Predictive Prosecution: Part 2
Read more from Critical Perspectives in Criminal Justice.
Andrew Guthrie Ferguson is Professor of Law at the David A. Clarke School of Law at the University of the District of Columbia. He has been featured extensively in law reviews and is the author of The Rise of Big Data Policing: Surveillance, Race, and the Future of Law Enforcement, which was honored as the best scholarly book on law and legal studies by the Association of American Publishers in 2018. Professor Ferguson is also Senior Visiting Fellow at the Harvard Law School Criminal Justice Policy Program and Policing Data Fellow at the NYU Law School’s Policing Project.
Note: This is Part 2 of a two-part series on predictive prosecution. This essay explores the impacts and future of predictive prosecution. Part 1 has more information on the history and details of predictive prosecution and was published in our last issue, here.
Predictive prosecution — data-driven policies that shape prosecution strategies — exists in an experimental phase. This Essay seeks to raise preliminary questions about an obviously nascent experiment. But, the questions are real, and will need to be answered soon. The hope of this brief Essay is to set forth the possible impacts, raise questions, and plan for the future of predictive prosecution.
Preliminary Questions about Predictive Prosecution
This section examines one big question surrounding predictive prosecution. How does predictive prosecution impact prosecutorial decision-making? Due to the constraints of the format, the ideas discussed are initial impressions, not full explorations of complex and important topics.
Predictive prosecution offers potential benefits in terms of prioritization, efficiency, and more informed judgments. Prosecutors must make difficult decisions every day, and more information might provide for better choices. In today’s legal system, prosecutors possess almost unlimited discretion (Podgor). Prosecutors decide whom to prosecute (Wayte v. United States). Prosecutors decide how to charge and how to structure plea bargains (Covey; Litman; Podgor). And prosecutors decide recommendations for sentences (Griffin). Adding information from sources such as the predictive policing “Heat List” (see Part I) or organically developed intelligence does not present any direct ethical or constitutional concerns.
If used to identify and proactively target actual crime drivers in a community, a predictive prosecution system could well provide an overall benefit to society. If resources could be redirected toward incapacitating more serious offenders (through bail, charging, and sentencing decisions), while concomitantly incapacitating fewer, less serious offenders, such a process could mean fewer overall people in jail. Such a system might also be more efficient, redirecting scarce prosecution resources. Of course, the current system of mass incarceration that has developed over the last several decades has not lacked for efficiencies in prosecuting and convicting defendants (Alexander; Chettiar). Mandatory minimums, harsh drug sentences, plea bargains, and other processing efficiencies have created an overly efficient process for incarcerating millions of people (Traum). But, the web of people caught up in this system has been overbroad, lacking a commitment to prioritize those most dangerous to society (Pfaff). Millions of nonviolent offenders, millions of misdemeanants, and millions of low-level figures in the drug world are serving significant time in jail (Natapoff). Individually, those persons might not be the chosen targets of our criminal justice resources, but systemically prosecutors have had few mechanisms to evaluate or rank relative danger or risk to society (Neyfakh).
Predictive prosecution offers a potential smart-on-crime counterweight to the tough-on-crime practices of over-incarceration. In fact, taken one step further, if prosecutors only sought to target those predicted to be of high risk of committing crime, then a huge majority of people would see reduced bail, better pleas, and more lenient sentencing. Such prioritization might significantly reduce pretrial detention costs, long term sentencing costs, and overall criminal justice costs.
The danger, of course, is that predictive prosecution might not reduce prosecution levels, but might, in fact, bring more people into the criminal justice system. Two obvious concerns arise within the Enforcer Model (see Part I). First, in the Enforcer Model, individuals are being linked to criminal activity by proxies for criminal activity. A gang member who has a friend who was shot may be added to the system because, statistically, the associates of dead gang members are more likely to themselves be involved in gun violence. The “two degrees of separation” analysis may both be accurate and yet overbroad when it comes to prosecutorial decisions (Papachristos & Kirk). The particular individual might not have done anything but be a victim of violence, or might remain a small time criminal actor. Further, that particular individual might be summoned to a call-in by a prosecutor and threatened that he may face harsher detention, charging, and sentencing decisions should he get in trouble in the future. So, that individual is in the first instance added to a prosecution list without criminal activity of his own, and in the second instance faced with the potential for a harsher criminal justice outcome because of that designation.
