1. Background

Legal license to treat an individual in certain ways—subjecting them to special scrutiny, detaining them, or ruling that they are liable to penalties—often depends on our having a sufficiently high degree of confidence given the available evidence that such treatment would be appropriate. When making these determinations, individual police officers or judges sometimes rely exclusively on their own observations and evaluation of the available information; sometimes they consult or defer to the judgment of an expert. Someone attentive to how vulnerable human decision-making is to cognitive biases of all sorts—not to mention the distortions introduced by prejudice—might hope to improve these decisions by basing them on algorithmic predictions.Footnote ¹ Rather than relying on a single agent’s personal assessments of whether a particular suspect is likely to reoffend, for example, we might hope to leverage historical trends in rearrest or reconviction data to yield some objective measures, insulated from individual irrationality or animus.

This, in its most optimistic frame, was the driving promise and aim of the risk-assessment tools first developed to aid parole decisions in Chicago in 1933. The original model leveraged factors like marital status, behavioral infractions within the detention facility, prior arrest record, and work history to sort inmates into nine rough cohorts, and furnished parole boards with the relative frequency of rearrest among past members of an offender’s cohort as an indicator of the probability that the inmate would reoffend (Burgess Reference Burgess1936, 499). Since then, statistical Risk and Needs Assessment (RNA) tools have been refined and proliferated; they are now used in the majority of jurisdictions in the United States to guide pretrial decisions relating to whether (and how high) to set bail, as well as posttrial determinations concerning whether to divert a defendant from incarceration and when to consider them for parole.Footnote ²

Most of these tools employ straightforward statistical analysis on historical arrest databases, seeking to isolate the strongest correlations between a relatively sparse set of recorded variables and a property representing the target outcome (e.g., failure to appear, another arrest, or arrest for violent offense). There is some variation in the variables used: third generation risk assessment measures improve on the original second generation modelsFootnote ³ by using not only static variables—properties that do not change over time, such as age at first arrest, having a prior conviction, sex, etc.—but also dynamic variables (e.g., years since last offense, employment status, present substance abuse) which are responsive to the subject’s current behavior, and can reflect reduced (or increased) risk over time. Simplifying a bit, these tools are ultimately algorithms taking the variable values as inputs, assigning them weights, and outputting an estimate of how often someone with those features in the database ends up with the target outcome.

There is now a new wave of tools with a somewhat different structure, more aptly described as applications of artificial intelligence. They leverage a wide array of information in vast databases to train a model to recognize patterns in an existing dataset in order to predict the outcome value for a new entry. This method allows the trained model to discover previously unnoticed correlations between the target outcomes and properties in the dataset. The hazard is that the correlations might be artifacts of the particular dataset, rather than robust connections in the underlying phenomena.Footnote ⁴ As these machine-learning techniques improve, programs using them have become increasingly effective at a wide variety of recognition and classification tasks, and have been pressed into service for a range of prediction tasks, too.Footnote ⁵ The allure of these tools is that they offer a chance not just to make an informed guess about how often an outcome will occur in some set, but to identify and intervene in a case before the predicted outcome occurs or becomes acute: removing precancerous tumors, connecting struggling students with extra resources, or, in the case of criminal justice, preventing a predicted victimization.

Buoyed by enthusiasm for data-driven policing and sentencing, both sorts of tools have made their way into law-enforcement at several points. To name just a few examples, PredPol and HunchLab are used by police departments across the United States to identify hotspots for property crime, assault, and auto theft. Several states use Palantir’s data-analysis program Gotham to leverage information aggregated from service and arrest databases in order to guide the allocation of police resources and aid in suspect identification and profiling. These include at least the Chicago Police Department’s Strategic Subjects Initiative (SSI), LAPD’s Los Angeles Strategic Extraction and Restoration (LASER) program, and the Northern California Regional Intelligence Center. Police departments in New Orleans and New York City had similar contracts.Footnote ⁶ On the postarrest side of things, the Correctional Offender Management for Profiling Alternative Sanctions (COMPAS) program, developed in 2002, is used both to make pretrial determinations about bail and posttrial determinations about sentencing and parole throughout Michigan, Wyoming, Wisconsin, California, and in an ever-increasing number of counties in other states.Footnote ⁷

