SHAMENA ANWAR, JOHN ENGBERG, ISAAC M. OPPER, LEAH DION
What Happens When
Judges Follow the
Recommendations of
Pretrial Detention Risk
Assessment Instruments
More Often?
T
he use of artificial intelligence (AI) methods to aid with decisionmaking in the criminal
justice system has widely expanded with the increased use of risk assessments, which are AI
tools that use formal, actuarial, and algorithmic methods to predict future misconduct (Ste-
venson, 2018). Nowhere has this shift been more dramatic than in the widespread adoption
of AI-enabled risk assessment tools to aid in pretrial detention decisions. Until recently, jurisdictions
relied primarily on a money bail system to determine pretrial release, whereby judges set a monetary
bail, and defendants who could not afford to pay were required to remain in jail until their criminal
charges were adjudicated in the courts. To increase fairness, and to reduce overall pretrial detention
rates and ensure that only the riskiest individuals are detained, criminal justice reform efforts have
focused on using risk assessments—in which individuals assessed to be low and medium public
safety risks are recommended for release and individuals assessed to be high-risk are recommended
for detention during the pretrial period—as an alternative to the money bail system. To eliminate
detention disparities based on wealth, pretrial risk assessment tools do not recommend monetary
bail. Roughly half of the counties within the United States currently use a pretrial risk assessment
(Hochstedler Webb, Riley, and Wells, 2023).
The promise of reforming the monetary bail system by using risk assessment tools is substan-
tial; recent studies have shown that moving from a situation in which judges make pretrial release
decisions without consulting the tool to one in which the recommendation of the tool is always
followed would result in many more individuals being released during the pretrial period, with
no negative public safety impacts (Kleinberg et al., 2018; Angelova, Dobbie, and Yang, 2023). For
Research Report
