Abolition by Algorithm

Peter N. Salib* March, 2025

In one sense, America’s newest abolitionist movement—advocating the elimination of policing and prison—has been a success. Following the 2020 Black Lives Matter protests, a small group of self-described radicals convinced a wide swath of ordinary liberals to accept a sweeping claim: Mere reforms cannot meaningfully reduce prison and policing’s serious harms. Only elimination can. On the other hand, abolitionists have failed to secure lasting policy change. The difficulty is crime. In 2021, following a nationwide uptick in homicides, liberal support for abolitionist proposals collapsed. Despite being newly “abolition curious,” left-leaning voters consistently rejected concrete abolitionist policies. Faced with the difficult choice between reducing prison and policing and controlling serious crime, voters consistently chose the latter.

This Article presents and analyzes a policy approach designed to accomplish both goals simultaneously: “Algorithmic Abolitionism.” Under Algorithmic Abolitionism, powerful machine learning algorithms would allocate policing and incarceration. They would maximally abolish both, up to the point at which crime would otherwise begin to rise. Results could be impressive. The best evidence evaluating modern machine learning models suggests that Algorithmic Abolitionist policies could: eliminate at least 40% of Terry stops, with high-end estimates above 80%; free a similar share of incarcerated persons; eradicate most traffic stops; and potentially remove police patrols from at least half of city blocks—all without increasing crime.

Beyond these practical effects, Algorithmic Abolitionist thinking generates new and important normative insights in the debate over algorithmic discrimination. In short, in an Algorithmic Abolitionist world, traditional frameworks for understanding and measuring such discrimination fall apart. Traditional frameworks sometimes rate Algorithmic Abolitionist policies as unfair, even when those policies massively reduce the number of people mistreated because of their race. And they rate other policies as fair, even when those policies would cause far more discriminatory harm. To overcome these problems, this Article introduces a new framework for understanding—and a new quantitative tool for measuring—algorithmic discrimination: “bias-impact.” It then explores the complex array of normative trade-offs that bias-impact analyses reveal. As the Article shows, bias-impact analysis will be vital not just in the criminal enforcement context, but in the wide range of settings—healthcare, finance, employment—where Algorithmic Abolitionist designs are possible.

Introduction

In many ways, prison and police abolitionism has been a surprising success. Amidst the 2020 Black Lives Matter protests, a small group of self-described radicals convinced a wide swath of mainstream liberals to accept a series of insurgent propositions: (1) Prison and policing are extremely harmful—more so than is ordinarily assumed, and especially for members of marginalized racial groups; (2) mere reforms cannot meaningfully redress those serious harms; (3) thus, the only way to make a serious difference is to substantially eliminate police and prisons.

The abolitionist movement struck an immediate chord. By June 2020, just a month after George Floyd’s murder, 41% of Democrats supported defunding the police, compared with 19% who supported increased funding.¹ Proponents of police and prison abolition were published in national newspapers and elite academic journals.² Progressive lawmakers with national profiles, including Alexandria Ocasio-Cortez and Ilhan Omar, endorsed the abolitionist argument for reduction over reform.³ Both New York and San Francisco planned large cuts to police budgets.⁴ And the city council of Minneapolis, where police killed Floyd, unanimously pledged to disband its police force entirely⁵

But seen another way, abolitionism has failed. Despite having cultivated a bloc of “abolition curious” left-leaning voters, it has produced little, if any, durable policy change.

The problem is crime. 2021 saw a modest nationwide uptick in homicides, which received significant political attention.⁶ Support for concrete abolitionist policies collapsed. Only 25% of Democrats supported defunding the police, and 3% supported increased funding.⁷ Cities that had cut their police budgets re-funded them at even higher levels than before.⁸ Minneapolis never got around to disbanding its force.⁹ As it turns out, even “abolition curious” voters fear that, whatever the ills of prison and policing, their indiscriminate elimination would do more harm than good.¹⁰ The empirical evidence suggests they are right.¹¹

Nevertheless, abolitionists have manifested an important political opportunity: If there were a policy approach that could abolish significant amounts of prison and policing without increasing serious crime, voters would support it. Or at least, major blocs of voters in Democrat-leaning cities might. The problem is that, so far, no such policy approach appears to exist.

This Article introduces such an approach: “Algorithmic Abolitionism.” These policies are, in one meaningful sense, abolitionist. Algorithmic Abolitionism’s primary purpose is to radically and immediately reduce, not reform, policing and incarceration—and thereby radically reduce their harms. Algorithmic Abolitionist policies are thus not principally designed to reduce crime below current levels. However, unlike untargeted abolitionist practices, for example, “defund the police,” Algorithmic Abolitionist policies would accomplish their reductions without increasing crime.

This Article proceeds in four parts. Part I describes the rise—and stall—of prison and police abolitionism. It begins by arguing that abolitionists are correct about their core empirical and normative claims. A substantial empirical literature documents the serious harms of being caught up in the American criminal enforcement system.¹² These harms, from unemployment to family destruction to early death, are so fundamentally tied to prison and policing as to be mostly irremediable via reform. Major reductions are possible only via less policing and incarceration. But, as Part I shows, voters also acted reasonably in rejecting untargeted police and prison abolition as harmful on net. This is because another robust empirical literature shows even moderate across-the-board cuts to prison and policing substantially increase serious crime.¹³ The sweeping cuts proposed by abolitionists would likely have increased it even more. Poor and non-white Americans, the very groups abolitionism seeks to help, would have suffered the most.¹⁴

Part II introduces Algorithmic Abolitionism as a solution to this impasse; it offers a comprehensive suite of Algorithmic Abolitionist policies, spanning the entire criminal enforcement system, to abolish large shares of prison and policing without increasing serious crime. Here is one concrete example: For over a decade, computer algorithms have predicted whether a given arrestee or convicted person will likely engage in further crime if not incapacitated.¹⁵ Such predictions could be leveraged into Algorithmic Abolitionist incarceration policies. The state could start with the lowest-risk individuals—the ones almost certain not to commit additional crimes. It could release as many people as possible, climbing the risk gradient up to the limit at which crime rates would begin to increase. This policy maximally reduces the harm of incarceration—by literally abolishing a portion of it—while retaining its present crime-controlling effect.

The extent of abolition depends on how accurately the operative algorithm can sort high-risk individuals from low-risk ones. The crime prediction algorithms of the past probably lacked much abolitionist potential. But the last decade has seen a revolution in machine learning, including major advances in computer vision, natural language processing, and more.¹⁶ Likewise, today’s best crime prediction algorithms have significant abolitionist potential. High-quality empirical evidence suggests that one algorithm published in 2018 could be used to reduce pretrial incarceration by as much as 41.9%, with a bare-minimum reduction of 18.5%.¹⁷ Similar evidence suggests that an even newer algorithm—published just last year—would facilitate releasing as many as 80%, and at least 50%, of parole-eligible prisoners.¹⁸ There are reasons to think that this result could generalize to the prison population as a whole.¹⁹ Using just these algorithms, then, Algorithmic Abolitionist policies might free millions of Americans per decade who would otherwise be incarcerated—all without generating an increase in crime.

Algorithmic Abolitionism has significant implications beyond incarceration. Point estimates from high-quality academic evaluations of the best existing algorithms suggest that such polices could: cut police street stops by up to 88% (and at least 42%) without affecting rates of dangerous contraband recovery; eliminate the police presence from 50% or more of urban blocks without increasing criminal victimization; and eliminate nearly all traffic stops, a major source of police violence,²⁰ while detecting as many traffic violations as status-quo procedures.²¹ They could literally abolish much or most of the police, but without the downsides of indiscriminate policies like “defund.”

All of this is feasible today. In fact, many jurisdictions already use algorithms to some extent in directing criminal enforcement.²² However, none have achieved the levels of abolition just described. Part II argues that this is because those jurisdictions have uniformly implemented the wrong policies and probably used the wrong algorithms. Part II closes by enumerating a few key principles for designing Algorithmic Abolitionist policies. It shows that the main deficiencies in existing plans could be remedied with a stroke of the legislative pen.

The Article’s second half turns to normative considerations. On top of its important practical effects, Algorithmic Abolitionist thinking opens new normative terrain. In particular, Algorithmic Abolitionism invites a significant rethinking of well-worn debates over algorithmic discrimination—which Part III takes up.

Historically, progressives, including prison and police abolitionists, have opposed the use of algorithmic technology in criminal enforcement with near unanimity, principally citing fears of racial injustice.²³ The literature on algorithmic discrimination has grown dense and thorny. Experts offer competing theories—and thus quantitative measures—of fairness. Worse, an impossibility theorem has emerged: For essentially all algorithms, becoming perfectly fair according to one theory mathematically entails discriminating according to the others.²⁴

Yet algorithmic decisionmaking is ascendant, in both criminal justice and elsewhere. This Article therefore asks whether algorithmic policies can be designed which not only avoid perennial racial justice critiques but actively promote equitable outcomes.

Part III begins by arguing that Algorithmic Abolitionist policies sidestep the impossibility theorem. Such policies would generally reduce—even if not eliminate—bias according to multiple theories of fairness simultaneously. This, the Part argues, should be sufficient reason to favor Algorithmic Abolitionist policies over alternatives.