Not everyone welcomes the increased use of algorithmic prediction tools. Attorney General Eric Holder, for instance, cautioned that:

Some of the obvious ethical concerns about using algorithms in criminal justice—bias in error rates, data looping, redundant encoding, etc.—have already been the subject of significant academic and media attention. COMPAS in particular has drawn substantial criticism for having racially biased error rates.Footnote ⁸ There are also a handful of epistemic concerns that have been raised about the probative value of risk scores, including worries that these systems base predictions on spurious correlations, are lacking in explanatory value, and give fixed datapoints from a person’s past too much weight to be epistemically reliable in predicting whether they will reoffend.Footnote ⁹ In this paper, I will set all of these aside in order to focus on a third type of concern visible in Holder’s remarks: whether pursuing criminal justice with these tools is consistent with treating the defendant as an individual.

The reason for this narrow focus is straightforward: while they are serious, the problems mentioned above are mostly problems with using algorithms badly. But no matter how we clean, debias, or supplement, all the tools in question trade in actuarial inference. The score assigned to an individual does not reflect any deep insight into his “true nature”; it reports the frequency of a type of outcome among people in the database who are similar to him with respect to the values of the predictor variables. Setting aside programs focused on predicting locations of crimes rather than individuals, the basis for risk predictions made by these algorithmic tools is, ultimately, observations about the behavior of other people. The reasoning structure is actuarial in that it moves from the conjunction of the subject has feature G and the relative frequency of feature F among others with G is x to confidence of approximately x in the subject has feature F. In a slightly different context—addressing the use of statistical or probabilistic evidence to establish liability in civil trials or settle sentencing questions in criminal trials—several legal theorists, philosophers, and judges have objected that inferences of this form functionally make it a “crime to belong to a reference class,”Footnote ¹⁰ violating the right to be “treated as an individual.” We might reasonably ask whether this right also forbids the use of any of the algorithmic tools mentioned above. If so, this would be a problem not just with using the tools badly, but with using them at all. But it isn’t immediately obvious what the right to be treated as an individual forbids, because it isn’t clear what it is a right to, exactly.

Rather than trace the constitutional grounds or legal interpretation of this right, my project in this paper is to explore its core: what moral interests might it protect, and are those interests threatened by relying on the outputs of algorithmic methods in determinations of probable cause, guilt, or sentencing? After exploring a few different interpretations of the right (in section 2), I ultimately propose understanding it as protecting agents’ claims to a fair distribution of the burdens and benefits of the rule of law (in section 3). What it forbids is not the use of probabilistic information or statistical methods but treating wrongdoing by some as justification for imposing extra costs on others; it demands that we respect the separateness of persons. This has significant implications for the administration of criminal justice (explored in section 4): it permits the use of predictive tools in principle, but only if the predictors are transparent, subject to agents’ deliberate control, and unavoidable burdens are fairly distributed or outweighed by benefits to the very same individuals. In section 5, I explain that since this bars relying on indices of distributive injustice or unchosen properties to determine whom to subject to extra costs associated with criminal justice, it precludes the most common applications of algorithmic tools for bail determinations or the regulation of street crime. However, they might be justifiable and effective in combatting white-collar crime.

I will, where possible, treat all the applications together, but the implications vary depending on the context in which an algorithmic prediction tool is used, so it will be helpful to have a sense of the variety of uses to which they are put. So, very roughly: some of the tools mentioned are used to guide the application of prearrest scrutiny, establishing probable cause to subject someone to additional search or surveillance, arrest, or detention. At the next stage, a risk assessment might be offered to support bringing charges, or as evidence about whether the subject will either reoffend or fail to appear if released on bail before trial. In theory (if not in practice), profile evidence could be offered at trial as evidence helping to establish beyond a reasonable doubt that the defendant is guilty of the crimes with which they are charged.Footnote ¹¹ Postconviction, it could be offered at the time of sentencing to indicate the probability of reoffending if given a short sentence, or fitness for diversion into noncarceral forms of punishment (e.g., probation, supportive housing, or mental health assistance programs). Finally, it can be used at any stage within the sentence duration to determine eligibility for parole.

