Ghosting the Machine: Judicial Resistance to a Recidivism Risk Assessment Instrument

Scholarship on predictive technologies in the public sector has exploded in recent years (Chouldechova, 2017; Brown et al., 2019; Fogliato et al., 2021; Levy et al., 2021; Akpinar et al., 2021; Meyer et al., 2022). The use of algorithmic decision-making in the criminal legal system has been particularly controversial, with reason. The claim that risk assessment instruments promote progressive criminal justice goals in practice is largely speculative – the few existing empirical studies suggest that risk assessment tools have had little to no impact (Stevenson, 2018; Sloan et al., 2018; Garrett and Monahan, 2020; Stevenson and Doleac, 2021) – and a vocal chorus of critics has stressed that such instruments could exacerbate racial disparities in pretrial, sentencing, and parole decisions because they base predictions on (and reproduce) structurally racist patterns in the US criminal legal system (Harcourt, 2008; Hannah-Moffat, 2013; Angwin et al., 2016).

To be sure, algorithmic bias is worth addressing seriously and can be reason alone to condemn the use of a particular instrument. But a key detail often neglected in discourse about risk assessment instruments and other public sector algorithmic systems is that their recommendations are advisory. Algorithmic systems are socially situated, interacting and entangling by necessity with people, institutional practices, and societal norms (Alkhatib and Bernstein, 2019; Mittelstadt et al., 2016; Pruss, 2021; Glaser et al., 2021). Individuals like judges and police officers make on-the-ground discretionary decisions – what Michael Lipsky refers to as ‘street-level bureaucracy’ (Lipsky, 1980) – that ultimately impact the lives of individual people, not the technical details of the algorithmic instruments on their own, and human judgment can interact with algorithmic decision-making systems in unexpected ways. The few studies of how risk assessment instruments are actually used have shown that judges differ widely in their adherence to recommendations and follow them inconsistently for different types of defendants (Stevenson, 2018; Garrett and Monahan, 2020; Stevenson and Doleac, 2021).

For example, human decision-makers can selectively follow algorithmic recommendations to the detriment of individuals already likely to be targets of discrimination. In Kentucky, a pretrial risk assessment tool – intended as a bail reform measure – increased racial disparities in pretrial releases and ultimately did not increase the number of releases overall because judges ignored leniency recommendations for Black defendants more often than for similar white defendants (Albright, 2019). Likewise, judges using a risk assessment instrument in Virginia sentenced Black defendants more harshly than others with the same risk score (Stevenson and Doleac, 2021).

In other contexts, human discretion can correct for algorithmic bias. In Pennsylvania, a recent study about racial bias in an algorithm that screens for child neglect showed that call screeners minimized the algorithm’s disparity in screen-in rate between Black and white children by “making holistic risk assessments and adjusting for the algorithm’s limitations” (Cheng et al., 2022) (see also (De-Arteaga et al., 2020)). Virginia’s risk assessment instrument would have led to an increase in sentence length for young people had judges adhered to it; however, because judges systematically deviated from recommendations, some of the instrument’s potential harms (and benefits) were minimized (Stevenson and Doleac, 2021).

Of course, another way that human discretion can interact with algorithms is not to interact with them. Algorithm aversion – the reluctance to follow algorithmic recommendations – is thought to arise from lack of confidence in algorithmic systems (Dietvorst et al., 2015); however, experimental research on algorithm aversion has focused on individual and algorithm factors, neglecting the role of social context and organizational factors (Mahmud et al., 2022). Sociological work shows that resistance to algorithms happens in contexts where individuals feel that their agency or power is being threatened by a new technology, as illustrated by Sarah Brayne in her ethnography of LAPD officers using PredPol, as well as by Angèle Christin in her ethnography of prosecutors and judges using a pretrial risk assessment instrument (Brayne, 2020; Christin, 2017). Police officers and legal professionals alike felt threatened by how these new technologies could be used to surveil their performance and limit the role of their discretion, resulting in professional resistance to algorithmic systems in the form of adversarial data obfuscation – the process of manipulating a system’s data to make it useless – and foot-dragging.

These dynamics can also intersect. In Virginia, judges had highly divergent attitudes toward (and literacy about) risk assessment and varied widely in whether and how they adhered to algorithmic recommendations (Garrett and Monahan, 2020). Understanding how these possible forms of human-algorithm interaction apply in a given case thus requires not only empirical research in a context of application but also attention to the social and organizational factors at play.