Part III also introduces three entirely new arguments into the literature on algorithmic fairness. Each argument arises from the main thing that makes Algorithmic Abolitionist policies unique: their ability to radically reduce levels of carceral harm. By contrast, traditional accounts—and metrics—of algorithmic fairness are concerned with distributions of harm. Part III shows that these distribution-focused frameworks often break down when applied to levels-reducing policies. A new evaluative approach is needed.

To see why, consider the following thought experiment: Imagine a country where one million citizens are locked away in prisons and jails.²⁵ Imagine that 33% of them are Black, despite Black citizens composing only 12% of the hypothetical nation’s population.²⁶ In this country, according to a simple theory of fairness,²⁷ Black citizens are locked up at nearly three times the fair rate.²⁸ But imagine that, in this country, a hypothetical Algorithmic Abolitionist sentencing policy could be implemented. Under it, crime would remain constant, but the number of incarcerated citizens would fall from one million to 100,000. Suppose, however, that, of those 100,000, 40,000 would be Black.

Standard, distribution-focused frameworks for measuring algorithmic fairness would score the Algorithmic Abolitionist policy much worse than the status quo. After all, the new policy would increase the Black share of incarceration from 33% to 40%.

Part III’s first new normative argument is that the standard approach is missing something crucial.²⁹ True, the hypothetical policy worsens incarceration’s racial distribution in the hypothetical nation. But counterintuitively, and much more importantly, the policy drastically reduces the number of people mistreated because of their race. Here are the numbers: Under the status quo, about 210,000 more Black citizens are locked up in our imaginary nation than the fair rate would dictate.³⁰ That is, 210,000 people are discriminatorily imprisoned because of the color of their skin. But under the hypothetical Algorithmic Abolitionist policy, the number of people imprisoned because of race would fall to 28,000—a reduction of 87%.³¹

This is a generalizable failure of the standard approach to algorithmic fairness. It occurs whenever traditional, distribution-focused tools are employed to evaluate a policy that radically reduces the level of a harmful outcome. In such cases, traditional tools miss what matters most: the discriminatory injuries borne by actual human beings.

Formalizing these insights, Part III introduces a new quantitative tool for measuring algorithmic fairness: “bias-impact.” Bias-impact measures the percentage reduction in the number of people who suffer discrimination—harmful outcomes unfairly imposed because of their race. It is a flexible tool. Bias-impact can be used to measure algorithmic discrimination under all major theories of algorithmic fairness. Moreover, as Part III shows, no matter one’s preferred theory, the actual Algorithmic Abolitionist policies proposed herein would significantly reduce unfairness, measured using bias-impact.

Moving on, Part III introduces a second novel argument into the literature on algorithmic fairness. It concedes that, in addition to individual discriminatory injuries (captured by bias-impact), pure distributional fairness (captured by traditional tools) carries some normative weight. The question becomes how to trade off distributional improvements against other benefits—like reductions in the number of people suffering discrimination or in the total number of people imprisoned and policed.

Part III proposes a new, strict framework for evaluating such tradeoffs. The framework is both demanding and highly solicitous toward distributional justice. So solicitous, the Part argues, that almost any proponent of pure distributional fairness ought to endorse the framework’s recommendations.³² The framework centers on a rule dubbed the “Super Difference Principle,” which borrows both prioritarian and Rawlsian distributional constraints—and then supercharges them. The Super Difference principle demands that a policy: (1) have, among practicable alternative policies, the highest “floor” of outcomes for the worst off and (2) deliver the largest total benefit to the group that, under the status quo, was the most disadvantaged. Part III shows that Algorithmic Abolitionist policies would almost inevitably satisfy the Super Difference Principle. This is again because Algorithmic Abolitionism radically reduces bad outcomes, delivering the biggest benefits to those who, under the status quo, would suffer the most.

Part III’s third novel argument about algorithmic fairness refocuses the debate over structural inequality through the lens of Algorithmic Abolitionism.³³ Structural inequality, it is often argued, arises when modest initial inequalities compound over time. But, Part III argues, existing accounts of that compounding are incomplete. They lack a model of differential growth: why, for some racial groups, but not others, policing begets ever more policing. Part III supplies such a model. It proposes that carceral harm, like a virus, has a reproduction rate.³⁴ One arrest may cause, on average, one more arrest, more than one, or less than one. If one group’s reproduction rate for arrests is above one, while another’s reproduction rate is below one, arrests among the former, but not the latter, group will compound. Part III gives reasons to think that ambient levels of prison and policing drive their reproduction rates. If so, then Algorithmic Abolitionist policies would, by substantially reducing those levels, push reproduction rates downward for all groups. Doing so, they would promote long-run structural equality, not inequality, in prison and policing.

Finally, Part IV offers brief treatments of several other normative concerns. Why the brevity? Because there is less to say. Algorithmic Abolitionist thinking, as just described, offers many new insights into debates over algorithmic discrimination. It offers some insights, but less decisive ones, elsewhere. Part IV describes how Algorithmic Abolitionism fits into conversations about alternatives to policing; the political economy of crime; constitutional constraints; and privacy and surveillance. Two themes recur throughout the Part. First, unlike other policy approaches, Algorithmic Abolitionism is positive-sum. It offers the possibility of simultaneous wins to groups whose goals are usually in conflict with one another. Algorithmic Abolitionist policies could, for example, reduce both crime and punishment for crime. Second, Algorithmic Abolitionism’s reduction in levels of prison and policing would, unlike other algorithmic enforcement policies, reduce the total amount of harm from invasions of privacy, constitutional violations, and more.

I. Prison and Police Abolitionism

The American movement to abolish police and prisons feels new. And it is, at least, newly prominent. Beginning in the mid-2010s, a series of police killings—mostly of unarmed Black men—brought the high costs of policing into public light. Each killing amplified the ever-growing wave of public protest against police violence.³⁵ Then, in 2020, George Floyd was murdered by Derek Chauvin in Minneapolis.³⁶ The ensuing protests were nationwide and massive. Polling suggests that tens of millions of Americans attended the rallies that summer, which would make them the largest in American history.³⁷

Atop this groundswell of public anti-policing sentiment, abolitionist ideas broke through into the liberal and progressive mainstream. Most prominently, activists across the nation called on governments to “defund,” “disband,” or “abolish” the police.³⁸ No longer were these calls sequestered to the radical fringes. Rather, prominent non-abolitionist-identifying liberals adopted them nationwide. “Defund” proposals were published in the nation’s paper of record³⁹ and endorsed by sitting congresspeople.⁴⁰ And in June of 2020, a substantial plurality of Democratic-identifying voters supported such policies.⁴¹ In Minneapolis, where Floyd was murdered, the city council unanimously voted to disband the entire police department and replace it completely with a to-be-defined “community-led public safety system.”⁴² Other major cities pledged to substantially cut police funding.⁴³

In truth, the prison and police abolitionist movement long predates 2020. Foundational thinkers Angela Davis, Ruth Wilson Gilmore, and Rose Braz launched Critical Resistance⁴⁴—an organization devoted to “eliminat[ing] the prison industrial complex”—in 1997.⁴⁵ Davis published her book, Are Prisons Obsolete?, in 2003.⁴⁶ And Critical Resistance launched its newspaper, The Abolitionist, in 2005.⁴⁷ Allegra McLeod’s Prison Abolition and Grounded Justice introduced the “first sustained discussion” of prison abolition into the law review literature in 2015.⁴⁸ A deluge of articles soon followed.⁴⁹ Thus, abolitionism is a diverse, multigenerational movement, encompassing thinkers who both agree and disagree on many points. This Article’s recounting of the movement is necessarily cursory.

The Article engages with prison and police abolitionism in terms familiar from its recent breakthrough into mainstream discourse. Whatever their other disagreements, in 2020, abolitionists succeeded in convincing a wide swath of mainstream liberals of two core beliefs: First, prison and policing are inherently harmful, such that mere reform cannot adequately address their evils. Second, consequentially, in order to reduce harm from prison and policing, both should be mostly or completely abolished.