2. Interpreting the right

We can start unpacking the right to individualized treatment by reviewing how judges have discussed the purpose and value of the requirement. Justice Stevens highlighted the rule’s protective function for establishing probable cause in his dissent in Samson v. California: “The requirement of individualized suspicion, in all its iterations, is the shield the Framers selected to guard against the evils of arbitrary action, caprice, and harassment.”Footnote ¹² In an opinion rejecting the use of actuarial evidence for sentencing in United States v. Shonubi, Judge Newman emphasized that the “specific evidence” requirement is only satisfied by “evidence that points specifically to [behavior] for which the defendant is responsible.”Footnote ¹³ A flat-footed reading of these comments might yield an interpretation contrasting individualization with generalized treatment, leading us to interpret the right as something like:

• • A claim that high-stakes legal decisions be personalized, rather than being subjected to “one-size-fits-all justice.”

But this is too simplistic. As Harcourt (Reference Harcourt2007) stresses, relying on actuarial data actually allows our determinations to be highly tailored to the individual. For instance, rather than having broad sentencing categories, these methods enable us to fit the sentence to the strength of the correlation between the subject’s specific features and rearrest or reconviction. We could in principle make similarly personalized judgments about probable cause or reasonable suspicion using algorithmic tools given the very large databases and high number of personalizing variables these methods allow us to take into consideration. But more personalized treatment isn’t necessarily better. As Lippert-Rasmussen (Reference Lippert-Rasmussen2011) points out, personalization may lead to our being treated worse than otherwise and is in some tension with other weighty principles of justice, such as the generality and equal application of law, and the fair social distribution of various burdens. More urgently, this form of individualization does not capture the connection to the individual’s responsibility stressed in Judge Newman’s comments. The problem is not that statistical determinations are inadequately personalized, but that they are not appropriately responsive: they treat the individual according to how we expect him to act based on our experience with others like him, rather than how he himself has acted. Personalization is not the point.

Moral theorists prefer to characterize the relevant obligation as a duty to be responsive to the individual’s responsible agency, grounded in the values of autonomy or respect. For instance, Dworkin (Reference Dworkin1977) contends that detaining a person based on actuarial prediction, however accurate, is unjust “because that denies his claim to equal respect as an individual.” Duff (Reference Duff, Ashworth and Wasik1998, 155–56) also anchors the claim in respect, holding that:

Nevertheless,

Walen (Reference Walen2011a) articulates the content of the state’s duty to respect the autonomy of its citizens in much the same way: “A state must normally accord its autonomous and accountable citizens this presumption [that they are law-abiding] as a matter of basic respect for their autonomous moral agency.’’ This is consonant with suggestions by Amour (Reference Amour1994), Duff (Reference Duff, Ashworth and Wasik1998), and Moss (Reference Moss2018a) that, in general, taking statistical generalizations as reason to conclude that an individual is probably dangerous runs afoul of the individual’s moral claims.Footnote ¹⁴

But does treating individuals with appropriate respect really require approaching them as a completely novel case without an expectation that our knowledge of other cases will give us reliable guidance concerning them? One might be skeptical whether merely predicting that an individual is likely to offend is a failure of respect or affront to their autonomy.Footnote ¹⁵ Some do argue that viewing someone as predictable in this way fails to regard them appropriately as an agent, rather than a thing determined by external pressures.Footnote ¹⁶ But we needn’t even embrace anything this strong to identify a moral problem with using actuarial predictions to warrant harmfully interfering with a person. Plausibly, it is morally negligent or reckless to intentionally harm someone unless we have not only reasonably high credence (e.g., above some threshold) that the action is morally appropriate, but this credence is resilient. Very roughly: our present evidence must be such that little if any new information consistent with it would cause our credence to drop below the threshold.Footnote ¹⁷ The more harmful the interference, the more resilient the credence must be to justify it. Even if making a prediction based on statistics is not a failure of respect for the individual’s agential freedom, using that prediction as grounds for harming them is a failure of respect for their agential status, because unless supplemented, statistical evidence cannot be adequately resilient.

