When Risk Models Disagree, Take the Lowest Score, Study Says

What does the lowest-score rule change for accuracy?

Nothing, according to the data — and that's the point. When the researchers applied the simple rule of taking the lowest risk score for every inmate, they found no loss in overall predictive performance. The models mostly agreed anyway; when they didn't, favoring the person cost the system nothing.

Singh put the principle bluntly: "The uncertainty involved in selecting a model should not be at the expense of the individual concerned."

The rule also reduces reliance on any single model, which matters for what the researchers call stability — whether an individual's assessment shifts depending on which model happens to be in use. Under the EU's AI Act, recidivism tools count as high-risk applications. The study argues that for such systems, accuracy and low error-rate disparities across demographic groups aren't enough on their own. Individual assessments need to be stable too.

What did the team actually build?

Working with researchers from Catalonia's Department of Justice, Singh and Castillo translated complex legal rules into an algorithm. That let them systematically determine, in the legal sense, whether a person reoffended after release. The result was a dataset covering more than 17,500 prison releases spanning over 15 years of system use.

On that data, they built interpretable linear models with improved predictive performance and smaller differences in error rates across demographic groups — comparing, among others, men versus women and people with versus without Spanish nationality.

They also added technical safeguards so that rehabilitative progress made during imprisonment lowers the risk score. "An assessment system should actually reflect positive changes as such," Singh said. "Rehabilitative progress should never lead to a higher risk assessment simply because of the mathematical structure of a model."

That sentence deserves a pause. It implies earlier model structures could, in principle, penalize progress — a limitation worth keeping in mind when evaluating any scoring tool.

Does a human still decide?

Yes, and the researchers insist on it. In Catalonia, as in most European countries, both algorithmic and human assessments run during imprisonment and get compared. The final call stays with a person. "It makes sense to draw on both assessments," Singh said. "However, since the data is subject to a high degree of noise, the decision-making responsibility must ultimately remain with a human expert."

Those decisions also involve trade-offs a score can't capture: public protection, the prisoner's circumstances, their progress behind bars, rehabilitation goals, and even institutional pressures such as prison occupancy rates.

The caveats

A risk score is a probability, not a verdict. It doesn't tell you whether someone will reoffend. And this is one study of one regional system, presented at a conference rather than appearing in a peer-reviewed journal. The lowest-score rule held up here; whether it generalizes to other jurisdictions and other tools remains untested.

Still, the finding offers something rare in algorithmic accountability: a free lunch. If picking the gentler score costs no accuracy, the burden of proof shifts to anyone who would do otherwise.

The paper: Ashwin Singh et al., "Model Multiplicity and Predictive Arbitrariness in Recidivism Risk Assessment," Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (2026). DOI: 10.1609/aies.v9i4.43311.