Similarly, in the “Investigative Model,” individuals are being targeted because they have been identified as the primary targets for removal (Papachristos & Kirk). The key, of course, is the process by which people are targeted. If limited to only those individuals with multiple convictions for violence, this incapacitation approach can be defended. Using minor crimes to incapacitate major criminal actors is aggressive, but defensible. However, if other factors such as a lack of cooperation with police, suspected but unproven violence, or low-level, non-violent crimes become the justification for being a target, then justification for aggressive incapacitation weakens. Using minor crimes to incapacitate minor criminal actors undercuts the value of targeting only the serious offenders.
Put another way, because the targeting mechanism of identifying the primary targets rests with the prosecution (in collaboration with police), and because there is no system to challenge or correct a targeting error, a risk arises about the data populating this system. Prosecutorial decision-making runs a real risk of being infected by bad data in these systems (Herring v. United States). Personal bias could influence who becomes a target. Political or economic pressure could shape the types of crimes addressed.
Even more generally, any data-driven system runs into concerns with data quality. Data can be inaccurate (M.D.M. Fan; Navid; Steinbock; Whalley). Data can be biased (Taslitz (a)). Data can reify the existing socio-economic inequalities in the criminal justice system (Llenas). Data can also be overwhelming, with little practical or technological checks on quality or accuracy (Mitnick). Yet, every day police and prosecutors collect more data on individuals, and systems are being designed to become more reliant on this data collection (Mitnick). In prior articles, I have laid out the concern of data error in the criminal justice system (Ferguson (a); Logan & Ferguson). From big data to small data — all data systems generate error. Human error, collection error, processing error, analytical error, application error, or sharing error all exist and cannot be minimized when this same data is used to determine human liberty. If prosecutors’ discretionary power involving bail, charging, and sentencing is informed by erroneous or merely poorly correlated data, then real injustice could occur.
The issue is not that prosecutors cannot rely on this data within their existing professional and ethical mandate, but whether they should. The subsequent part of this Essay will address how prosecutors should minimize the real risk of using bad or biased data.
Principles for Predictive Prosecution
Predictive technologies are not new to the criminal justice system (Harcourt (a)). Since the 1920s the lure of predictive insights has led the criminal justice system to try to forecast the future. Predictors for recidivism (Hamilton (a); Sidhu), pretrial detention (Baradarna & McIntyre; Williams), sex offenders (Hamilton (b); Janus & Prentky), juveniles (Fagan & Guggenheim; Roberts & Bender), and a host of actuarial solutions have been proposed (Ferguson (b)). Predictive policing, and now predictive prosecution, fit that pattern.
For almost as long as their creation, the critiques of these predictive technologies have identified the same concerns over and over again. Predictive correlations become mistaken for causation (Underwood), validation studies fail to validate (Grove & Meehl; Harcourt (b)), analytical mistakes infect the legitimacy of the conclusions, and error — small and systemic — pervades all data-driven systems. The concept of predictive prosecution provides the same promise and potential critique. Yet, because of the prosecutor’s special role in the criminal justice system, there may be some cause for optimism. If designed carefully, a predictive prosecution system might provide an accountability mechanism to police data error and moderate blind reliance on data-driven predictions.
While a full descriptive framework is beyond the scope of this Essay, any predictive prosecution system must be built on four related principles: ownership, accuracy, transparency, and fairness. These principles are explained below, with recognition that significant additional discussion and debate is needed before the adoption of any predictive prosecution program.
First, prosecutors must accept ownership of the data underlying predictive prosecution systems. If bail determinations, charging decisions, or sentencing is impacted at all by data correlations, then that underlying data must be trustworthy enough to withstand scrutiny of judges inquiring about the bases of the lists or reasons for the decisions. Whether from a predictive policing system or organically developed by prosecutors, once used in court, prosecutors must take responsibility for the data. Integrating police and prosecutorial systems, even informally, means that prosecutors must take on a data management duty that they previously did not have to accept.
