Closing the Quality Assurance Feedback Loop: Independent Recommender System VALidator

Recommender systems are endemic, increasingly used for social decision-making processes, reducing a complex human to a single data item, and are unfortunately known for bias particularly when used in the context of criminal justice. Administrative demands, ‘doing more with less’, ensure that recommender systems will remain in use to varying degrees in different jurisdictions. Consider the potential circumstances of a judicial officer using a false positive bail recommendation to release a serial rapist into the community, or a false negative bail recommendation to remand a defendant where they have carer's responsibility for elderly parents or children. Quality assurance including quality control activities are standard in other fields, namely medicine, however software quality assurance for recommender systems supporting social decision-making processes has not yet matured, focusing on technical internal system improvements. When the outcome of a data-driven recommender system can erroneously drastically impose on freedom, we need to have reliable quality assurance alongside the ability to invoke our “right to be an exception” [1]. This paper briefly outlines the co-design and testing of a novel quality control validation tool (VAL *©) that highlights such exceptions, independent of recommender systems, with case studies. VAL re-humanizes the defendant for the judicial officer or prosecutor, providing key specific circumstances alongside the recommendation, enabling better informed outcomes.