A. A Convincing Diagnosis

Abolitionists are correct about the problem. Prison and policing are inherently harmful, and no amount of reform can eliminate the relevant harms. Certainly, high-profile tragedies like the police killings of unarmed Black men are part of the problem—and perhaps emblematic of it. But they are just the tip of the proverbial iceberg. As abolitionists rightly contend, even in the best-case scenario—if the system were working as intended without overt corruption, conscious racism, or evil intent—policing and imprisonment would be extraordinarily costly.⁵⁰

Many of these costs are, in fact, the point. Punishment, including imprisonment, is intentionally harmful. It is designed to impose costs on social bad actors. Imprisoned persons are separated from their families and friends, made to live under poor conditions, and restricted in their physical movement and activities precisely because these things harm them. Policing, too, is costly on its face. In the ordinary course of detecting and preventing crime, police must stop, search, seize, and arrest citizens. Even if unaccompanied by violence—actual or threatened—these impositions of state power over individual liberty are damaging. It is the reason the Fourth Amendment exists—to place limits on the amount of damage the police can do when fighting crime.⁵¹

These observations are not unique to abolitionism. Liberal thinkers since Bentham have classed punishment as an evil “in itself.”⁵² Under Benthamite Utilitarianism—and derivative consequentialist theories—such evil is permitted only to the extent necessary to prevent even greater harms from crime.⁵³

But, as abolitionists emphasize, the heavy costs of prison and policing go far beyond the baked-in harms necessary to achieve deterrence. Prisons, for example, are not actually sterile sites of moderate isolation, constrained freedom, and modest living—secure enough to incapacitate and bad enough to deter, but no more. They are instead sites, as McLeod puts it, of “intense brutality, violence, and dehumanization.”⁵⁴ Currently, tens of thousands of imprisoned people are locked in solitary confinement, caged in small cells, and completely isolated for twenty-three or twenty-four hours per day.⁵⁵

Again, the abolitionists are not alone in raising such concerns. A bevy of quantitative social science research shows that the costs of policing and prison go well beyond the intended ones. The empirics show that the costs of an encounter with the criminal justice system can redound for a lifetime, or beyond. First, there are the economic consequences. Contact with the criminal justice system seriously impairs one’s ability to find a job and earn a living.⁵⁶ This effect worsens—in both severity and persistence—as the contact grows.⁵⁷ This is true even for those to whom the presumption of innocence applies. Being held in jail pretrial—even briefly before charges are dropped or an acquittal is entered—can be economically disastrous.⁵⁸ Employers need not and do not hold jobs open during such periods of incarceration. Worse, since incarcerated people are disproportionately poor, losing a job can tip off a spiral of economic hardship—high-interest loans, unpaid debts, eviction, and more. Other economic costs of the criminal justice system are imposed directly. Criminal defendants are legally obligated to pay certain fees arising from their cases—sometimes including tens of thousands of dollars in “room and board” for their prison stays.⁵⁹ If these bills go unpaid, perhaps due to lack of employment, the penalty is often additional fines and fees.⁶⁰

Criminal punishment also destroys relationships. Even after incarceration ends, couples with an incarcerated member are more likely to divorce.⁶¹ The children of incarcerated people are thus more likely to live in a single-parent household both during and after incarceration.⁶² Beyond family, incarcerated persons generally have fewer strong social connections than the population at large.⁶³

The criminal justice system can also kill you, even long after your last encounter with it. Former prisoners have substantially worse health outcomes than other groups.⁶⁴ They are more likely to die early.⁶⁵ The effects can stretch beyond the grave. The children of incarcerated persons have worse life outcomes than those of similarly situated nonincarcerated persons.⁶⁶

Finally, incarceration is criminogenic—that is, it causes crime. Quasi-randomized studies show that, beyond some minimum, each marginal month spent imprisoned increases the expected number of crimes a person will commit over their lifetime.⁶⁷ This paradoxical result is discussed in greater detail below,⁶⁸ but the intuition is easy enough to grasp: If being imprisoned destroys your earning power, saddles you with unpayable debt, weakens your social network, and worsens your health, it might for those same reasons drive you to commit more crimes. This induces a vicious cycle—prison causes harm, which causes more crime, which causes more prison, and so on.

As abolitionists regularly emphasize, these costs are not evenly distributed. At every juncture—search, seizure, arrest, jail, charging, trial, sentencing, and parole—the harms of prison and policing are disproportionately imposed on already-disadvantaged groups. Black men bear the brunt.⁶⁹ But so do other people of color, the poor, sexual minorities, and other marginalized communities.⁷⁰ Yet again, quantitative social science supports the abolitionist diagnosis. Robust empirical evidence shows that these racial and other social disparities cannot be fully explained by differences in base rates of crime commission between groups. Rather, discrimination is at work across the criminal justice system.⁷¹

Thus, the abolitionists are clearly right about something: Prison and policing are extremely costly, both because that is their point and because of their many downstream, legally unintended consequences. None of that is seriously contestable.

But abolitionists go even further, arguing that these costs are so deeply entrenched in policing and prison that no amount of reform could satisfactorily ameliorate them. Framed a certain way, this claim is contestable. For example, as with other kinds of violence, a large proportion of police violence and other misconduct is committed by a small proportion of officers.⁷² Yet, due to special protections written into law and police union contracts, those officers are almost never fired or disciplined.⁷³ Thus, violent police encounters might be reduced simply by making police officers subject to ordinary mechanisms of legal and professional accountability.⁷⁴ Reforms might also improve racial unfairness in policing. Evidence from some cities suggests that Black and Hispanic officers are less biased in making stops and arrests or using force.⁷⁵ And these officers initiate fewer stops and arrests with force than their white counterparts.⁷⁶ Thus, reforms that diversify police departments, too, could make a difference.

But framed another way, the abolitionist claim that reform cannot meaningfully address the serious harms of prison and policing is obviously true. As just discussed, the problems with policing and prison are not limited to excesses of the system. Much of what is bad about prison and policing is inherent in both—a feature, not a bug. And that which is not inherent is exceedingly unlikely to be completely—or even mostly—fixed via reform. For example, it is unlikely that reforms to incarceration could prevent the downward economic spirals that so often result from even short periods of confinement. Most jobs cannot be done remotely from a prison cell, even if employers were willing. And employers are not willing. Moreover, the hiring stigma against individuals who have had contact with the criminal justice system remains intense.⁷⁷ It is not clear that it can be reduced via policy. Previous attempts have backfired.⁷⁸ The same goes for family, interpersonal, and intergenerational consequences. As long as the criminal enforcement system continues to involve taking people out of their families and communities and placing them in new criminogenic communities, there will be serious social effects.

The best version of the abolitionist argument is thus not that reforms can do nothing to blunt these effects. It is that they cannot do enough. They fall far short of a cure. Both logic and empirical investigation show the abolitionists have all of this correct.

B. Bitter Medicine

If the lion’s share of prison and police’s costs cannot be redressed by reform, a different solution suggests itself: abolish prisons and police. That is exactly what abolitionists now demand. This position is perhaps exemplified by the movement’s most successful policy campaign to date, “defund the police.” Almost immediately after George Floyd’s murder, a nationwide chorus of activists called for police departments to be literally and immediately eliminated—either in whole or in large part.⁷⁹

As already noted, these proposals were supported by a plurality of Democrat-identifying Americans—the majority party in most major cities with large police forces. Those cities took their voters’ preferences seriously. As already noted, Minneapolis’s city council almost immediately pledged to “dismantl[e]” the city’s police department.⁸⁰ And other cities, to a lesser degree, followed suit. San Francisco’s mayor promised to cut police funding by $120 million—approximately 14% of the projected budget.⁸¹ New York likewise pledged to cut its policing budget by $1 billion, or about 17%.⁸² Political figures, media outlets, and academics lent their support to the cause. Ilhan Omar—whose congressional district encompasses much of Minneapolis—said of the city’s “defund” plan that “you can’t really reform a department that is rotten to the root.”⁸³ Alexandria Ocasio-Cortez wrote that policing is “an indefensible system” and that “reform measures do not ultimately solve what is a systemic problem.”⁸⁴ The New York Times ran an opinion piece headlined, Yes, We Mean Literally Abolish the Police.⁸⁵ In it, activist Mariame Kaba advocated—as an intermediate abolitionist goal—immediately halving police budgets and officer counts.⁸⁶ In an essay in the Georgetown Law Journal Online, India Thusi contends that the demand to defund the police “deserves serious scholarly engagement.”⁸⁷

Then, abolitionists lost their momentum. Their early policy successes in places like Minneapolis, New York, and San Francisco proved illusory. The reason, it seems, was crime. In 2021, the United States saw a modest—but nontrivial—nationwide uptick in homicide, widely discussed in the runup to the midterm elections.⁸⁸ Public support for policies like “defund the police” collapsed, even among Democrats.⁸⁹ Neither New York nor San Francisco made the promised deep cuts to police budgets.⁹⁰ And Minneapolis quietly shelved its plan to disband its police force.⁹¹

Abolitionists, it seems, managed to convince left-leaning voters of the diagnosis. But when it came time to administer the cure, it proved too bitter to swallow. Plausibly, the nationwide increase in homicides reminded ordinary citizens that prison and policing, while extraordinarily costly, are there to try to prevent the even larger costs of serious crime. Perhaps, they reasoned, that for all of their ills, prison and policing were a necessary evil; even moderate across-the-board cuts would do more harm than good, including for the groups policing and prison burden most.

This is all correct, too, as another robust body of empirical findings shows. Serious crime disproportionately affects low-income people, people of color, and members of sexual and gender minority groups.⁹² Such crimes cause grievous injury or death to their victims. They also have economic effects, significantly reducing victims’ earnings, increasing their reliance on public benefits, and reducing their health.⁹³

Moreover, prison and policing are vital tools for controlling crime. Even modest-across-the board cuts cause significant increases in serious crime. A handful of decades-old, empirically unsophisticated studies once suggested otherwise.⁹⁴ And, occasionally, some abolitionists will repeat those old claims.⁹⁵ But since the 1980s, a raft of studies with much better causal identification strategies have shown that, for better or worse, the old view was wrong. Here is a necessarily incomplete survey.

