The Influence of Error on Perceptions of Machine Learning vs. Clinician-Based Risk Assessments

Risk assessments are key legal tools that can inform a number of legal decisions regarding parole sentencing and predict recidivism rates. Due to assessments being historically performed by humans, they can be prone to bias and have come under various amounts of scrutiny. The increased capability and application of machine learning technology has lead the justice system to incorporate algorithms and codes to increase accuracy and reliability. This study researched laypersons’ attitudes towards these algorithms and how they would change when exposed to an algorithm that made errors in the risk assessment process. Participants were tasked with reading two vignettes and answering a series of questions to assess the differences in their perceptions towards machine learning and clinician-based risk assessments. The research findings showed that individuals lent more trust to clinicians and had more confidence in their assessments when compared to machines, but were not significantly more punitive when it came to attributing blame and judgement for the consequences of an incorrect risk assessment. Participants had a significantly more positive attitude towards clinician-based risk assessments, noting their assessments as being more reliable, informed, and trustworthy. Participants were also asked to come to a parole decision using the assessment of either a clinician or machine learning algorithm at the end of the study and rate their own confidence in their decision. Results found that participants were only significantly less confident in their decision when exposed to previous instances of risk assessments with error, but that there was no significant difference in their confidence based solely on who conducted the assessment.