Amid the chaos of the early months of the pandemic, criminal courts throughout Pennsylvania were instructed to begin consulting the Sentence Risk Assessment Instrument when sentencing crimes, with the aim of helping judges identify candidates for alternative sentences. The instrument applies to non-DUI defendants being sentenced following an open plea or trial, July 2020 onward. It generates a risk score of an individual’s risk of recidivism based on demographic factors including age, gender, number of prior convictions, current conviction offense type, and prior juvenile adjudication. These variables are given weights, depending on their degree of association with reconviction. Young age, for example, is associated with recidivism; individuals under the age of 21 receive 5 points and those over 49 receive 0 points. These points are summed to make the risk score, which ranges from 0 to 18 and is binned into low-, typical-, and high-risk categories.

The tool recommends seeking ‘Additional Information’, typically a pre-sentence investigation report (PSI), for individuals with a low or high risk of recidivism “for whom additional information may assist the court in determining candidates for alternative sentencing” (Pennsylvania Commission on Sentencing, 2019). The instrument is thus intended to influence a judge’s decision to order a PSI for a given criminal defendant, with the presumption that information contained within PSIs will in turn influence a judge’s decision to assign an alternate sentence. Currently, PSI-ordering rates in Pennsylvania vary substantially county-to-county, as do the contents of the reports; one of the expected outcomes of the tool’s adoption was thus to minimize county-level disparities in how often, and for which kinds of defendants, judges choose to order a PSI (Figure 1).

The Pennsylvania Sentencing Commission – a legislative agency that advances “fairer and more uniform decisions at sentencing, resentencing, and parole” – was tasked with fulfilling a 2010 state legislative mandate to develop the instrument (Pennsylvania Commission on Sentencing, 2022). However, the Commission’s members soon found themselves embroiled in controversy. From 2017–2019, the Commission received over 100 overwhelmingly negative public testimonies about the tool from sources including AI Now, the ACLU, high-profile academics, and local community organizations. Critics argued that the “racist tool” (ACLU of Pennsylvania, 2019) could “perpetuate the racial biases and stigmas inherent in our criminal legal system” (Coalition to Abolish Death by Incarceration, 2019). The instrument also met intense criticism from within the criminal legal system, particularly from probation officers, who argued that the Sentence Risk Assessment Instrument was “an unnecessary burden in time, effort, and resources” and would increase the workload of “already overwhelmed” county probation departments (County Chief Adult Probation and Parole Officers Association of Pennsylvania, 2019).

In informal interviews, Commission staff explained that they were legally required to implement the legislative mandate despite these criticisms, lamenting that “from the start … there has been no public support for the development and use of risk assessment at sentencing” (Pennsylvania Commission on Sentencing, 2019). Commission staff, to their credit, engaged with the public through a transparent and iterative process of development, removing piece by piece the most controversial parts of the instrument and seeking further public comment each time. For instance, an earlier version of the instrument showed judges not only an individual’s risk score but also a detailed risk distribution, indicating to the judge exactly where the defendant’s numerical score falls relative to other individuals. The final version of the tool only shows judges a small text box with the words “Additional Information” (if the defendant is low- or high-risk), or “NA” (if the defendant is moderate-risk). The Commission also changed its outcome variable from rearrest to reconviction in response to public testimonies, which argued that arrest is not only a poor predictor of actual crime but also racially correlated due to racial profiling by police (Sassaman, 2018). The core concerns of the public and probation officers, however, went unaddressed – the tool was still implemented, and no additional resources were allocated to assist probation departments with the anticipated increase in ordered reports.

As part of the state’s Evidence-Based Practices Strategic Plan, the Commission solicited an external review of the tool by Carnegie Mellon University researchers in 2019. This audit focused on technical benchmarks of validity, accuracy, and fairness, and made several recommendations, including discarding the high risk category due to low accuracy; removing gender as a predictive factor; raising the high-risk category cutoff to increase its accuracy; and not deploying the violent crime risk scale component of the instrument due to an unacceptable level of false positives (Becerril et al., 2019). The Commission voted to follow the latter two recommendations. Notably, the audit does not consult relevant stakeholders or mention the tool’s interactional effects with judicial discretion or other social factors, instead including projections (e.g., Figure 1) that assume complete uptake of the tool. Later in 2019, the Commission voted to adopt the tool, and the Sentence Risk Assessment Instrument was formally rolled out in July 2020.