1. On Policing

A clever 2005 study uses exogenous variation in the terror alert level—which in turn caused variation in the level of policing in Washington, D.C.—to measure the effect of policing on crime.⁹⁶ The study finds significant effects: a 6.6% citywide reduction in crime on “high” alert days, compared with “elevated” alert days.⁹⁷ The reduction was over twice as high—15%—on the National Mall, where police presence likely increased the most during periods of high terrorism alert.⁹⁸

The crime-preventing effects of policing are not limited to minor, quality-of-life crimes. Policing prevents murder. A study published in 2022 leverages variation in the timing of federal block grants to measure the effect of police on homicides.⁹⁹ It finds that hiring ten to seventeen police officers saves, on average, one life per year that would otherwise end by homicide.¹⁰⁰ The per-capita effect is significantly larger for Black victims.¹⁰¹ A similar study from 2019 compares cities that barely qualified for federal policing subsidies with those that barely failed to qualify.¹⁰² It finds similar results.¹⁰³ Hiring ten additional officers prevented approximately one murder, five rapes, and twenty robberies annually.¹⁰⁴

A 2016 paper investigates what happened when New York directed surges of police officers into high-crime areas.¹⁰⁵ It finds that the surges caused an 11% reduction in all violent felonies, including a 12% reduction in assaults and a 15% reduction in robberies.¹⁰⁶ Major property crimes fell, too—and by even larger amounts. Overall, property felonies went down by 26%, and burglary fell by 46%.¹⁰⁷

2. On Prison

Punishment—and incarceration in particular—prevents crime at the societal level. In 2011, California was forced by a federal court to reform its corrections code and enact a “very large and relatively sudden decline in incarceration rates.”¹⁰⁸ A 2016 study relied on county-level variation in pre- and post-reform incarceration rates to estimate the change’s effect on crime.¹⁰⁹ It finds that each additional release generated as many as 0.04 violent crimes each month for every 100,000 California residents.¹¹⁰ An earlier, similar study—likewise estimating the effects of broad decarceration forced by litigation—finds similar jurisdiction-wide results.¹¹¹

Studies quantifying incarceration’s individual-level, rather than society-level, effects often find net reductions in crime. One recent study leveraging Ohio’s random assignment of defendants to judges—who vary in their average harshness in sentencing—finds a modest net reduction in crime from incapacitation.¹¹² Another study from 2021 uses a sudden change in North Carolina’s sentencing guidelines to estimate the dose-response effect of incarceration.¹¹³ That is, it quantifies how much each marginal year of incarceration reduces the incarcerated person’s likelihood of being reincarcerated in the future.¹¹⁴ The study finds a large crime-reducing effect from the first year of incarceration—44% less incarceration over the next three years—with effects diminishing for each additional year.¹¹⁵

This finding is consistent with the small number of studies finding that incarceration is, on net, criminogenic as to the incarcerated individuals.¹¹⁶ These do not suggest that prison lacks any incapacitative or specific deterrent effect. Rather, it just suggests that those effects diminish as sentences grow longer and can thus, under certain conditions, be overtaken by criminogenic effects.¹¹⁷ More importantly, none of this contradicts the claim that incarceration decreases crime at the level of society—as opposed to at the level of individual incarcerated persons. Even if it increased crime for those actually incarcerated, this relatively small crime-producing effect could be swamped by the general deterrent effect on everyone else. And that is exactly what the studies—discussed above—¹¹⁸of prison’s total societal effect find.

Hence, the turn among ordinary liberals against abolitionist policies was reasonable, assuming that they cared about both the harms from prison and policing and the harms from serious crime. The turn can be understood as a rejection of policies that would reduce one set of important harms only by increasing another. Voters may simply have determined that abolitionist proposals’ costs would, on balance, outweigh their benefits.

This suggests that ordinary liberals’ turn against prison and police abolitionism might well be limited to rejecting just the particular policies on offer—not the ideology wholesale. There is good reason to believe that abolitionists’ normative arguments retain their full force. Nothing about the 2021 homicide spike undermines the arguments for policing and imprisonment’s heavy costs, nor their incurability via reform. Thus, it stands to reason, the voters who supported abolitionists in 2020 might support them again, if only there were a way to substantially reduce prison and policing without increasing crime.

II. Algorithmic Abolitionism

Algorithmic Abolition offers such a solution. It is a way to maximally act on prison and police abolitionists’ forceful normative critique while responding to concerns about serious crime harbored by even “abolition curious” liberal voters. Algorithmic Abolitionism is abolitionism in that it radically reduces carceral harm by elimination, not reform. But unlike previous abolitionist proposals, Algorithmic Abolitionism would accomplish its elimination without allowing crime to increase.

To be clear, Algorithmic Abolition is not total abolition. The approach would not eliminate all prison and policing on its own. It is thus perhaps best described as what Amna A. Akbar has dubbed a “non-reformist reform,” the kind of important intermediate step toward fundamental transformation that radicals should embrace.¹¹⁹ In particular, Algorithmic Abolitionism would satisfy Akbar’s key criterion of helping to shift the “balance of power” away from carceral authorities.¹²⁰ Likewise, Algorithmic Abolitionist policies could directly achieve the goals of prominent abolitionists like Mariame Kaba, reducing various kinds of policing by well over her target of 50%.¹²¹

Algorithmic Abolitionism has become possible only recently—in the last five to seven years. Algorithms, however, have been used in criminal enforcement for much longer than that. The infamous¹²² COMPAS tool, for example, has been used by local governments to make bail decisions since at least 2008.¹²³ Other actuarial—and thus algorithmic, in a sense—approaches to risk scoring were first used for the same purpose in the 1960s.¹²⁴ But there is little reason to think that these earlier algorithms could accomplish the feats of decarceration and de-policing described herein.¹²⁵

For algorithms to unlock significant abolitionist potential, modern machine learning was required. The world is currently in the midst of a “Machine Learning Explosion.”¹²⁶ The explosion was triggered by a confluence of related technological factors: breakthroughs in algorithmic design, massive increases in data, and ever more processing power.¹²⁷ Over the past twenty years, each of these preconditions for rapid algorithmic advancement has improved by at least one order of magnitude and as many as three.¹²⁸

The result has been a blinding proliferation of algorithmic accomplishments. Modern machine learning algorithms power large language models, like GPT-4, that can perform a range of text-based tasks from coding to passing the bar exam.¹²⁹ They safely pilot the autonomous taxis currently carrying riders across San Francisco.¹³⁰ They can generate creative, beautiful, and completely unique images in any artistic style, based on simple natural-language prompts.¹³¹

Modern machine learning techniques were perhaps first brought to bear on criminal enforcement by Jon Kleinberg, Himabindu Lakkaraju, Jure Leskovec, Jens Ludwig, and Sendhil Mullainathan, through their algorithm published in a 2018 paper.¹³² Their algorithm’s goal—like many algorithms before it—was to determine which defendants would be likely to commit crimes if not incarcerated while awaiting trial.¹³³ As with other applications of next-generation machine learning to difficult problems, the result was a quantum leap forward. As discussed in detail below, their algorithm performed remarkably better than the status quo.¹³⁴ Today, high-quality, peer-reviewed empirical evidence shows that, by using algorithms to make various decisions about prison and policing, dramatic reductions—on the order of 40%, 80%, or more, depending on the context—are possible.

This Part lays out—for the first time¹³⁵—a comprehensive vision of how algorithms could begin abolishing police and prisons. It goes beyond narrow discussions of, for example, pretrial risk assessment, considering algorithms’ potential across the whole range of criminal enforcement functions. The Part is quantitative. Relying on the best available empirical evidence, it shows both that prison and policing could be dramatically reduced using algorithms and—for each function—attempts to estimate by how much. Law review literature has started to absorb some of the operative evidence—especially the Kleinberg et al. algorithm—but much of this evidence is too new to have been properly incorporated.¹³⁶

This Part goes beyond merely collecting existing evidence. It represents the first attempt to think systematically about how modern predictive algorithms could be implemented to maximum abolitionist effect. As the Part describes, many jurisdictions have already begun to experiment with the use of algorithms in policing and incarceration. But nowhere have those algorithms yet produced the dramatic reductions in either institution described here. This Part explains why. Doing so, the Part lays down a set of key principles for the successful design and implementation of Algorithmic Abolitionist policies.

A. Prisons

Of the policies proposed in this Article, the one that has come closest to having been tried is the Algorithmic Abolition of prisons. In recent years, many jurisdictions have incorporated algorithmically generated risk scores into their pretrial incarceration decisions.¹³⁷ And a few states have incorporated them into sentencing post-conviction.¹³⁸ Section II.C returns to these experiments to suggest that none of them has, in fact, been either truly abolitionist or truly algorithmic.

Let us first explore how much decarceration would be possible under a genuine Algorithmic Abolitionist regime. Begin with pretrial incarceration: Here, the Algorithmic Abolitionist goal is to eliminate as much jailing as possible without increasing either of two bad outcomes. First, if released, certain defendants might flee the jurisdiction and thus avoid responsibility for crimes already committed.¹³⁹ Second, if released, certain defendants might commit further crimes.¹⁴⁰ If algorithms could predict who will jump bail or commit crimes with greater accuracy than status-quo decisionmakers—judges—then much needless incarceration could be avoided.