At a bare minimum, civic respect and equality still requires that the default orientation of law enforcement to any member of the political community not be one expressive of suspicion or disrespect. Considering a person to be probably law-abiding orients police to respect and protect them; considering them to be probably lawbreaking activates a very different script. There are reasons to doubt that, in practice, this difference in default orientation is primarily responsive to evidence rather than to stereotypes or group-based prejudice.Footnote ¹⁸ But even if it tracked group-level rates of arrest, default suspicion would fail to treat citizens as they are entitled. Generalizations about “types of people” or trends in broad demographic categories aren’t sufficient to justify suspending the civic respect owed to a particular person, which demands treating them as probably law-abiding. Borrowing heavily from Duff’s language, we might articulate this as:

• • A claim to be respected as a presumptively law-abiding citizen, unless and until one defeats this presumption through one’s own action and behavior.

The central role this gives to respect and autonomy seems on the right track, and it seems to capture much of the intuitive moral core of the demand that treatment be individualized. But it doesn’t explain what is objectionable about actuarial inferences after guilt has been established—once the presumption has been defeated by admissible, individualized evidence. If the right requires nothing more than that we treat agents as law-abiding until we have adequate particularized evidence that they aren’t, then there is no conflict at all between it and the use of algorithmic risk scores in making sentencing determinations. While one could simply accept this conclusion, it seems to me that the “presumption of law-abidingness” does not exhaust the obligations grounded in civic respect.

What else might it entail? Perhaps:

• • A claim to not be subject to extra burdens simply on account of one’s social identity.

This is the interpretation naturally suggested by Colyvan, Regan, and Ferson (Reference Colyvan, Regan and Ferson2001) and rejected as unrealistically idealistic by Tillers (Reference Tillers2005). There are two ways to develop the thought that equality of standing or respect entitles individuals to be free from extra burden, and I find both plausible. On the one hand, we might be concerned about being subject to disproportionate burdens associated with law enforcement relative to other groups; this is the central animating idea behind Bambauer’s (Reference Bambauer2015) explication of why statistical evidence should not be used to establish probable cause. On the other hand, we might worry about being subject to burden that they would not be subject to were we to hold all else except their social identity fixed; something like this the centerpiece of Underwood’s explanation of why racial membership and other protected categories are an inappropriate base for statistical prediction. She grounds protection against the use of these and other unalterable features in the value of autonomy: “Of all the factors that might be used for predictive purposes, those beyond the individual’s control present the greatest threat to individual autonomy. Use of such factors in a statistical prediction device is particularly undesirable if the device is to be used in a context in which autonomy is highly valued” (1979, 1436). This emphasis on preserving the individual’s control gives us reason to also object to adding burdens to social identities that aren’t themselves protected, but are either unchosen (e.g., socioeconomic status) or reflect important personal choices (e.g., marital status).

Some offer a more procedural gloss of the right, arguing that it is actually a proxy for the right to a certain kind of explanation for the state’s decisions in her particular case:Footnote ¹⁹

• • A claim to an explanation for the state’s exercise of coercive powers.

Vredenburgh (Reference Vredenburghworking paper) compellingly argues that the value of explanations of this kind is instrumental. Access to such explanations is a prerequisite for agents’ ability to act on the political system, to hold it accountable, and form the rules which characterize the basic structure of society. The right so understood requires both more and less than that the subject be given a true account of why a legal decision concerning them has been made; the explanation offered must equip them to act intentionally to hold the decision-making body accountable. It therefore must both be intelligible to them and bear some relation to the actual decision-making procedure employed. In understanding the moral core of the right to individualization as a right to the information necessary to form and reform the legal policies, this gloss aligns closely with Justice Stevens’s comment that the right is a shield against arbitrary uses of state power.

Each of these interpretations highlights something of value in the content of a right to be treated as an individual. Rather than offer a competing interpretation of “individualization,” I suggest that, in fact, the right doesn’t protect a unique interest—rather, the entitlement to “be treated as an individual” simply demands that the procedures of the criminal law be justifiable not merely in the aggregate but to each individual subject to them. Rephrased, it is:

• • A claim to fair distribution of the benefits and burdens of public law.