Second, and relatedly, prosecutors must ensure the accuracy of the data. In adopting theories of intelligence collection to augment traditional prosecution roles, prosecutors should also examine how intelligence agencies test and assess the data collected. In the national security context, thousands of intelligence analysts work for the United States government because of a healthy distrust of the raw intelligence coming in from sources (FBI). Intricate internal systems exist to evaluate the reliability of data, recognizing that actionable data for targeting cannot be relied upon without critical analysis. So, too, with intelligence-driven prosecution, prosecutors must establish systems to assess the value of the data coming in through community sources, detectives, social media, and other sources.
In addition, this push for accuracy means developing systems to audit existing data-collection systems, including mechanisms for removal and alteration of bad or outdated data. The danger of a high-volume data collection enterprise is that it is much easier to simply collect everything, accurate or not (Lapp). Going back to correct errors involves time, money, and technological sophistication (Westland). But, without such checks, the data becomes unworthy of use in criminal courts. Direct connection to criminality, not mere correlation, should be required when an individual’s liberty is being decided. Processes must be created to ensure that personal bias or corruption does not distort the targeting. Further, the data collection and analysis must be scrutinized for implicit or explicit bias (Taslitz (b); Gove). Disproportionate minority contacts, high incarceration rates, and harsh sentencing have been clearly demonstrated throughout the criminal justice system (Sterling). Any data-driven system built on top of that inequality will likely reify the inequality unless explicit steps are taken to address the issue.
Third, any data system must be transparent (Zarsky). This involves a two-fold transparency, both to the prosecutor using the data and the community legitimizing the use of the data. Prosecutors are lawyers trained in law, not technology. In large offices the data will be compiled by colleagues and assistants. In systems of “extreme collaboration,” data will also be compiled by police. So, mechanisms must be created so that prosecutors can understand the source of the data. Prosecutors need to be able to not only trust, but understand and defend the data. Arguments cannot be along the lines of “judge, I am asking for a no bond bail determination because the pre-printed form told me to ask for it,” but because of particularized, verifiable facts that can be obtained through a data-driven system. Arguments cannot be “judge, the defendant is on the SSL, so we ask that he be held,” but based on the actual underlying facts that might have led some individual to be on that list. Prosecutorial transparency requires understanding why individuals have been chosen to be marked by predictive technologies. This understanding may also require knowledge of the provenance of the data, the currency of the data, and the reliability of the data.
The other aspect of transparency focuses on community acceptance of predictive prosecution outcomes. The Orwellian nature of government lists of predicted targets rightly causes suspicion. Any predictive prosecution system needs to be able to explain, in a relatively open and clear way, how people are placed on predictive lists, and why the criteria is legitimate. This presents a challenge in that most prosecution or police methods also need to be relatively opaque in order to avoid undermining ongoing investigations. This balance between transparency and operational secrecy presents real tensions. But, as the creation of custom notification letters demonstrate, prosecutors can develop a process to show and explain why someone is targeted. Custom notification letters are “customized” and include the target’s specific criminal history and risk factors. The reasons for the targeting are thus particularized and individualized and open for inspection. Similarly, in call-ins, prosecutors can explain in specific detail why the particular targets have been contacted. This process provides transparency and legitimacy to the process (albeit after the fact).
This type of customization also needs to be applied systemically. Prosecutors need to be able to explain why certain communities have been targeted, and how they have attempted to avoid class- or race- based impacts. Using crime mapping, visual displays of historic criminal activity, and other accessible media, the argument can be made for why certain areas were chosen and not others. Discriminatory impacts need to be monitored and studied. Communities may accept a higher prevalence of prosecutorial interest in an area, but it must be explained and defended in a transparent manner.
Finally, predictive prosecution systems must build in mechanisms to ensure fair process. An emphasis on fairness must address concerns that citizens might hold in being targeted by predictive techniques. A process will need to be developed to challenge a target designation on a police list (Hu). A method to account for possible racial or class discrimination will need to be created (Ajunwa et al.). Clear procedures to use and validate the predictive target list need to be developed. And, a general emphasis on procedural justice must continue. Due to the influence of some of the academics who provided the early inspiration for the Chicago projects, procedural justice has been a key organizing principle behind the intervention strategy, but such an emphasis must continue to be prioritized (Meares).
References available in full pdf.