Indeed, algorithms vastly outperform humans. Consider again the aforementioned algorithm—a gradient boosted decision tree—trained by Kleinberg et al.¹⁴¹ This algorithm was designed, in the first instance, to predict failure to appear.¹⁴² But Kleinberg et al. show that their algorithm would perform even better if asked to predict freestanding crimes.¹⁴³

The algorithm’s predictions—and incarceration decisions—were compared in a simulation against the actual decisions judges made. The algorithm came out far ahead. The 1% of defendants that the algorithm identified as having the highest risk were rearrested 62.7% of the time.¹⁴⁴ Yet judges released nearly half of those individuals.¹⁴⁵ The judges who did so were not simply being lenient. The strictest subset of judges jailed more people—but an essentially random collection from across the risk spectrum.¹⁴⁶ These judges were not meaningfully better at predicting and detaining the individuals most likely to commit crimes upon release.

What kind of Algorithmic Abolitionist policy could be implemented using such predictions—and to what effect? Suppose high-performing algorithms, rather than low-performing judges, imposed pretrial incarceration only on those individuals most likely to commit crimes if released. What risk level would be sufficient to justify incarceration? Here, we can appeal to the definition of Algorithmic Abolition outlined above: Algorithmic Abolitionism seeks to eliminate as much incarceration as possible without increasing crime. Thus, the crime-risk cut score would be set at the status quo rate of recidivism. Anyone under that score would be released and, thus, crime would be held constant.

What impact would such a policy have? Kleinberg et al. tell us. They performed exactly this simulation—ranking pretrial defendants from riskiest to least risky and release as many as possible without increasing observed crime. The result: pretrial incarceration could be reduced by as much as “41.9% with no increase in crime rates.”¹⁴⁷ Hence, if algorithmic, risk-based incarceration decisions completely replaced judges, the American level of pretrial incarceration would be permanently reduced by roughly 185,000 people annually.¹⁴⁸ In the first decade alone, millions of prison years would be avoided. And the resulting predicted increase in crime—from bail jumping to misdemeanors to murder—would be zero.

Some might worry that these New York-based results would not generalize to the rest of the country. On the contrary, the Kleinberg et al. pretrial risk algorithm also works on a national dataset, to similar effect.¹⁴⁹

Others might worry that the massive effect reported above is overstated as a best-case scenario. True, a reduction of over 40% is the high end of the reported range. But the low-end estimate is still unprecedented: an 18.5% reduction in incarceration—a total of 82,000 people.¹⁵⁰ Importantly, this low-end estimate is conservative in the extreme. It assumes that every single person in the dataset that New York judges incarcerated—and whose actual post-release behavior was thus unobservable—would have committed crimes if released.¹⁵¹ Given the extremely high rate at which judges incorrectly released high-risk defendants, it seems extraordinarily unlikely that their error rate incarcerating people could have been anywhere near zero. Thus, the optimistic scenario is much more likely than the pessimistic one, and even the pessimistic one would constitute remarkable abolition.

Pretrial detention is, of course, only half the story of incarceration. Or more precisely, it is a bit less than a quarter of the story. In the United States, about 400,000 people are currently jailed awaiting trial. But 1.9 million in total are incarcerated.¹⁵² The difference—about 1.5 million people—are imprisoned because they were convicted of a crime.

How could an Algorithmic Abolitionist approach transform post-conviction incarceration? The approach here is roughly the same. As with pre-conviction incarceration, post-conviction imprisonment would be eliminated to the maximum extent possible without causing crime—and especially serious crime—to increase.

Evidence suggests that, as with pretrial incarceration, algorithms could be used to abolish large proportions of post-conviction imprisonment. In an article in the Journal of Quantitative Criminology, Hannah S. Laqueur and Ryan W. Copus report the performance of a Super Learner ensemble algorithm trained to make parole decisions.¹⁵³ Their approach to simulating the algorithm’s ability to outperform judges was broadly similar to that of Kleinberg et al.¹⁵⁴ Their data likewise comes from New York state.¹⁵⁵

Their results are even more striking: The public officials currently tasked with deciding incarceration are, once again, extremely bad at it, at least if the goal is minimizing crime. The crime-risk distribution—including violent crime—of the people New York parole boards release is very similar to those they send back to prison.¹⁵⁶ That is the result one would expect if parole decisions were made by coin flip. Thus, many people with a high probability of committing violent crimes are being released;¹⁵⁷ many more with almost no risk of harming anyone are being locked up.¹⁵⁸

How much imprisonment could be eliminated under an Algorithmic Abolitionist regime? In the period observed, New York parole boards granted parole to just 20% of people eligible for it.¹⁵⁹ They thus denied 80% of petitions. Algorithmic decision-making would have inverted those numbers. That is, the algorithm could have released 80% of people up for parole—again, with no increase in crime.¹⁶⁰ On net, then, algorithms would have increased release among parole-eligible individuals by 60% without any tradeoff in serious crime.

As with Kleinberg et al., Laqueur and Copus also report a highly conservative worst-case scenario. Even assuming that unobserved crimes for unreleased defendants would have been twice as high as estimated, the Laqueur–Copus algorithm could still release about 50% of defendants.¹⁶¹ This would have more than doubled the number parole boards freed.¹⁶²

Parole-eligible individuals, of course, constitute only a subset of the total imprisoned population. Most criminal defendants must serve a substantial period of incarceration—often years—before they first become eligible for parole.¹⁶³ Thus, an Algorithmic Abolitionist policy that implemented the Laqueur–Copus algorithm only for parole eligible defendants, and only on the current parole eligibility schedule, would reduce incarceration. But it would fall short of reducing total incarceration by 60 or 80%. Many people would languish in prison while they awaited their first chance to be released by the algorithm.

Better, then, to institute Algorithmic Abolitionist sentencing from the get-go. Rather than waiting for months or years, an optimal Algorithmic Abolitionist policy would release low-risk offenders as soon as possible—immediately following conviction.

If a Laqueur-Copus style algorithm performed similarly at sentencing as it does for parole, the abolitionist effect would be impressive. Incarceration would be reduced by a full 80% as everyone deemed low risk by the algorithm was immediately released.¹⁶⁴ But is such comparable performance at sentencing likely? At this stage, lacking direct empirical evidence on the question, it is hard to say for certain.

But here are some arguments, based on what we do know, suggesting that Algorithmic Abolitionism at sentencing might look similar to Algorithmic Abolitionism at parole. Recall first that incarceration has a criminogenic effect, which, as sentences lengthen, can outpace its initial crime-reducing effect.¹⁶⁵ Shifting algorithmic release from parole to the moment of sentencing avoids years of potential criminogenesis. This means that average crime risk among those algorithmically assessed at sentencing would be lower than those assessed by Laqueur and Copus at parole. The result would be an even greater expected rate of release than Laqueur and Copus report.

Some may object that this is flawed reasoning. True, shifting algorithmic release from the moment of parole to the moment of conviction does avoid prison’s criminogenic effects for the released individuals. But it also eliminates prison’s crime-reducing effects for those individuals—especially those strongest effects at the beginning of the sentence.¹⁶⁶ Thus, one might argue, the shift in timing would increase average crime risk in the target population and therefore reduce the rate of release.

This is not quite right. Not everyone has the same risk of committing future crimes. And the Algorithmic Abolitionist sentencing algorithm would release only those individuals it judged to have little or no risk of committing future crimes. Incarceration, by and large, could not possibly reduce those individuals’ risk of committing future crimes. Most would have no risk to reduce. Thus, for them, incarceration would be purely criminogenic, and the above-described argument for the shift to release at sentencing holds.

Indeed, algorithms could reap a double-bonus from heterogeneity. Even among individuals with nontrivial crime risk, incarceration’s crime-reducing effect may be highly heterogenous.¹⁶⁷ Strong specific deterrent effects from incarceration might either be correlated with high crime risk or otherwise predictable by algorithms. If so, a sentencing algorithm like Laqueur and Copus’s could release even more people without increasing crime. It would accomplish this by focusing incarceration on those for whom it would have an unusually large crime-reducing effect. All of these factors suggest that a Laqueur-Copus style algorithm, applied to sentencing, could enable release rates similar to the 80% it would release at parole—and perhaps more.

There are, however, other countervailing factors worth considering. Not all incarcerated persons in New York are eligible for parole.¹⁶⁸ The population on which the Laqueur-Copus algorithm was evaluated might therefore be fundamentally different from the full incarcerated population. But because New York only instituted determinate sentencing in the late 1990s and 2000s, many non-parole-eligible incarcerated persons in fact committed the exact same crimes as parole eligible ones.¹⁶⁹ And even after the change, a full 40% of New York state prisoners are serving parole-eligible sentences.¹⁷⁰ This shift in sentencing law also suggests that parole-eligible incarcerated persons in New York might be older on average than non-parole-eligible ones. The crime-reducing effects of age are well documented.¹⁷¹ On the other hand, the same legal shifts suggest that people eligible for parole have been incarcerated substantially longer than those not eligible. Thus, the crime-reducing effects of ageing are countervailed to at least some extent by the above-discussed criminogenic effects of long sentences. The best way to see how all of these effects net out in Algorithmic Abolitionist sentencing is to try it. Even a small-scale experiment would be highly informative.

