The Present and Future of AI in Pre-Trial Risk Assessment Instruments

From music and romantic partner recommendation, to medical diagnosis and disease outbreak detection, to automated essay scoring,“Artificial intelligence”(AI) systems are being used to tackle prediction, classification, and detection tasks that impact nearly every sphere of our lives. Since the fundamental task of pre-trial risk assessment instruments is one of prediction, we anticipate that the success of AI technology in these other domains will inspire an increase in the availability of AI-based risk assessment instruments in the coming years. The purpose of this critical issue brief is primarily to equip practitioners considering adopting an AI-based pre-trial risk assessment tool to consider the relevant questions relevant to determining whether adopting such a system will result in better predictions and ultimately move their jurisdiction towards fairer, more just and decarceral pre-trial decision-making that respects civil and human rights.

AI technologies are not likely to achieve considerably greater predictive accuracy than currently available risk assessment instruments. The primary obstacle is that the behavioral outcomes these tools seek to predict, outcomes such as future arrest or court non-appearance, have an inherently high degree of uncertainty or randomness. Existing evidence suggests that any significant increase in predictive ability would need to come from uncovering hereto unknown and unused highly predictive risk factors. It is reasonable to doubt whether such factors exist. Furthermore, to the extent that incorporating new types of data does produce gains in predictive accuracy, those gains must be weighed against the ethical concerns …