There is a document that arrives at the bench before sentence is passed on most defendants who have pleaded guilty or been convicted in a magistrates court. It is called the pre-sentence report. For decades it was the work product of a probation officer who had interviewed the defendant, reviewed the circumstances of the offence and formed a professional view about risk, remorse and the most appropriate disposal. Its authorship was human. Its limitations were human. It could be questioned, challenged and probed. The probation officer who wrote it could be asked to clarify. That report still arrives. But the professional view it contains is now substantially shaped by something that cannot be questioned in court, cannot be cross-examined and whose inner workings are protected by commercial confidentiality. It is called the Offender Assessment System known as OASys.
That is where we are but a little history of how we arrived here is of some interest. OASys was developed by the Home Office through three pilot studies before being rolled out across the entire prison and probation system in England and Wales between 2001 and 2005. It did not spring from a single moment of invention. The Home Office had commissioned its first statistical predictive tool, the Offender Group Reconviction Scale (OGRS), which was deployed in 1996 across probation offices in England and Wales. This simple algorithm has since been incorporated into OASys which grew to incorporate additional machine learning algorithms. The intellectual framework behind it was the "What Works" movement in criminology; the idea that recidivism could be reduced by matching offenders to evidence based interventions in place of adopting either of the two assessment tools then in use. This revised study of sentencing was based to some extent on pioneering innovation in Minnesota USA in 1978. In 1984 the United States Supreme Court required all judges to use those sentencing guidelines which had resulted. The link between the American systems and the English sentencing guidelines is real although the influence was intellectual rather than structural.
The grading of offences, lists of aggravating and mitigating factors and sentence recommendations traceable to publications by Professor Andrew Ashworth in 1983 and 1987, was ultimately adopted by the Sentencing Guidelines Council and then its successor the Sentencing Council established in 2010.
By January 2025 OASys was producing more than nine thousand assessments every week. The benches of England and Wales are reading its outputs on a daily basis without, in the great majority of cases, knowing quite what they are reading. However, the National Offender Management Service describes its risk scores as the most influential document in the sentencing, planning and management process.
OASys combines what its designers call structured professional judgment with risk prediction algorithms. It weighs static factors such as age and criminal history alongside dynamic ones including accommodation, employment prospects, relationship stability and drug use. Each of those dynamic variables is a proxy for poverty. A defendant who rents insecurely, has no steady job and lives in a postcode where drug use is prevalent will score higher risk not because of anything intrinsic to their character but because the circumstances of their life which the law is supposed to assess as mitigation have been re-encoded as danger. The algorithm does not make moral distinctions. It processes inputs.
The bench that receives this report cannot interrogate the algorithm. The defence solicitor cannot see the weighting applied to each variable. The probation officer who compiled the report may not fully understand how the score was derived. The number arrives with institutional authority, dressed in the language of professional assessment and it influences bail decisions, sentencing outcomes, the choice of disposal and, further down the line, which prison a defendant is sent to and what rehabilitation programmes they are permitted to access. This is, to use an appropriately judicial word, consequential.
When I was sitting the pre-sentence report was a document I read critically. My colleagues and I had been trained to question it, to look for inconsistencies between the circumstances described and the recommendation made, to consider whether the probation officer had taken adequate account of the defendant's own account of themselves. Indeed with my colleagues` assent I often spent almost as much time questioning the report`s writer as I did on the sentence pronouncement. That critical engagement was part of the magistrate's function. It is not clear how a magistrate in 2026 is supposed to exercise the same critical engagement with a risk score generated by a system they have never been told exists, let alone been trained to evaluate although I quite understand how some magistrates with long memories might find themselves at odds with current thinking.
The theoretical appeal of algorithmic decision-making in criminal justice is straightforward and not without merit. A machine, it could be argued, does not have bad days. It does not form unconscious impressions based on how a defendant presents. What an algorithm does inherit is the data on which it was trained. The Lammy Review of 2017 established that at every stage of the criminal justice process in England and Wales people from black, Asian and minority ethnic backgrounds face worse outcomes than white defendants and that those disparities cannot be explained by the nature of the offences with which they are charged. When a predictive algorithm is trained on historical conviction and sentencing data it learns that pattern. It does not learn that the pattern might reflect injustice. It learns that the pattern is the baseline.
There is a wider context. The Court of Appeal's handling of the Post Office Horizon appeals, completed in 2021, established with the force of authority what common sense had always suggested: convictions cannot stand if based on unexamined software outputs. The great majority of those prosecutions was heard in magistrates courts. The benches that convicted those defendants were not acting in bad faith. They were operating within a system that presented computer evidence as reliable, offered no mechanism for challenging it and proceeded on the institutional assumption that if a machine said it, it was probably true. A bench can be challenged on its reasoning. A risk score cannot.
The Horizon software was not artificial intelligence in the modern sense. It was a financial management system that produced erroneous data and whose errors were systematically denied by those with an institutional interest in the system's reliability.
But the principle is identical to the one that now applies in every courtroom where an OASys score informs a sentencing decision. If the output of a computer system is accepted without scrutiny, justice is not being done. It is being performed.
In December 2023 the Courts and Tribunals Judiciary issued guidance to judicial office holders including magistrates on the use of artificial intelligence. The Judicial College identified preparing for innovation and change as a key objective in its activities report for 2023 to 2024. The Ministry of Justice's AI Action Plan, published in July 2025, committed to rolling out enterprise grade AI assistants to every member of Ministry of Justice staff by December of that year. Although substantial progress has been made that aspirational hope has still some way to go. However Microsoft 365 Copilot has been made available to leadership judges following a successful pilot. But to date even so called “leadership magistrates” have not been allocated Copilot licenses. Instead AI use in the magistrates’ courts is described to be at an operational/pilot level.
Whilst the above offers political sound bites [bytes?] for the MOJ for the fourteen thousand or so lay magistrates of England and Wales there is currently no attempt to offer systematic understanding of the algorithmic tools that are already shaping the documents they read in court. There is no requirement that a pre-sentence report disclose the weighting applied by OASys to the variables it has assessed. There is no mechanism by which a defence advocate can challenge a risk score on behalf of their client.
The argument for the lay magistracy has always rested on the principle that local people, drawn from the full range of community life, bring a judgment to the lower courts that is both democratically grounded and practically wise. But that principle requires that magistrates be in full possession of all the material that bears upon the decision before them. A bench that is reading, without knowing it, the output of a commercial algorithm is not in full possession of the material. It is in possession of a summary that someone or something has already partially made up their mind about.
It is a human not a machine which sits in the middle chair. But that ephemeral machine has acquired a significant influence over what happens to the person who stands before it. That influence is invisible, unaccountable and, as far as the magistracy is concerned, but a phantom in time and space. It is now apposite for reality, transparency and confidence in humanity to once again oversee the court process.