Here is another important question about Algorithmic Abolitionist sentencing: What effect would it have on general, as opposed to specific, deterrence? Under the incarceration policies contemplated here, large proportions of people convicted of crimes would never see jail time. Even if algorithmic precision in the allocation of incarceration kept crime from rising among the convicted, what about everyone else? As discussed above, incarceration appears to have an important general deterrent effect that goes well beyond those individuals actually caught and sentenced.¹⁷² The literature thus shows that, if incarceration is indiscriminately reduced, its deterrent effect is likewise reduced, and crime increases.¹⁷³ Would the huge reductions in total punishment wrought by Algorithmic Abolitionism, thus, massively reduce general deterrence and increase crime?

Not necessarily. Algorithmic Abolitionism is not like indiscriminate abolitionism. While it does reduce incarceration dramatically, it does not do so uniformly. The reductions come from releasing the large share of low-risk defendants who, under the status quo, are imprisoned. But the status quo also frees many among the comparatively small set of individuals who will predictably commit further crimes. Under Algorithmic Abolitionism, they would be incarcerated. Thus, for this latter group, the switch to Algorithmic Abolitionist policy increases, rather than decreases expected punishment. Moreover, individuals performing a reflective cost-benefit analysis of their potential crimes seem much more likely to fall into the latter category than the former. If this is right, and potential defendants realize it, then Algorithmic Abolitionism would, at a first cut, increase, not decrease general deterrence.

What about in equilibrium? Consider the rational potential crime committer at the margin. Suppose the relevant algorithm would score him as low risk, and he knows this. Such knowledge, it should be noted, could be quite hard to obtain, given the “black box” nature of many machine-learning algorithms.¹⁷⁴ But if the marginal individual could predict his low risk score, this would decrease the punishment he expected from crime. Then, if rational and inclined to crime, he would commit more crimes.

But we are not yet in equilibrium. Algorithms need not be static. If the initial version of an algorithm induced some marginal individuals to commit more crimes, some of them would be caught. Then, their data could be used to further train the algorithm, such that similar marginal individuals would no longer be scored low risk. If they knew this, they would then expect more punishment and revert to a lower level of crime commission. This would then produce more data on which to update the algorithm. And so on until an equilibrium was reached. Formal microeconomic models of the expected equilibria for various Algorithmic Abolitionist policies would be useful, but they are beyond the scope of this Article.

Here is one more factor favoring the idea that Algorithmic Abolitionism would maintain, rather than reduce, general deterrence. Some evidence suggests that dedicated crime-committers overestimate new technology’s ability to single them out for increased punishment. For example, economists of crime have found that the general deterrent effect of DNA databases is larger than pure rational-actor models would predict.¹⁷⁵ Potential offenders may believe that DNA technology is even better at solving crimes than it really is. The same effect could hold for committed bad actors’ estimate of algorithms’ ability to single them out.

Algorithmic Abolition could also maintain general deterrence by shifting criminal policy toward detection, rather than punishment. Robust criminological evidence shows that increasing the certainty of apprehension for a crime is a much stronger deterrent than increasing the severity of punishment.¹⁷⁶ Thus, if Algorithmic Abolitionism’s large reductions in incarceration tended to increase crime, this could be offset by modest increases in apprehension. Such increases in apprehension rates could be achieved using the Algorithmic Abolitionist policing policies described below.¹⁷⁷ And because—as also described below—Algorithmic Abolitionism can, as a baseline, eliminate such huge shares of policing, modest deviations from that baseline would blunt the overall abolitionist effect only moderately.

B. Police

The previous Section described Algorithmic Abolitionism’s potential as applied to incarceration. This Section turns to policing. It estimates the potential effect of Algorithmic Abolitionist policies for three major police functions: traffic enforcement, street stops, and patrols. These functions are selected for two reasons. First, they constitute the vast majority of the public’s unwanted interactions with the police and, thus, the majority of potential harm from disruptive and violent policing.¹⁷⁸ In 2015, 10.8% of Americans over age sixteen reported having an interaction with the police that they did not request.¹⁷⁹ 8.6% reported that the type of unsolicited interaction was a traffic stop where they were the driver, and another 2.4% reported a traffic stop where they were the passenger. A full 1% of Americans reported being stopped by the police on the street.¹⁸⁰ These interactions arose from police patrols, not, for example, from the execution of preexisting warrants.¹⁸¹ By contrast, just 1.1% of Americans reported any other type of unwanted police interaction.¹⁸² That comprises arrests—including those with a warrant—that did not start with traffic or street stops.¹⁸³ The second reason for focusing on these police functions is that, as with incarceration, algorithms already exist that can allocate them much more efficiently than humans. Significant abolition is thus possible using existing technology.

Begin with street stops. Algorithmic Abolitionist policies could, plausibly, eliminate the vast majority of them. In 2016, Sharad Goel, Justin M. Rao, and Ravi Shroff developed an algorithm, trained on data from real New York street stops, to identify individuals likely to be carrying illegal weapons.¹⁸⁴ As with other crime prediction tasks, human assessments here are extremely inaccurate. In New York, between 2011-12, 43% of stops made by police had less than 1% probability of revealing an illegal weapon.¹⁸⁵

Thus, as with other risk estimation tasks, superior algorithmic performance could enable significant amounts of abolition. Within the set of actually made New York stops, the Goel et al. algorithm identified a subset of just 6% stops from which the majority of weapons could have been predictably recovered.¹⁸⁶ This suggests that, if applied population-wide—to both those people whom the New York police actually stopped, and those who it did not—an algorithm-based policy might have reduced stops by 88% while recovering the same number of weapons.¹⁸⁷ At a bare minimum, the Goel et al. results suggest that roughly 42% of stops could be abolished without meaningfully reducing weapons recovery.¹⁸⁸ Even better, implementing this Algorithmic Abolitionist policy would not require officers to fire up a computer application before making every stop-and-frisk decision. Instead, the Goel et al. algorithm reveals a set of simple observational heuristics that nearly replicate its performance and can be easily deployed in the field.¹⁸⁹

Algorithmic Abolitionist policymaking could reduce traffic stops even more. Much of this abolition could be accomplished without advanced algorithms. According to a recent nationwide survey of police, speeding is by far the most common reason for traffic stops.¹⁹⁰ Many jurisdictions already employ speed cameras in certain locations that automatically detect speeding and send citations to the car’s registered owner.¹⁹¹ The Algorithmic Abolitionist version of this existing policy is to implement speed cameras broadly enough that human enforcement is no longer necessary—and then to ban it. This might sound expensive. But, on the contrary, robotic traffic enforcers are much cheaper than human ones.¹⁹²

The automation of other traffic enforcement functions would require algorithmic technology, but nothing novel. The next most common reason reported by police for stopping vehicles is equipment violations.¹⁹³ Algorithms are already highly effective at correctly identifying the content of images of all kinds,¹⁹⁴ and they could be fine-tuned to identify the whole range of visually identifiable automotive equipment violations presently detectible by human police. Examples—including an algorithm that detects illegal window tints—already exist.¹⁹⁵ Other enforcement actions currently carried out via live police stops could be automated, too. Various algorithmic approaches already exist for detecting moving violations, including dangerous and impaired driving.¹⁹⁶

Indeed, if anything, algorithmic traffic enforcement might threaten to replace more than 100% of current traffic stops. Automated systems, after all, would detect all violations in their vicinity, not, as under the status quo, just those that occur in sight of a police officer. Likewise, automated systems might enforce violations strictly—citing everyone driving over the speed limit, rather than just those driving, say, more than ten-ish miles per hour over it. Insofar as this is a problem, solutions are available: Even if automated systems could cite many more traffic offenders than under the status quo, they need not. No rule of law requires that all detected violations result in a citation; even police often give warnings. With automated systems, enforcement could be set at desired levels in any number of ways. Systems might issue warnings to occasional violators, with citations accruing only to repeat offenders. Systems could be turned on and off at random, mimicking the presence or lack of enforcing officers at different times of day. Or they could be left on at all times, with the decision to cite or not determined probabilistically, to similar effect. All of these would likely be fairer than the status quo, wherein non-random, albeit inconsistent, enforcement affects racial and socioeconomic groups differentially.

What about policies that would not merely reduce unnecessary police stops, but eliminate unnecessary police presence entirely? How much Algorithmic Abolitionism could be possible for police patrols, broadly construed? There are two potential crime-controlling effects of patrols. The first is general deterrence. Empirical evidence shows that mere police presence—even with no citizen interactions—in areas that would otherwise experience crime has a significant deterrent effect.¹⁹⁷ Such presence reduces crime on net; it is not merely displaced to other geographic areas.¹⁹⁸ Patrols’ second crime-controlling effect is the incapacitation that results when police avert a crime in progress. For both effects, physical police presence in a given area has value only insofar as that area would otherwise experience high crime risk. In areas where crime is unlikely to occur, the disruptions, intimidation, and violence police encounters can cause constitute needless harm.

As with individual crime risk, humans are bad at predicting spatio-temporal crime risk. An Algorithmic Abolitionist approach can do better, directing police patrols to only that subset of places and times where they are truly needed. The approach here is the same as with Algorithmic Abolitionist incarceration policies: Begin with the lowest-risk places and eliminate police patrols up to the point at which crime would begin to rise.

