For some years now, in many courtrooms around the world, a number has been appearing alongside the case file: a risk score, computed by a system trained on tens of thousands of past cases, estimating the probability that a person will offend again. This is so-called predictive justice, and it is one of the most delicate grounds on which artificial intelligence meets people's lives. In the chapter «Guilt and the Algorithm» of Beyond Turing - The Incalculable Remainder, Antonio Fabbrizio addresses the subject with the tools of information-systems engineering and a clear-cut thesis: a score can be computed; guilt must be judged.
The argument is not ideological - it is technical. A recidivism risk model is, in essence, a big data pipeline: input variables, supervised training on historical outcomes, calibration, a score as output. The chapter dismantles its three critical points with engineering precision.
The three structural limits of the risk score
The first limit is statistical: the score is a statement about a class of people, not about an individual. It says that, among similar profiles, a certain fraction has had a certain outcome in the past; it says nothing necessary about the single case. The second is semantic: judicial archives record administrative events - arrests, complaints, convictions - which also depend on where and how enforcement looked, not a person's moral inclination; a model trained on that data learns the geography of policing together with the phenomenology of crime, and returns both, indistinguishably, as risk. The third is decisional: every threshold applied to the score produces false positives and false negatives, and in this domain neither error is just a number - on one side a person suffering an unjustified restriction, on the other a potential victim left unprotected.
The empirical research cited in the chapter confirms the need for sobriety: after the journalistic investigations into actuarial recidivism tools, studies formally demonstrated that the different mathematical definitions of fairness are mutually incompatible when base rates differ across groups, and that the accuracy of one of the most widely used tools is comparable to that of human assessors with no specific training.
The judge, cognitive bias and the AI Act
The chapter has a second, more surprising movement: it does not set an idealized human judgment against the algorithm. Cognitive science has documented for half a century that the human mind, at every level of excellence, is exposed to involuntary conditioning - numerical anchoring, media pressure, cognitive fatigue - acting on all the actors of a trial, in every direction. And precisely for this reason, the author observes, legal civilization has built over the centuries what an engineer would read as a fault-tolerant architecture: adversarial proceedings, the impartiality of the judge, the levels of appeal, the duty to state reasons, reasonable doubt.
«The algorithm may inform judgment; it may not, at any point in the pipeline, replace it.»- Beyond Turing, ch. 15, Guilt and the Algorithm
Within this frame, the algorithmic score must be treated for what it is: the latest arrival among the factors that can condition the decision-maker, made more insidious by automation bias - the perceived authority of the number, greater than its real reliability. European law has drawn the boundaries: Regulation (EU) 2024/1689 (the AI Act) prohibits systems that assess the risk of offending on the basis of profiling alone (Article 5), classifies AI systems for the administration of justice as high-risk (Annex III) and, with Article 14, demands that the decision remain in the hands of the one who judges. It is the same lesson as State v. Loomis: the actuarial score never as the determining factor of the decision.
Digital prejudice and the right to be forgotten
Finally, there is a conditioning that acts outside the courtroom and that the chapter brings into sharp focus: digital prejudice. The web indexes the accusation and ignores the silent tail of the acquittal; the result is a penalty no code provides for - permanent reputational conviction, striking the acquitted innocent with the same efficiency with which it files away the guilty. The right to be forgotten (Court of Justice of the EU, Case C-131/12; Article 17 GDPR) has opened a breach, but the underlying question is anthropological: man can change, data cannot. Legal orders know rehabilitation; the digital profile freezes the person in their worst instant.
It is a chapter that speaks to jurists, engineers and citizens together, in a measured register and with rigorous documentation - from Beccaria to Ferrajoli, from Tversky and Kahneman to Directive (EU) 2016/343 on the presumption of innocence. And it delivers a conclusion that stands as a design principle for anyone building information systems for justice: the system must know when it does not know - and stop.
