Statistics Quarterly: January to March 2026 (25 June 2026) https://www.gov.uk/government/statistics/criminal-court-statistics-quarterly-january-to-march-2026/criminal-court-statistics-quarterly-january-to-march-2026
[7] Supreme People’s Court of the People’s Republic of China, ‘Chinese Courts Must Implement AI System by 2025’ (12 December 2022) https://english.court.gov.cn/2022-12/12/c_1053712.htm accessed 12 August 2026.
[8] R (Johnson and others) v Secretary of State for Work and Pensions [2019] EWHC 23 (Admin) (11 January 2019)
[9] ‘Home Office Suspends Use of Digital Streaming Tool for Visa Applications after Legal Action by JCWI and Foxglove’ (9 August 2020), [20]–[29], https://tinyurl.com/54cwujpt, accessed 12 August 2026
[10] Sir Brian Leveson, Independent Review of the Criminal Courts: Part II, vol 1 (Ministry of Justice 2026) 68 [23], 189 [152].
Leave a Reply