So interpreted, the right entails that a person must not face disproportionate burden or suspicion except as a consequence of their own responsible action, and that agents of the state default to respectful engagement. It is grounded ultimately in the preconditions for laws to be fair, both in their content and administration. The crowning virtue of the rule of public law is its ability to shape citizens’ practical reason and ground reliable expectations, enabling them to hold each other to standards which they had fair opportunity to meet. Its benefits include many of the values articulated by the interpretations we’ve surveyed: expressing respect for autonomy, constraining the exercise of coercive power, ensuring that sanctions are responsive to responsible agency, and ensuring that those subject to law are in a position to challenge or reform it. The burdens, meanwhile, are the various costs associated with the scrutiny and punitive sanctions applied in the course of enforcing the laws. What fair distribution of burdens and benefits demands depends on context: preconviction, all individuals must have fair opportunity to avoid hostile encounters with law enforcement; at trial, they must not face disproportionate likelihood of false conviction; postconviction, they must not be subject to disproportionate punishment.

To afford all individuals a fair opportunity to avoid hostile encounters with law enforcement, the laws must be public, clear, and prospective.Footnote ²⁰ These are necessary conditions on its ability to structure citizens’ relationships to each other and the state in a way that expresses respect for their autonomy and equality as agents. Citizens can’t know what the law forbids if its requirements are secret or inscrutable; if it is retroactive, they cannot act intentionally to avoid violating it.Footnote ²¹ Importantly, these requirements take lexical priority over considerations of administrative efficiency: the value of the rule of law is lost if the laws are applied in ways that do not facilitate the mutual accountability of citizens and state. Choices about the administration of public law have drastic implications for individuals’ freedom, ability to pursue their life projects, and participation in the political community. Legal transgressions expose a person to the coercive power of the state in various forms, ranging from asset forfeiture to deprivation of liberty and loss of civic rights. General appeals to the aggregate value of efficient crime reduction cannot justify or compensate a particular person for their loss of crucial protections against suffering the state’s coercive imposition of these harms. It is good to lower the average risk of suffering victimization, but if an overall reduction is achieved by dramatically increasing the burdens borne by particular members of the population in a way that neither tracks their responsible action nor is offset by equally weighty benefits to them, the burdens and benefits of the rule of law are not fairly distributed.

I suggest, then, that the right issues an injunction not against the use of probabilistic information or generalizations, but against a familiar form of moral aggregation. Just as invocations of the separateness of persons are in other contexts made to assert that benefits to some cannot offset harms to others, the demand that we treat people as individuals here asserts that wrongdoing by some does not weaken others’ moral claim against the imposition of extra costs, even if they belong to the same demographic group. The right to individualized treatment is the right of those subject to a criminal law that its procedures be justifiable to them individually, not only by appeal to average outcomes. This lens is both unifying and clarifying: it explains why each of the earlier glosses feels partly—but only partly—right. And unlike the other interpretations, which identify relatively all-purpose goods or moral interests, this reading of the right gives it a content specific to criminal law.Footnote ²²

Accepting this interpretation has relatively revisionary implications for the administration of criminal law. While I am mainly exploring the limitations this right imposes on algorithmic prediction tools, it is worth noting that it constrains nonpredictive administrative decisions, too. Consider the practice of cash bail (allowing individuals to avoid pretrial detention conditional on paying a sizable fee, typically $10,000, refunded if they appear on their scheduled court date). The immediate effect is to ensure that some of the most severe burdens of an encounter with the law—lengthy pretrial detention during which the defendant incurs a variety of costs, often including job loss due to forced absence—fall disproportionately on the very poor.Footnote ²³ The median yearly income for people who are detained pretrial because they cannot post a bail bond is $15,109; 37 percent make less than $9,489 per year. This practice leaves those below the poverty line exposed to disproportionately severe burdens without affording them offsetting benefits adequate to justify the imposition. Whether bail determinations are made in highly personalized ways or in deference to an algorithmically produced risk score, insofar as they distribute the burdens and benefits of the rule of law unfairly, they violate the moral interest that animates the right to be treated as an individual.Footnote ²⁴

