Predictive Algorithms In The Justice System

Predictive Algorithms In The Justice System: Implications Of Neutrality Logics Presented by: Dr. Ugwudike Date: 19 February 2021 Predictive Algorithms In The Justice System: Implications Of Neutrality Logics - This presentation: • Focuses on data-driven predictive algorithms applied in the justice system. • Draws on insights from the sociology of race to demonstrate the implications of race neutrality logics. • Proffers corrective measures. Predictive Algorithms In The Justice System - - Predictive algorithms are applied in key areas of criminal justice decisions making: • Policing (e.g. bail decisions, forecasting crime locations) • Sentencing • Probation (e.g. programs & early revocation) • Prison (e.g., parole) History & evolution of risk prediction and management (Feeley and Simon 1992; Garland 2001; O'Malley 2001: O'Malley and Smith, Forthcoming. 2021). Implications of Race Neutrality Logics: Masking Discretionary Decision Making - The algorithms are ostensibly race neutral with no overt references to race. - They reflect the hallmarks of liberal progress towards 'post racial" race neutrality (described as unrealistic abstract liberalism - e.g., Mills 2017, see also, Bonilla-Silva 2015; Feagin 2013; Golash-Boza 2016) - Structural conditions can sustain racial bias regardless of apparent race neutrality (Obasogie 2013; Feagin 2013; Mills 2017). - E.g., conditions that enable discretionary decision making (Murakawa & Beckett 2010). Implications of Race Neutrality Logics: Masking Discretionary Decision Making - Potential conduits of Bias: • Selecting data • Selecting risk predictors • Selecting recidivism descriptors (see also Eaglin 2017) Implications of Race Neutrality Logics: Masking Discretionary Decision Making - Selecting Data • E.g., Biases in administrative data can artificially inflate risk  'Predictions reflect the data used to make them. If black defendants are arrested at a higher rate than white defendants in the real world, they will have a higher rate of predicted arrest as well. This means they will also have higher risk scores on average, and a larger percentage of them will be labelled high-risk' (Hao and Stray (2019)  Angwin et al. (2016) o 'There's software used across the country to predict future criminals. And it's biased against blacks.'  Lum and Isaac (2017): o Predictive policing algorithms can over-predict crime in BAME locations. - Selecting risk predictors - Commonly-used predictors include: • Criminal history & deficiencies in education & employment (See, Bonta and Andrews 2007; Desmarais et al. 2016; Hamilton 2015). • Black people experience exclusion from education and employment at a higher rate.  (Department of Education 2016) (Office for National Statistics 2011). - Selecting risk predictors: • - 'Hospital 'risk scores' prioritize white patients' (Price 2019). Selecting Recidivism Descriptors: • The creators define recidivism. • Some definitions can introduce bias. • E.g. Biases in administrative/criminal justice decision-making can artificially inflate risks. The Alan Turing Institute - Research Project: A Multidisciplinary Study of Predictive Artificial Intelligence Technologies in the Criminal Justice System - P.I. Dr Pamela Ugwudike (Criminology), Co-I Dr Adriane Chapman (Data Science), Research Fellow Dr Phillip Grylls (Software Engineering) - The study: • Objective: Identify potential conduits of bias • Method: Largescale tests of data to observe impact on predictions • Finding: Evidence of self-reinforcement Implications of Race Neutrality Logics: Masking Uneven Distribution of Algorithmic Harms: - Digital capital • The power or ability to acquire resources required for creating & operationalizing influential digital technologies (e.g. Van Dijk 2005). - Digital exclusion • Affects those who lack requisite digital capital • Their values and preferences are excluded and they are vulnerable to digital harms (e.g. Mervyn et al. 2014). • - Digital divide • - Digital divide Separates those with digital capital from others (e.g. Horton 2004). (Halford and Savage 2010; Horton 2004; van Dijk 2005). Possible Solutions? - Diversify & democratize creational processes. - Introduce legal frameworks. - Require audits & Impact assessments. - (See generally, CDEI 2020; EDRI 2021; House of Commons 2018; Law Society 2019; Ada Lovelace Institute 2020)