How good are algorithms at predicting when and where crime will occur? In a 2015 randomized trial, researchers pitted an adapted earthquake prediction algorithm against the top human crime forecasters in the Los Angeles, CA and Kent, England police departments.¹⁹⁹ Even when compared with human forecasters using hotspot data and qualitative criminal intelligence, the algorithm did about twice as well.²⁰⁰ Other, newer, and more complex, machine-learning models might do even better. A deep neural network published in 2017 was trained to predict incidents of 31 types of crime in Chicago down to a spatial resolution of just a few feet.²⁰¹ It achieved almost 85% accuracy.²⁰²

How much abolition of police patrols can be achieved using modern machine learning? It is difficult to say with the same precision as was possible for Algorithmic Abolitionist incarceration policies. Unlike with incarceration, no available empirical evaluations specifically simulate the minimum amount of police patrols necessary to maintain status-quo crime rates. Nevertheless, plausible estimates of patrol reduction can be extrapolated from related empirical findings. These estimates are, however, necessarily more speculative than those based on studies attempting to measure algorithms’ abolitionist potential directly.

Algorithms’ potential for abolishing police patrols depends on the status quo distribution of policing: What percent of a given city’s neighborhoods are patrolled on a given day? And what percent of crimes are captured by those patrols? The more algorithms can successfully predict the same number of crimes in a more limited geographic—or temporal—range, the more abolition is possible.

In the randomized trial in Los Angeles and Kent, the algorithm was about twice as good as the best human forecasters at identifying hotspots with unusually high levels of crime. It could capture twice as many crimes using hotspots covering the same percentage of the city.²⁰³ This finding—that algorithms are roughly twice as good as humans at predicting geographic crime risk—is consistent with recent findings from Chicago.²⁰⁴ Using that estimate, then, one could design an Algorithmic Abolitionist policy expected to eliminate 50% of risk-targeted police patrols while maintaining the status quo deterrent effect. A simple version would be to dispatch patrols randomly to half of the geographies the algorithm selected.

This estimate of a 50% reduction in police patrols is not the maximum. It focuses just on the number of patrols that could be eliminated which would otherwise have been deployed to high risk hotspots. However, in many urban neighborhoods, police presence is not primarily allocated on the basis of risk.²⁰⁵ Substantially more abolition should therefore be possible when one includes the many geographies where police are not currently trying to patrol according to risk-based need.

To get a rough of the total abolitionist potential, however, further speculative assumptions are necessary. Suppose optimistically that, under the status quo, police patrols in the average city are present in the immediate vicinity of 30% of crimes when they happen. And suppose that to do this, police are daily deployed to neighborhoods covering 60% of the city’s geography. Studies using data from Baton Rouge and Chicago suggest that algorithmically directed patrols could achieve the same crime coverage while deploying to only about 5% and 10% of the city, respectively.²⁰⁶ An Algorithmic Abolitionist policy using these algorithms could—under these assumptions—eliminate between 83% and 92% of a given city’s police patrols with no increase in crime.

C. How to Design an Algorithmic Abolitionist Policy

The idea that algorithms might inform the administration of criminal justice is not novel. Many jurisdictions already use at least some algorithms resembling those described above.²⁰⁷ Pretrial incarceration is a common application. Over 60% of the U.S. population now lives in a jurisdiction that uses some kind of risk assessment tool in the pretrial incarceration process.²⁰⁸ Yet despite this, the dramatic abolitionist results described above have not emerged anywhere. Why not?

The answer is that no existing algorithmic policy can rightly be called Algorithmic Abolitionism. Each suffers from at least one of three important flaws. Each flaw is straightforwardly remediable. This Section describes these common failures and, based on them, articulates a set of principles for designing Algorithmic Abolitionist policies.

The first widespread problem is that no existing algorithmic criminal justice policy is abolitionist. Put simply, none are designed around the core Algorithmic Abolitionist goal: reducing as much prison and policing as possible without increasing crime. Consider Virginia, a rare state that adopted risk scores for sentencing with the explicit goal of reducing incarceration—rather than, for example, reducing crime.²⁰⁹ Yet even in this most propitious of circumstances, the legislature set a goal of reducing imprisonment by just 25%.²¹⁰ This is a substantial figure, and a laudable goal, but it falls far short of the abolitionist frontier.

Here, then, is the first lesson: Algorithms will not achieve dramatic abolition by accident. Polices must be built for that purpose from the ground up. The details of policy design—the cut score at which a prisoner is released, the risk score at which a neighborhood is patrolled—determine the results. For Algorithmic Abolitionism to be achieved, these must be calibrated for maximal elimination, rather than around lesser goals.

The second widespread defect is likewise one of policy design. Virginia’s algorithmic sentencing law set a goal of reducing incarceration by 25%.²¹¹ But it did not achieve even that. Indeed, a recent analysis shows that the policy generated “no net change in the incarceration rate or length of sentence[s].”²¹² The same was true in Kentucky following its adoption of risk scoring for pretrial detention. As with Virginia, despite an explicit legislative goal of lowering incarceration, the policy “led to only a trivial increase in pretrial release.”²¹³

The problem was not with the algorithms. Rather, it was that even in these facially algorithmic policies, algorithms were given very little role to play. In both states, judges were allowed to simply ignore algorithmic recommendations. In Virginia, for example, judges were shown both a given defendant’s risk score and an indication of whether it was low enough for the algorithm to recommend release.²¹⁴ But judges could decline to follow the algorithmic recommendation without consequence.²¹⁵ And decline they did. Judges either ignored algorithmic risk-based recommendations entirely or deviated from them so as to increase incarceration.²¹⁶

Thus, the second lesson for designing Algorithmic Abolitionist policies: Algorithmic decisions must have bite. What this means may vary by context. In the context of police patrolling, for example, algorithmic direction of patrol patterns should be mandatory just like patrol orders given by ranking human officers. In the domain of sentencing, there are various options for giving algorithmic determinations bite.

Here are three possibilities, ordered from least bite to most. First, judicial deviations from algorithmic sentencing decisions could be made costly. In federal criminal cases, for example, judges who deviate from the U.S. Sentencing Guidelines range must explain why in writing—and do so on penalty of reversal.²¹⁷ This approach appears to be relatively effective at disincentivizing deviation. Even after United States v. Booker²¹⁸ made Guidelines recommendations discretionary—constrained only by the written explanation requirement—deviations did not rise much.²¹⁹ Judges who wished to depart from algorithmic sentencing determinations could likewise be required to bear the burden of producing a written explanation.

Second, a prediction requirement could be added to the written explanation requirement. If judges believe they are better at estimating who will and will not commit crimes than machine-learning algorithms, they could be forced to put their money where their mouths are, supplying their own probabilities. Then, their aggregate accuracy could—as with algorithms—be assessed against real-world results. The judges could be either praised for their superior wisdom or held accountable for their hubris.

Finally, algorithmic determinations could be the final word—not subject to judicial override for any reason. For optimal results, courts should only impose this approach after public experiences with the more flexible approach convince citizens that judicial discretion, on net, does more harm than good. This policy design, of course, raises serious concerns about errors—the harmless defendant whom the machine sentences to years of imprisonment. But such errors are also rampant under the current system.²²⁰ And as just described, allowing humans to override algorithms increases, not reduces, them.²²¹ Thus, at some point it may be worth considering whether keeping humans in the loop actually serves humanity.

Returning to algorithmic policy design, the final widespread flaw—or at least potential flaw—in existing policies is technical. Where algorithms are already used in criminal enforcement, they are rarely the best available algorithms. As mentioned above, the COMPAS tool, for example, is now almost a decade and a half old.²²² Yet it is the primary algorithm used for pretrial risk assessment in at least eleven counties with a combined population of over 4.3 million people.²²³ As already discussed, in recent years, algorithmic capabilities have exploded across essentially every application. Most of the high-performing algorithms described herein use techniques popularized during that period.²²⁴ If older, widely used algorithms cannot predict crime risk with the same accuracy as these newer ones, then they cannot be used to abolish as much policing or prison.

Hence, the final lesson for designing Algorithmic Abolitionist policy: Use the best available technology, with a plan to update as technology improves. Note that this maxim will not inevitably imply discarding more traditional statistical models in favor of newfangled machine-learning approaches. For some questions—and some datasets—traditional approaches work as well as the newer ones.²²⁵ The point, however, is to test. Older algorithms should be pitted against new approaches, and the best performer should be adopted into public use. Great feats of abolition can now be achieved using modern techniques. There is no reason for policymakers to settle for less.

Existing experiments in algorithmically driven policing, as opposed to incarceration, have likewise, not dramatically reduced policing. But the reasons are now familiar. First, and most importantly, reducing policing was not these policies’ aim. Algorithmic policing experiments have instead generally sought to reduce crime.²²⁶ On this metric, algorithmic policing has perhaps seen more success than algorithmic sentencing. Two recent empirical evaluations of Chicago’s predictive policing program—from RAND and the University of Chicago Crime Lab—show that it did meaningfully reduce crime.²²⁷ But those evaluations also show that actual crime reduction fell far short of the algorithmic tools’ potential.²²⁸ The primary culprit was again policy design, not algorithms’ predictive power.²²⁹ Algorithmic determinations again lacked bite, so that many Chicago precincts were simply “not following the playbook” and “ignoring [the algorithm’s] recommendations.”²³⁰

III. Algorithmic Abolitionism Upends the Bias Debate

Despite algorithms’ considerable abolitionist potential, and despite prison and police abolitionists’ success in spreading their argument for elimination over reform, both progressives and abolitionists have vigorously opposed the use of algorithms in criminal justice.²³¹ In the past, they were perhaps right to do so. Their most common objections—to algorithmic bias and structural racial inequality—have been genuine concerns. Moreover, until now, the potential abolitionist upside of algorithms has not been widely understood.