3. Individual treatment and algorithmic tools

On my analysis, the right to be treated as an individual is, in principle, consistent with the use of algorithmic tools in criminal law. But rather than make a particular policy proposal, I think it more fruitful to fill in the general contours of the constraint, articulating more specifically what it means for a policy to be justifiable to those subject to it. As I have unpacked it, what the right requires is not that the application of legal sanctions be personalized, but that they be justifiable to each individual subject to them. This is necessary for the administration of criminal law to express appropriate respect for each individual’s autonomy and preserve the mutual accountability of citizens and state. Applying this specifically to predictive tools, I claim that if a factor f is used as a predictor in the administration of criminal law, at least three conditions must be met:

• • [control]: f must be subject to agents’ deliberate control

• • [transparency]: It must be transparent that f is used as a predictor such that the basis for decisions is sufficiently clear to facilitate civilian criticism or reform.

• • [burdens]: The unavoidable extra burdens imposed by using f as a predictor (increased hassle, risk of false conviction, or severity of punishment) must be outweighed by the benefits it yields to the individuals who must bear these burdens.

I suspect that, in practice, there are few applications within the administration of criminal law where predictive algorithms can be deployed while satisfying all of these conditions.

Let’s start with the [control] condition. We said earlier that to avoid subjecting anyone to more than their fair share of burden, legal sanctions must be tied to responsible agency. It follows that it must at least be in-principle possible for an individual of any permissible social identity to avoid suffering a downside cost that is not born by everyone. So, if a high-risk score suffices to justify extra scrutiny, conviction, or a lengthier sentence, the predictor variables can’t be tied to unchangeable identity-tracking properties like race or gender. But mere in-principle avoidability is not sufficient to ensure that extra costs track agency. So, the properties used to indicate criminality—and thus to determine the distribution of costly legal sanctions—must also be ones that at least the actually law-abiding agents could act to avoid. They can’t be things that subjects have little real chance of escaping, like residence in high-crime neighborhoods, poor educational background, or an unstable family environment.Footnote ²⁵ We will likely find robust correlations between these features and criminal offending rates—especially in historical databases—but it would violate the right to individual treatment to leverage such correlations to justify the predictive application of criminal sanctions, because doing so concentrates extra hassle and risk on individuals on the basis of factors over which they lack agential control.

The [transparency] condition articulates a precondition for the rule of law. Law expresses respect for subjects’ autonomy only when it enables citizens to anticipate what compliance requires from them. So, in addition to the brute ability to avoid properties that would lead to having a high-risk score, subjects must also be able to act intentionally to avoid them—which means they need to able to know which variables are used and roughly how. Finally, the [burdens] condition acknowledges that some costs are unavoidable and cannot always be distributed perfectly evenly across the population. It allows the imposition of these costs, but only if they are offset by benefits to the individuals who bear them.

4. The space for prediction

One might worry that my interpretation of these constraints is too strict—that the population-level gains to efficiency or accuracy justify violating at least one of them in some contexts. For instance, can’t the deterrence benefits of using algorithmic predictions to guide reasonable suspicion outweigh individual subjects’ moral complaints against failures of transparency provided that the algorithms are sufficiently accurate?

5. Some upshots

We began with a simple question: Is the use of algorithmic prediction tools in criminal law consistent with the right to be treated as individual? And we have arrived at a highly qualified “maybe.” On the interpretation I have offered, this right does not preclude the use of statistical methods in principle but does significantly constrain their design and application. Law enforcement is fundamentally different in its orientation than some other applications of predictive algorithms: the law does not—must not—aim to detect “social cancers” before they manifest. It rather must function to announce expectations for behavior using the coercive apparatus only to hold agents accountable to those very expectations. When legal decisions are made in ways that do not afford subjects a fair opportunity to avoid hostile encounters with law enforcement, or that impose costs disproportionately, this constitutes an unfair distribution of the burdens and benefits of the rule of law. The impulse toward secrecy must be resisted; where predictions are made, they must be based only on factors that are within agents’ deliberate control and not core to valuable exercise of autonomy.