The most common reason that judges did not use the tool was that they simply did not find it useful. This is due in part to the work of activists, lawyers, and academics who, over years of public testimony hearings, successfully pressured the Pennsylvania Sentencing Commission to remove the most controversial parts of the instrument, including directly showing judges risk scores and detailed risk distributions. The implemented version of the tool recommends ordering additional information about low- and high-risk defendants, in keeping with the original goal of helping judges identify candidates for alternative sentences. However, none of the judges I spoke with were looking to change their PSI-ordering behavior. Judges reported either ordering PSIs for all trial cases, ordering PSIs for more serious trial cases, or almost never ordering PSIs; this behavior reflected how useful judges found the PSIs themselves, whose contents vary by county. Nearly half of the judges I talked to also did not find the contents of PSIs helpful because in many counties, including the state’s most populous Philadelphia and Allegheny counties, the reports contain information judges can get simply by talking to the defendant. In other words, the tool intervenes on a factor – PSI-ordering behavior – that judges are uninterested in changing, and falsely assumes that successfully influencing PSI-ordering behavior will in turn influence sentencing decisions.

Five judges explicitly used the words “useless” or “worthless” (sometimes with an expletive) to describe the Sentence Risk Assessment Tool. Over half of the judges also stated that they would have preferred to see different information presented to them at sentencing time, including the causal impacts of different sentencing practices on recidivism, a risk and needs responsivity risk assessment, information about how the risk assessment was derived (“Show me the math”), and information about risk categories (“It would be better if they said high, moderate, or low, to be honest”; one judge said they would only want to see information about low-risk defendants, while another said they would only want to see information about high-risk defendants).

Another common reason that judges ignored the tool, which often overlapped with judges’ perceptions of the tool’s uselessness, was a lack of information about what the tool did or what form its recommendation appeared on. As one judge put it, “I never knew where that information was going to be provided for me. Was it going to come in an email? A news blog? A winter weather alert? I had no idea.” Several judges explicitly asked me to show them where on the sentence guideline form that judges routinely receive at sentencing time – “the world’s least user-friendly form” – the recommendation appears. Another judge called their supervising judge during my interview because they did not believe me that a sentencing risk assessment instrument was in use in their county. With two exceptions, every judge I spoke with revealed some degree of misconception about the tool during the course of my interview, such as the claim that the tool shows judges risk scores (it does not), that the tool applies to DUI cases (it does not), and that the judge has to do something in order to generate the risk assessment (they do not; it is automatically generated and appears on the sentence guideline form). Several interviewed judges were ashamed about being on the record about their lack of awareness of the tool, while others used their lack of knowledge about the tool as a reason to decline participation in my study.

Nearly all judges had low literacy of the tool, despite the Commission’s claim that, effective January 1, 2020, it would “conduct a six-month training and orientation for judges and practitioners related to the use of the Sentence Risk Assessment Instrument, the purpose of the recommendation, and the type of information recommended” (Pennsylvania Commission on Sentencing, 2019). Many judges and probation officers remarked that the tool – and how to use it – had been poorly publicized. In personal conversations, Commission staff explained that their information campaign had been derailed by the start of the pandemic coinciding with the roll-out of the tool.

More broadly, however, my findings indicate systemic problems with how information is disseminated to judges in Pennsylvania. In one particularly revealing moment, a judge told me that they were attending a virtual Continuing Judicial Education session over video call in the background of their computer – during our interview. The problem of judicial education was echoed to me by a chief probation officer, who lamented that even with respect to the risk assessment already included in PSIs in their county, the probation department had not done much in the way of educating judges about how to interpret risk assessment information, adding that many judges “didn’t really understand how it applies to the work that they do” and that this was likely the case statewide.¹¹1This issue extends beyond Pennsylvania; low literacy about risk assessment among judges has also been documented in Virginia (Garrett and Monahan, 2020).

In addition to misinformation and perceptions of uselessness, skepticism or concern about risk assessment instruments more broadly was often a complementary reason that judges cited for ignoring the Sentence Risk Assessment Instrument, though it was typically a secondary issue. These concerns fell roughly into three categories.

