LAUSANNE. Three cantonal courts are quietly testing software that proposes sentencing ranges based on thousands of past judgments, and a case now heading to the Federal Supreme Court will decide how far that experiment may go.
The appeal arises from a burglary conviction in a German-speaking canton, where the trial judge acknowledged in the written judgment that the sentence of thirty months sat within the range proposed by a statistical tool fed with a decade of cantonal case law. The defendant contends the tool converted precedent into presumption, and that his punishment reflects the average of other men's crimes rather than the particulars of his own.
The legal question is framed narrowly but reaches far. Swiss law guarantees an individualised sentence weighing culpability, prior record and personal circumstances. The appellant argues that when a judge adopts an algorithmic recommendation without independently justifying the outcome, the sentencing decision is no longer the court's own, violating both the right to a lawful judge and the prohibition on arbitrariness.
The tools at issue are modest compared with predictive policing systems elsewhere. Built by cantonal IT units with university partners, they ingest anonymised judgments and produce a distribution: for a given offence category and offender profile, the median sentence, the interquartile range and the outliers. Judges say they consult the output the way they might consult sentencing statistics compiled by hand.
Proponents inside the judiciary point to a genuine problem the tools address. Studies by the Swiss Centre of Expertise in Criminal Justice have found unexplained disparities between cantons and even between courtrooms in the same building, with comparable offences attracting sentences differing by a third or more. Consistency, the argument runs, is itself a component of equality before the law.
The canton involved reports that judges followed the tool's proposed range in about 60 per cent of eligible cases during the pilot's first year, departing upwards in 25 per cent and downwards in 15 per cent. Court administrators emphasise that no sentence was issued without a reasoned judgment, and that the software has no access to names, addresses or ethnicity.
Defence lawyers are unconvinced. They note that the training data embodies the very disparities the tool is meant to cure, and that anchoring on a computed median risks hardening historical practice into law. The Swiss Bar Association has asked that defendants be told whenever a tool was consulted, a disclosure the cantonal pilot does not currently require. Several defence teams have begun adding a standard question about software assistance to their appeal checklists.
A professor of criminal law at the University of Lausanne said: “The danger is not that the machine sentences, but that the judge's reasoning quietly migrates towards the machine's number.” Empirical research on anchoring effects among judges, she added, suggests even experienced decision-makers drift towards suggested figures.
Cantonal justice departments defend the pilot as an efficiency measure at a time of chronic understaffing. Several courts face backlogs exceeding a year for non-custodial cases, and administrators argue that faster access to sentencing patterns frees judicial time for the cases that genuinely demand it. They also note that participation was voluntary, and that roughly a third of eligible judges declined to use the tool at all.
The Federal Supreme Court is expected to hear the appeal in the autumn session. Its ruling will likely turn less on the technology than on reasoning: whether a judgment that cites a statistical range alongside the statutory factors can still be reviewed meaningfully, or whether review requires knowing how the range was produced.
Whatever the outcome, the case has already fixed the terms of the debate. Legislatures in at least four cantons have paused their own digitisation plans pending the Lausanne decision, and the Conference of Cantonal Justice Directors is drafting guidelines that would require disclosure, audit rights and a documented judicial override whenever a tool's suggestion is not followed. A ruling is expected before the end of the year.