Requiring a fair distribution strongly constrains which variables can be used as a basis for applying extra scrutiny or criminal sanction. It rules out reliance on a great many static factors (age, gender, race), as well as a number of indexes of disadvantage (zip code or neighborhood, income level, previous police contact, number of acquaintances with police contacts or arrest records, education level). The former factors are ruled out because they are unavoidable, the latter because using them to justify the imposition of yet more costs on the victims of upstream distributive injustice—this time in the form of increased risk of suffering unjustified state coercion—is patently unfair. But while it is clearly unjust to base the distribution of burdens on unavoidable factors, you might think the same cannot be said of distributing benefits. Suppose that rather than use high-risk scores to apply sanctions, we were to simply use low-risk scores to exonerate, shorten sentences, or waive bail? There is cause for concern here, too. A policy of this kind channels goods toward those who lack the markers of disadvantage that yield a high-risk score, and, so, can still be expected to entrench racial and economic inequalities and compound disadvantage. It may be an improvement even so, but there is a dark side to reforms which succeed in alleviating injustice for many and concentrate the remaining costs on people who are comparatively vulnerable or politically powerless. Once only marginalized groups face the worst costs, it is much more difficult to muster the political will to enact the reforms necessary to correct the remaining injustice. A partial fix may well be worse than doing nothing then, because it allows the majority to simply look away.

Where this leaves us depends on the application. For street crime—particularly property offenses like auto theft, burglary, or mugging—the social value of predicting any particular future offense is low, especially as compared to the cost of a false positive prediction to each individual who is misclassified. This is because any given offense in this category is either quite difficult to predict with accuracy greater than chance (e.g., homicides) or imposes only relatively minor compensable harm (property damage or loss) on a small number of victims. Insofar as these sorts of crimes are also driven by inelastic social causes, a predictive proxy is unlikely to have strong deterrent effects and is likely to track socioeconomic disadvantage. Even if it is possible in principle to design risk-assessment or crime-prediction algorithms independent of these variables, it is at best unclear what evidential or predictive value a truly unbiased tool would have. Of the extant tools, those that conditionalize on static variables alone presently outperform those that also incorporate dynamic variables.Footnote ³⁵ We can expect that both would outperform prediction based only on the subset of dynamic variables that are not ruled out by the considerations just raised. So, while it may be possible to constrain the data used so that an algorithmic risk projection is consistent with the moral interests protected by the right to individualized treatment, it is unclear whether such predictors will have evidential value sufficient to justify their use.

However, white-collar crime, wage theft, and financial fraud generally may be appropriate arenas for the use of predictive tools. These tend to be more predictable and have a higher victims-per-offense ratio, and consequently there is higher social value to predicting or identifying any single instance. They are also most commonly perpetrated by a relatively well-resourced portion of the population, for whom additional scrutiny presents little more than a hassle. The subpopulations subjected to extra scrutiny, higher risk of false conviction, or longer sentences due to reliance on algorithms in financial crimes are also less likely to overlap with either a stable subgroup (like racial or socioeconomic category) or with populations already subject to intersectional disadvantage and distributive injustice, so the presumptive reasons against using a secret proxy are far less weighty for this application. So, of the possible applications for algorithmic predictions in criminal law, white-collar crime enforcement looks most promising.

In closing, let me revisit the optimistic aim of using algorithmic tools to improve the high-stakes decisions of criminal law. It is laudable to try to make determinations less biased, and to reduce the number of people subjected to unjustified or disproportionate costs in the course of law enforcement. Maybe we could make some progress toward this aim by supplementing the judgment of police officers, judges, juries, and parole boards with algorithmic assessments across the board. But this says more about how badly distorted our unassisted decisions are than about the accuracy or fairness of the algorithmic tools. Whether it is wise to embrace these tools as an incremental improvement depends on several factors that we haven’t had space to work through in this paper, including what the alternative is, and how decision makers would be instructed to incorporate the risk predictions into their deliberations. Without going into detail now, it’s worth emphasizing that the most natural instruction to give—that a high-risk score may be sufficient for an adverse judgment but isn’t necessary—will yield the worst of both worlds. If adverse judgments are still permitted in the absence of a high-risk score, then the algorithmic tool does not constrain any extant bias the decisionmakers may have toward (e.g.) giving disproportionately long sentences to defendants of color. But if a high score is sufficient, then any bias in the false-positive error rates of the algorithm simply combines with the extant distortions—and worse, the whole decision process has a veneer of being evenhanded and objective.