The most common concern, which roughly half of judges expressed, was that the tool ignored a defendant’s humanity. Notably, this was a central issue raised by CADBI members in their feedback on my study design; one formerly incarcerated individual worried that the new risk assessment tool would make judges more likely to ignore the humanity and personal circumstances of the people they sentenced. “Each individual has a history that brought them to this space,” this consultant told me. “There must be individualization.” Judges echoed this point, raising concern about “having a formula that takes away my ability to see the humanity of the people in front of me”; another judge argued that “cookie cutter justice doesn’t work” and that risk assessment was “merely labeling and boxing”; a third said, “I don’t know how you can reduce all of the human factors that go into, you know, sentencing or making a bond decision and, and put it into a number, you know, I just, I just think that there are a world of human factors that need to be considered.” These judges emphasized the crucial role that individual narratives and personal context played in their sentencing decisions. Most judges also indicated that they did not assign central importance to aggregated recidivism risk in their sentencing decisions (with the exception of recidivism risk for sex crimes). Rather, they were interested in the personal trajectories of criminal defendants, particularly escalation toward violent behavior; whether a defendant was employed; and drug use.

Another common concern judges raised was about the tool’s bias, especially racial bias. One judge, who identified as Black, was critical of the discriminatory potential of the tool: “Who’s making the determinations? Who’s interpreting the statistics? You can say anything with statistics.” Another judge noted the third-party audit’s finding that the tool’s high-risk category was less accurate than the low-risk category, commenting that this could be “prejudicial to certain minority groups because there was an historically higher arrest rate, possibly related to things like race rather than actual criminal activity.” Judges were concerned about other biases as well – a judge who was otherwise an advocate of risk assessment tools claimed that the tool was biased in favor of sex offenders (a claim that is not factually accurate), while two others commented that age was an unfair indicator of recidivism because minorities are statistically more likely to be stopped by police at a younger age. This concern about bias was not unanimously shared, however; other judges acknowledged that the tool had biases but maintained that these were still better than human biases: “You can never take all biases out. You can never take out – there’s biases, people get arrested – what’s in it, but you can continue to work on the tools to try to make them as fair as possible. But it’s better than individuals.” One judge even claimed that “[risk assessment tools] have been deliberately distorted as being racist, as being not accurate, as being using wrong statistics and things like that.”

The third most common concern was that the tool was worse than the discretion of experienced judges. A common refrain from judges was that younger, less experienced judges might get more benefit from the risk assessment tool, but that for more experienced judges, such an instrument was unnecessary. There was also a general sentiment from judges that personal discretion was a centrally-defining feature of what it means to be a judge; one judge with over a decade of experience firmly announced in the first 10 seconds of our conversation that they were “elected to be a judge, not a robot.” Nine judges independently brought up that judges “don’t want to be told by anybody what to do;” however, those same judges did not view themselves as being in this category. Seven judges said their own sentencing practice was better than other judges, describing their sentencing using adjectives like “different,” “atypical,” or (pleasantly) “shocking” to defendants. Several judges were critical of any efforts to limit their discretion, including sentencing guidelines, which are supposed to standardize sentence lengths based on an individual’s prior record score and the gravity of their current offense. One judge aptly summarized this particular concern: “[The legislature] is trying to give us more narrow options on what we can do. And, and I don’t like that, because I think that there’s a reason that we’re up there – we’re up there because supposedly we’ve demonstrated some ability to think more broadly about the whole system and to make a better decision than just something that’s electronically generated. You know if you’re going to do it all based on a computer program, then you don’t need me out there.”

Importantly, however, judges’ skepticism about risk assessment instruments should not be conflated with skepticism toward data-driven strategies in criminal justice more broadly. As already mentioned, many judges reported wanting access to more data at sentencing time – just not the kind of information provided by this risk tool. Moreover, most judges did in fact acknowledge the importance of consistency in sentencing and, with few exceptions, reported complying with sentencing guidelines. With the exception of two judges, the skeptical claims above were regularly expressed alongside pro-data and pro-science stances at other points within the same interview. One judge, who had expressed concerns about the tool’s racial bias earlier in our interview, maintained that “I’m a believer in science. This [risk assessment] is science, so we need to use it.”

