Audit slams RisCanvi justice algorithm as opaque, unreliable and unfair
Audit slams RisCanvi justice algorithm as opaque, unreliable and unfair
Occurred: July 2024
Report incident 🔥 | Improve page 💁 | Access database 🔢
An audit revealed serious shortcomings associated with an algorithm used in Catalonia's criminal justice system to predict how likely prisoners are to reoffend.
An "adversarial" audit of RisCanvi conducted by algorithmic auditing company Eticas concluded that the algorithm suffered from “substantial deficiencies” in the reliability of its decisions by failing to standardise outcomes and for being potentially biased against specific types of people, particularly those with challenging backgrounds.
Eticas also found that the system had not adhered to Spain’s regulations on the use of automated decision-making systems, which have required audits since 2016, and that it had not undergone an Algorithmic Impact Assessment or rigorous testing.
Furthermore, the audit concluded that inmates, who are the people most affected by the system, were unaware of its existence, while staff using it lacked proper training and understanding of how it worked.
RisCanvi's shortcomings appear to stem from the poor design and training of the system, and from inadequate oversight and poor transparency.
It is unclear whether its transparency and accountability shortcomings constituted a deliberate attempt to conceal the system's inadequacies.
The case highlights the need for greater transparency and accountability of so-called "high-risk" systems that can have a major impact on people's lives.
It also suggests the RisCanvi algorithm needs to be comprehensively overhauled and more rigorous testing and oversight introduced.
RisCanvi
Operator: Catalonian Department of Justice
Developer: Antonio Andrés Pueyo
Country: Spain
Sector: Govt - justice
Purpose: Predict prisoner recidivism risk
Technology: Risk assessment algorithm
Issue: Accountability; Accuracy/reliability; Bias/discrimination; Transparency
Eticas. RisCanvi Adversarial Audit
Eticas. Case study
Page info
Type: Incident
Published: February 2025