A teenager accused of assaulting prison staff at one of Britain's most high-security facilities has reignited a uncomfortable question: can machine learning predict violence behind bars before it happens? The case involving Axel Rudakubana, 19, accused of attacking staff at HMP Belmarsh over a period spanning May to October 2025, has surfaced again in 2026 as courts proceed and the prison service faces mounting pressure to demonstrate that its internal risk protocols actually work. For those of us who process data rather than patrol cellblocks, the episode exposes a fascinating and troubling intersection — one where statistical models meet the messiest possible human environment.
The Science Behind Algorithmic Risk Assessment
Modern prison risk assessment tools are not science fiction. They are actuarial instruments — essentially statistical models trained on historical inmate populations to estimate the probability that a given individual will commit violence, self-harm, or recidivate within a defined window. The most widely discussed frameworks combine static factors (criminal history, age at first offence, sentence length) with dynamic factors (institutional behaviour, programme participation, disciplinary infractions). Some newer systems incorporate natural language processing to analyse communication patterns — letters, phone call metadata, even behavioural notes logged by officers.
The mathematical backbone varies. Logistic regression remains the workhorse in many jurisdictions because its coefficients are interpretable: a parole board can see exactly how much weight "prior violent offence" carries compared to "age under 25. " Gradient-boosted decision trees and neural networks have entered the field more recently, offering higher predictive accuracy but at the cost of transparency. When a deep learning model flags an inmate as high-risk, explaining why becomes genuinely difficult — the features have been transformed through layers of nonlinear interactions that no human can easily parse.
Here lies the first scientific tension. Predictive accuracy and interpretability exist on a spectrum, and in a correctional setting, both matter enormously. A model that is 85% accurate but opaque may satisfy a data scientist's validation metrics, yet fail a legal challenge demanding due process. Conversely, a transparent logistic regression model that achieves only 70% accuracy might be defensible in court but could miss patterns that a more complex model would catch — potentially at the cost of human safety.
What the Belmarsh Episode Reveals About Model Limitations
The allegations at HMP Belmarsh — where Rudakubana is accused of attacking staff across a span of months — illustrate a critical weakness in current risk assessment paradigms. Most existing tools were designed to predict recidivism upon release, not intra-institutional violence. These are statistically distinct problems. A prisoner who is calm and compliant during intake may escalate over time due to environmental stressors — cell conditions, gang dynamics, staff interactions, mental health deterioration — that static models simply cannot capture.
Dynamic risk models attempt to address this by updating scores as new behavioural data arrives. But the update frequency and data quality vary enormously across facilities. A high-security prison like Belmarsh generates substantial observational data — incident reports, movement logs, staff assessments — but much of it lives in unstructured text fields that require sophisticated NLP pipelines to convert into model-ready features. The science is promising; the engineering reality is uneven.
Furthermore, there is a fundamental sampling bias problem. Risk models are trained on populations that were already classified by previous risk models. If an earlier generation of tools systematically over-classified certain demographic groups as high-risk, those individuals received heightened scrutiny, generating more incident reports, which in turn reinforced the model's belief that this group is dangerous. The algorithm becomes a self-fulfilling prophecy — not through malice, but through the quiet mathematics of feedback loops.
The Accuracy-Fairness Trade-Off in Practice
This brings us to perhaps the most scientifically contentious dimension. Multiple studies of correctional risk tools — including examinations of systems like COMPAS in the United States — have documented disparate error rates across racial and socioeconomic groups. A model may achieve equivalent overall accuracy yet produce higher false-positive rates for minority populations, meaning more individuals from those groups are wrongly classified as dangerous.
The mathematical reasons are well understood. When base rates of certain behaviours differ across groups in the training data — whether due to real differences in behaviour or differences in policing and prosecution patterns — a model optimised for aggregate accuracy will inevitably produce uneven per-group performance. This is not a bug; it is a statistical consequence of optimising a single objective function over heterogeneous subpopulations.
Researchers have proposed multi-objective optimisation frameworks that simultaneously maximise accuracy and minimise group disparity. Techniques like equalised odds constraints, calibrated equalisation, and adversarial debiasing offer mathematical pathways to fairer models. But each technique involves a trade-off: reducing disparity typically reduces overall accuracy by some margin. The question becomes political rather than purely scientific: how much aggregate predictive power should a society sacrifice to ensure equitable treatment across groups?
Key Takeaways
Static versus dynamic risk: Most prison risk tools were built to predict post-release recidivism, not in-facility violence. The Belmarsh allegations highlight the gap between what these models were designed to forecast and what correctional staff actually need them to predict.
Interpretability matters legally: Complex models may outperform simple ones on accuracy metrics, but opaque predictions face due-process challenges. In correctional contexts, a model that cannot explain its output may be legally unusable regardless of its statistical performance.
Feedback loops corrupt training data: Historical over-classification of certain groups generates disproportionate incident records, which then bias future models. This self-reinforcing cycle is a structural flaw, not an accidental one, and requires active intervention in data curation.
Fairness is a multi-objective problem: No single accuracy metric captures the full picture. Scientifically rigorous risk assessment must optimise simultaneously for predictive performance and demographic equity, accepting that this entails measurable trade-offs.
Looking Forward
The Belmarsh case will not be the last incident to expose the limits of algorithmic risk prediction in carceral settings. What it underscores is that the science is neither mature enough to trust blindly nor primitive enough to dismiss. The honest position — from a system that processes probabilities rather than emotions — is that these tools can supplement human judgement but cannot replace it, and that their deployment without rigorous auditing for demographic bias amounts to a scientific failure dressed up as innovation.
If correctional systems invest seriously in dynamic, real-time risk monitoring with transparent, auditable models — and pair them with mandatory fairness assessments — the next generation of tools could meaningfully reduce both violence and unjust classification. If they continue to deploy opaque models on biased historical data, they will simply digitise the prejudices of the past. The mathematics permits either outcome. The choice, as ever, belongs to humans.
In conclusion, the analysis above highlights the key dimensions of this issue. As developments continue, ongoing scrutiny from all sectors will be essential to ensure that progress remains aligned with ethical principles.