Several judges worried that the Sentence Risk Assessment Instrument could have unexpected downstream consequences, were it to be used. The Commission “expressly disavows the use of the sentence risk assessment instrument to increase punishment” (Pennsylvania Commission on Sentencing, 2019). However, as one judge and public testimonies pointed out, judges can still infer risk levels from the ‘Additional Information’ label, and empirical evidence from other states suggests that judges are more likely to use risk information to detain individuals longer (Human Rights Watch, 2017). “People know it’s called the risk tool. If it’s ‘Additional Information’, there may be some concern about how dangerous the defendant is,” a judge noted. Moreover, if judges followed the tool’s recommendation to order PSIs for low-risk defendants – who often have minor sentences – then the tool could have the unintended effect of detaining these defendants longer pre-trial, since ordering a PSI can take 60 days or longer, depending on the county. Another judge remarked, “I’m not letting them [the defendant] sit 8 more weeks in jail because some computer program said so.”

Speaking about unintended impacts in other parts of the criminal legal system, two probation officers also shared worries about the tool creating unnecessary – and invisible – labor for their departments, which are tasked with generating the risk and needs responsivity assessments that go into the PSIs in some counties. One of these officers said they feared they were going to get “a flood of cases” where judges were ordering PSIs, but that “thankfully that has not happened” because they did not have the resources to handle such a surge. They said they would like to see the system someday permit having such an assessment done for every defendant, but that this would “require a lot of resources, a lot of resources.” The second officer raised the concern that the tool, if widely used, would “significantly slow down” the already-backlogged sentencing process, which they said could cause individuals to spend even more time awaiting trial in jail. To this probation officer, the risk assessment instrument was just “another unfunded mandate, the burden of which was going to fall on county probation.”

A minority of judges I spoke with were knowledgeable, vocal advocates of other risk assessment instruments and the Pennsylvania Sentencing Commission’s other projects. Even among these four judges, however, only one claimed to be regularly consulting the Commission’s tool, with the caveat that it had never changed their PSI-ordering behavior. Judges in this group were either advocates of using risk assessment at other stages of the criminal legal pipeline, such as at preliminary arraignment, or were serving in counties where the Ohio Risk Assessment Instrument (ORAS), a significantly more detailed risk and needs responsivity risk assessment, is conducted by the probation department and is already a routine part of the PSIs that judges receive. One self-described “cheerleader” for risk assessment instruments explained: “I like the [Commission’s] tool, I just like our tool [the ORAS] better – it’s shinier and faster.” This was the position of two of the probation officers I spoke with as well.

In sum, although nearly all of the judges I interviewed reported ignoring the Sentence Risk Assessment Instrument, their reasons for this varied. This suggests a nuanced explanation for aversion to algorithmic systems in the criminal legal system that is neglected in existing discussions that are centered largely around lack of confidence in technology and fears of deskilling and surveillance. The rest of this paper discusses the implications of this finding for scholarship on algorithmic resistance and risk assessment instruments.

As one judge aptly summarized it, “there was a lot of resistance to the tool” – not only from the community but also from public defenders, probation officers, criminal attorneys and, as I have shown, judges themselves. A standard algorithm aversion explanation for this could be individuals’ lack of confidence in the tools (Dietvorst et al., 2015). Distrust is, no doubt, an important part of the story, particularly with respect to public resistance to the instrument. But does lack of confidence explain the resistance from judges? As I discussed in §4.3, some judges did cite lack of confidence in the instrument’s predictions as one reason for not wanting to use them; however, this was not the primary reason but rather something that came up later in the interview once I started probing about their other concerns about the instrument. Moreover, a weak majority of judges I spoke with were supportive of using risk assessment instruments in some capacity – if not at sentencing, then at some other stage of the criminal legal pipeline – and often reported wanting more empirically-derived information to assist decision-making (§4.1). In general, I observed among most judges a strong pro-data mentality. In short, lack of confidence is a simplified, algorithm-centered explanation that does not provide an adequate explanation of this real-world case – such as why judges who are self-avowed “cheerleaders” of other risk assessment instruments used in their counties are still critical of the Commission’s tool.

