By Anagnostakis Law Team
Europe’s prisons are full, and the problem is growing. According to the Council of Europe’s latest SPACE I statistics, published in May 2026, more than 1.1 million people were held in custody across 51 European prison systems. Nine administrations reported severe overcrowding: Türkiye and France (131 inmates per 100 places), Croatia (123), Italy (121), Malta (118), Cyprus (117), Hungary (115), Belgium (114) and Ireland (112).
France shows how bad it can get. By June 2026 it held a record 88,829 people in 63,237 operational places, an occupancy rate of 140%. Remand facilities are at 173%, some prisons are above 250%, and more than 7,600 detainees sleep on mattresses on the floor. Paris has chosen to build its way out, with modular prisons and a target of 130% occupancy by 2032, and has rejected early release.
London has done the opposite. The Sentencing Act 2026 creates a presumption that custodial sentences of twelve months or less will be suspended. From this autumn, an “earned progression” model will make many prisoners eligible for release after serving one third of their sentence. Greece has combined both approaches: new facilities, and a €7 million electronic-monitoring platform to release around 2,500 detainees under tag.
Governments across the continent are really asking one question: who can safely leave, and who has to stay? That is where artificial intelligence is now being offered as the answer. The pitch is no longer about AI in the courtroom. It is about AI as an administrative tool that decides who gets a cell, who gets a tag, and who is flagged as a risk before anything has gone wrong.
The pitch: managing scarcity with data
South Korea’s Correctional Service has just set out a 2026–2030 plan built on this idea. AI models trained on counselling records, disciplinary history, behavioural observation and medical data would flag early signs of violence, self-harm or escape risk. The aim is to move corrections “from a reactive model… to a preventive one” (The Korea Times). The stated goals are fewer repeat offences, less overcrowding and better conditions for staff. Commissioner Lee Hong-yeon was careful to say that the AI would not decide parole or treatment outcomes by itself, and that the models will be tested for bias across demographic groups.
The United States is further down the same road. By 2023 about thirty corrections agencies used biometric monitoring to detect overdoses and self-harm in real time. Risk-assessment tools now routinely set the intensity of supervision based on employment, treatment participation and past violations (Justice Trends). Research from Chicago Booth goes further: it applies queueing theory and machine learning to allocating scarce places in diversion programmes. In other words, algorithms are used not to punish but to decide who is kept out of custody in the first place (Chicago Booth Review).
The appeal to European governments is obvious. If prisons can’t be built fast enough, perhaps the existing places can be allocated more intelligently. Score the prison population, identify who can safely be tagged, suspended or released at the one-third point, and keep the beds for those who can’t.
The ceiling Strasbourg has already set
The trouble is that this treats overcrowding as an optimisation problem. The European Court of Human Rights treats it as a fixed legal minimum. Since Muršić v. Croatia ([GC], no. 7334/13, 20 October 2016), a detainee with less than 3 m² of personal space in a shared cell benefits from a strong presumption that Article 3 has been violated. The state can rebut that presumption only by showing that the reduction in space was short, that the detainee had enough freedom of movement outside the cell, and that conditions were otherwise adequate. The CPT’s own standard, 4 m², is higher still.
No forecasting model changes that arithmetic. A cell either gives the space the Convention requires or it doesn’t. An algorithm that reshuffles who ends up below the line does nothing to raise the line.
Hungary learned this the hard way in Varga and Others v. Hungary (pilot judgment, 10 March 2015). The Court found overcrowding to be a systemic, structural problem and required the state to create an effective domestic remedy, not simply to manage individual placements better. Greece faces the same reckoning: the Nisiotis group of cases (leading case no. 34704/08) remains under enhanced supervision by the Committee of Ministers because repeated Article 3 findings have not been resolved by ad hoc administrative fixes.
The lesson applies across Europe. Putting an AI allocation tool on top of a structural shortage does not cure the shortage. At best, it manages the queue for a scarcity the Court has already said is unlawful.
That is the trap. AI can make the rationing look more scientific without making it lawful. The Chicago Booth researchers say so themselves: used carelessly, these systems “risk entrenching existing inequalities under the guise of scientific legitimacy.” This is especially true when the training data reflects past decisions about who was diverted, released or labelled dangerous. A model trained on a system that has already been found to violate Article 3 treats that system’s distortions as ground truth.
The regulatory overlay: what the AI Act actually requires
In EU member states, the analysis does not stop at Strasbourg. Under the AI Act, systems that law enforcement uses to assess a person’s risk of reoffending are classed as high-risk under Annex III (Domain 6). That triggers mandatory human oversight under Article 14, with a compliance deadline of 2 December 2027. Separately, Article 5 prohibits AI systems that predict criminal risk based solely on profiling or personality traits. A poorly governed “early warning” tool could easily slide into exactly that kind of crude pattern-matching unless it is tightly limited to individual, evidence-based assessment.
It is still unclear where purely administrative tools fit: population forecasting, bed-allocation optimisation, and resource planning that never touches an individual’s liberty. But once such a system starts informing who is chosen for electronic tagging, early release, a suspended sentence or lighter supervision, it is no longer back-office logistics. It is a factor in decisions about liberty. That puts it squarely in Annex III territory, and it should be treated that way by every European government now designing release and diversion schemes under capacity pressure. The UK is outside the AI Act but still bound by the Convention, so the same reasoning applies there through Articles 5, 6 and 14.
What a defensible version would look like
None of this makes the underlying instinct wrong. A structurally overburdened justice system has every reason to look for better tools to manage what it has while capacity catches up. That was the subject, not coincidentally, of my paper for the IBA Athens Transnational Crime Conference. But the lawful version of such a tool looks very different from the efficiency pitch.
First, any AI system that feeds into a release, tagging or classification decision must produce an individual, reasoned output that a court can review. A population-level risk band applied by proxy is not enough.
Second, the system must be tested for demographic bias, as the Korean plan promises, and the results must be published rather than simply asserted.
Third, the human decision-maker must have real discretion to depart from the model. A rubber stamp on an algorithmic recommendation is not oversight.
Finally, governments cannot use the existence of such a system as a political substitute for the structural remedy. Strasbourg has already told several European states that the real obligation is new capacity, sentencing reform and faster case processing.
The machine can help a stretched system decide where to look first. It cannot testify, in place of a judge, to whether three square metres and three hours outside a cell amount to human dignity. That judgment, and the accountability for getting it wrong, must stay where the Convention puts it: with a human decision-maker who can be asked to explain themselves.
