Threats and Opportunities in the Use of AI for Lawyers
Lawyers must understand the flaws in machine decision making in order to challenge the use of tools in the criminal justice context – just as they would challenge problematic human reasoning or actions.
Plentiful critiques of the use of machine learning and big data analytics in policing and criminal justice system processes are now emerging. A number of these critiques take as their starting point issues concerning the use of predictive analytics more generally. The way the software has been built and trained may be problematic and therefore result in unreliable or biased outputs. There are also broader questions about exactly how ‘fairness’ in such systems is being defined. Finally, there are enormous issues related to lack of transparency. The most widely used programs, such as PredPol and COMPAS, are proprietary and there is therefore no transparency as to how they operate. For this reason, it is difficult for anyone – defendants, lawyers, journalists, academics, even judges – to independently evaluate whether or not a system’s outputs are indeed reliable.
Eric Loomis appealed to the Wisconsin Court of Appeal and his case – that use of the COMPAS risk assessment in sentencing violated his due process rights – was heard by a bench of three judges. Loomis’s appeal on this point had three elements. The first was that the workings of COMPAS are opaque and unknowable, and therefore he could not be certain that he had been sentenced according to accurate information. Secondly, Loomis claimed that the use of the tool had violated his right to an individualised sentence. Finally, because one feature utilised by COMPAS is gender, he claimed that this had been improperly considered, in violation of constitutional protections guaranteeing equality before the law. The Wisconsin Court of Appeal unanimously found that Loomis’s due process rights had not been violated by the use of COMPAS.
Subsequently, in 2017, the US Supreme Court denied Loomis’s petition for certiorari – in other words, declined to hear a further appeal. Though it rejected Loomis’ claims, the Court of Appeal did note that the use of COMPAS for sentencing raises due process concerns. In response, it set out some parameters or guidance as to the future use of comparable risk assessment tools. The separate judgments of Roggensack CJ and Abrahamson J, in particular, stress that COMPAS should be used cautiously and selectively. Evidence suggests that this is probably inadequate, yet risk assessment tools are becoming ubiquitous in the US criminal justice system. In many other jurisdictions it seems likely that the use of such tools would not be permissible – or would be permissible only with safeguards such as demonstrable accuracy and transparency.
Author(s): Michael Legg & Felicity Bell