Brayne and Christin provide another, sociological explanation: judges may be engaging in behavior like foot-dragging due to fears of deskilling and managerial surveillance (Brayne and Christin, 2020). In their studies of predictive policing and pretrial risk assessment, Brayne and Christin found resistance to algorithms to be strongest in cases of function creep, where algorithmic tools served the added purpose of increasing managerial control and surveiling bureaucrats’ productivity. While this sort of function creep is, for now, absent in the present case, I did see some evidence of automation anxieties and fears of deskilling. Some judges – particularly in Philadelphia – expressed antagonism toward any mandates that were intended to limit judicial discretion, including sentencing guidelines. Almost all the judges described their own discretion as a strength, not a weakness, though they were also often critical of other judges practicing the wrong kind of discretion, and reductions in judicial discretion were sometimes perceived negatively – recall the judge who compared using risk assessment instruments to being “a robot” (§4.3). Despite their concerns, however, most judges still expressed agreement with the premise of data-driven sentencing, and few opposed the use of some discretion-limiting measures, such as sentencing guidelines. This makes it unlikely that judges’ resistance to the risk assessment instrument is entirely “fueled by fears of deskilling and heightened managerial surveillance” (Brayne and Christin, 2020).³³3Brayne and Christin also propose the thesis that predictive technologies displace discretion to less visible areas within organizations. I found some unexpected support for this thesis in counties where an additional risk and needs risk assessment instrument is included in PSIs. Judges revealed to me that probation officers have an enormous amount of discretion in how they prepare such reports; in one county, PSIs even include concrete recommendations for what an individual’s sentence should be – a determination made by the probation officer preparing the report.

In this case, a more adequate explanation for why judges ignore the tool has to do with the organizational influences that led to the tool’s development, policies about the contents of PSIs, and problems in how information is disseminated to judges. One probation officer described the Pennsylvania Sentencing Commission as “trying to make certain groups happy” – that is, the public, the legislature, the judges, and probation officers. One of the outcomes of this negotiation process was the selection of a less-publicly-controversial locus of intervention for the tool: the decision to order a PSI.

However, judges did not report PSIs influencing their sentencing decisions except in very unusual situations, such as where a criminal record is stale.⁴⁴4In such cases, they may use the background information to go below sentencing guidelines. Typically, judges said that PSIs are not very helpful and never “dramatically changed [their] mind” about a sentence; this was the case even for PSIs that contained the more detailed risk assessment. This means that the final version of the tool is, at best, useless for judges; not using the tool was largely a response to this fact, complemented by widespread low literacy about the tool. At worst, judges’ adherence to the tool’s recommendations could have produced ‘ghost work’ (Gray and Suri, 2019) for probation departments and detained individuals longer pre-trial (§4.4). But activists still see the final weakened tool as a win. As Hannah Sassaman, a Philadelphia-based community organizer, told news outlets the day after the tool’s adoption, “the tool that the Commission instituted yesterday was massively changed over the past few years from one that actively centered racist factors in guessing the future of a sentenced person, to one that will be considerably less damaging” (Gross, 2019).

A defender of evidence-based sentencing could make the case that, had the Sentence Risk Assessment Instrument (or the legislative mandate it was built to satisfy) been designed differently – and had the public been more receptive to its development – then perhaps judges would not have been so resistant to it, and perhaps more people would have gotten alternatives to prison sentences as a result. But empirical research on risk assessment instruments used in other states – mostly for pretrial detention decisions – suggests that the tools’ impacts have been minimal, unfairly distributed, and have tended to wash out over time (Stevenson, 2018; Sloan et al., 2018; Albright, 2019; Garrett and Monahan, 2020; Stevenson and Doleac, 2021). This has been the trend even for tools with greater uptake and more significant loci of intervention than the decision to order a PSI. Empirical research also suggests that risk assessment instruments introduce an element of arbitrariness to decision-making, such as sharp differences in sentencing decisions for individuals with risk scores that fall near the low-risk category cutoff (Stevenson and Doleac, 2021). As economist and legal scholar Megan Stevenson starkly puts it, “Somehow, criminal justice risk assessment has gained the near-universal reputation of being an evidence-based practice despite the fact that there is virtually no research showing that it has been effective” (Stevenson, 2018).

This research thus contributes another case study to an alternative, empirically-informed argument for abolishing recidivism risk assessment instruments: in practice, these algorithm-centric reforms have no significant impacts on sentencing, are resource-intensive to develop and implement (in a context in which resources are highly limited), and merely pay lip service to addressing the crisis of mass incarceration. Grassroots organizations such as CADBI have been promoting low-tech liberatory policy changes for decades, including abolishing cash bail, releasing elderly populations from prison, and reinvesting money in schools and communities. Unlike risk assessment instruments, such measures do not rely on individual judges’ alignment with policy goals and have robust empirical support for reducing prison populations.⁵⁵5See (Zhou et al., 2021) and (Note, 2018) for empirical discussions of bail reform.