But Algorithmic Abolitionism presents a new normative landscape. Even if racial-equality objections to previous uses of algorithms in criminal enforcement had merit, they lose their force as applied to Algorithmic Abolitionism. This is in part because such objections have, as a general matter, weakened as algorithmic design has improved. But it is also because an Algorithmic Abolitionist approach changes the normative terrain.

To that end, this Part outlines four normative arguments favoring the use of algorithms, despite longstanding objections based on racial justice. Three of them—the latter three—are novel contributions to the literature. The first argument is comparative. To the extent algorithms are biased, they are much less so than the best available alternative: human decisionmakers. This point has been made before,²³² but the Part shows how strong it is as applied to the state-of-the-art algorithms discussed herein.

The remaining three arguments are new. Algorithmic Abolitionist thinking is necessary to reveal them. The arguments arise when one considers the racial justice impact of policies that dramatically reduce levels of harmful outcomes—like being policed or incarcerated. Traditional thinking about algorithmic bias, by contrast, has been concerned almost exclusively with ensuring equal distributions of harm.

The first of these three novel arguments shows, counterintuitively, how concerns about racial distributions of harm and concerns about total harm from discrimination diverge. A levels-reducing policy can dramatically reduce the number of people being discriminated against, while maintaining or even worsening traditional distributional measures of fairness. To capture this insight, overlooked in existing accounts of algorithmic discrimination, the Part introduces a new, quantitative metric of algorithmic fairness.

The Part’s second new normative argument is about tradeoffs. It begins by endorsing the assignment of at least some normative weight to pure distributional, as opposed to absolute, measures of justice. This raises the question of how to trade distributional benefits against absolute goods—like reducing harm from discrimination or reducing total incarceration. Pure cost-benefit analysis, the Part argues, cannot satisfy the progressive critics of algorithms who prioritize distributional concerns. The Part therefore proposes a new framework for evaluating such trade-offs—one that stacks the deck in favor of distributional values. It then shows why the tradeoffs implied by Algorithmic Abolitionist policies would almost inevitably pass even such a stringent test.

The Part’s third novel normative argument relates to structural inequality. Structural inequality is sometimes argued to be the cumulative result of bias over time. But the Part argues that bias is not the sole—or even primary—driver of structural inequality stemming from criminal enforcement. Rather, levels are. Thus, the key to breaking cycles of structural inequality is reducing levels of harm—the main effect of an Algorithmic Abolitionist approach.

A. Comparative Bias

Arguments about racial bias dominate the debate over algorithms’ use in criminal enforcement.²³³ But they are also common in broader academic²³⁴ and popular²³⁵ debates about algorithms’ many uses. In the context of criminal enforcement, the dialectic inevitably runs as follows: First, a jurisdiction introduces some algorithmic tool into its criminal justice apparatus, promising greater efficiencies, improved fairness, and sometimes less policing or incarceration.²³⁶ Following introduction of the tool, high-profile academics or media organizations argue that the tool must not be used, because it is biased—i.e., that it treats one group (usually white Americans) better than another group (usually Black Americans).²³⁷ Then the algorithm’s creators rejoin that their model is, in fact not biased, so long as “bias” is correctly understood.²³⁸ Here, the debate explodes, but it is dominated by second-order arguments, built atop the initial ones: What is the right understanding of bias? Is it possible, even in theory, to produce an algorithm that would qualify as perfectly unbiased?

This Section does not seek to relitigate those well-worn disputes. Rather, it argues that, at least as applied to Algorithmic Abolitionist policies, they miss what matters most. For any use of algorithms—including Algorithmic Abolitionism—the policy-relevant question is not, “is this algorithm racially just?” It is instead, “is this algorithm more or less racially just than the alternatives?” The former, widely debated, question turns out to be a difficult one. The latter often turns out to be easy.

Determining whether an algorithm is racially biased, full stop, is less tractable than it sounds. There are many dimensions along which algorithms could be equal or unequal, and the answer thus depends on what one means by “biased.” Three popular measures of bias cover much of the available conceptual ground. First, there is “predictive parity.” Here, an algorithm is unbiased if a given risk score implies the same set of expected outcomes for members of different racial groups.²³⁹ That is, if an algorithm labels some individuals, say, “highly likely” to commit a crime, the actual rate of crime commission is the same for the Black subgroup as for the white one.²⁴⁰

Some argue that predictive parity does not do enough to promote equality. Even when an algorithm has achieved it, different groups may experience different kinds of errors at different rates.²⁴¹ Different kinds of errors can have different costs.²⁴² Incorrectly classifying someone as low risk—and thus releasing them—imposes little cost on the released person.²⁴³ Conversely, incorrectly classifying someone as high risk—and thus incarcerating them—is very costly to them.²⁴⁴ Thus, some bias-oriented critics of algorithms argue that fairness means achieving not just predictive parity, but rather “error-rate” parity. Under this theory, an algorithm is unbiased if members of different racial groups are subjected to costly errors—for example, unnecessary incarceration—at the same rate.²⁴⁵

These two conceptions of fairness—predictive parity and error-rate parity—are not only different; they are in direct tension. For most populations, it is mathematically impossible to achieve both simultaneously. The only exception is the exceedingly rare circumstance in which the two populations have identical underlying distributions of actual crime.²⁴⁶

Some theorists go further yet, arguing that algorithmic fairness means equalizing outcomes. Under this conception, an incarceration algorithm is racially unbiased only if it recommends imprisonment at the same rate for all racial subgroups.²⁴⁷ Then, the racial composition of the incarcerated population will mirror the racial composition of the total population. This, too, is almost always impossible to accomplish simultaneously with either of the other two versions of fairness.²⁴⁸

Thus, critics are right. Essentially all algorithms are biased, if “biased” means “unable to achieve parity under all three conceptions.” For essentially any real-world crime-risk algorithm, the achievement is mathematically impossible. And even if one agreed to evaluate algorithmic fairness using just one of the conceptions, the dispute over which one mattered most would remain.²⁴⁹

But maybe the question of whether an algorithm is biased is the wrong one. We could instead ask how much bias a given algorithm displayed—and compared to what? Shifting the inquiry in this way—from absolutist to comparative—causes the impossibilities just described to melt away. Although an algorithm cannot be perfectly unbiased according to multiple theories simultaneously, it is perfectly possible for one to improve according to multiple theories simultaneously.

And algorithms do improve, a lot, compared with the alternative. In criminal enforcement, that alternative is human decision making. Humans are extraordinarily biased. Algorithms regularly perform much better, no matter how one conceives of fairness.

Here is why. Consider a straightforward model in which Black defendants are unjustly incarcerated at higher rates than white defendants. Suppose that human judges believe that all defendants’ risk of crime is higher than it actually is and incarcerate them according to that (mistaken) assumption. This produces suboptimally high levels of incarceration, but it does not induce a racial disparity. However, if the upward skew of human risk assessments is uniformly even bigger for Black defendants than for white ones, the posited unjust disparity results.²⁵⁰

Now consider the effects of an Algorithmic Abolitionist incarceration policy on this model. Suppose the algorithm is, unlike humans, very good at identifying individuals who will commit crimes. Imagine that, among those it predicts will commit them, 99 percent do—irrespective of racial group. Likewise for its predictions of who will not commit crimes. The algorithm has achieved predictive parity—perfect fairness according to one of the three accounts.

Having achieved predictive parity, this algorithm cannot also be perfectly fair according to either of the other conceptions of bias. But it can improve things—perhaps a lot—according to both of them. The human judges’ uniformly stronger bias against Black defendants than against white defendants implies serious unfairness in the distribution of errors. Black defendants are much more likely than white ones to be erroneously imprisoned. By eliminating the additional human-induced penalty for Black defendants, the well-calibrated algorithm would necessarily reduce—if not eliminate—this disparity. The quantum of improvement is proportional to the quantum of irrational human discrimination that existed in the first place.

Likewise for the “equal outcomes” theory of fairness. The human judges’ overestimates of risk were uniformly larger for Black defendants than white ones. Thus, the algorithm’s elimination of human-induced error necessarily brings the rates of Black and white incarceration closer together. Here again, the quantum of algorithmic improvement is inversely related to the quantum of human discrimination.²⁵¹

The key insight here is that improved accuracy can often be a powerful driver of fairness. It should be no surprise, then, that the accuracy-improving algorithms powering Algorithmic Abolitionism would reduce bias as compared with the human alternative. And they would often do so according to multiple theories of bias simultaneously. The Laqueur–Copus algorithm, for example, could be used to eliminate present racial disparities in incarceration rates entirely, even while radically reducing total incarceration rates.²⁵² This would constitute perfect fairness under an equal outcomes conception. Though Laqueur and Copus do not report further equality metrics, their algorithm would likely make improvements under other theories of fairness, too. Laqueur and Copus report that achieving equality of outcome requires almost no tradeoff in terms of either the total reduction in incarceration or the resulting rates of crime.²⁵³ This is at least consistent with improvements under both the predictive-parity theory and the equal-errors theory. Thus, the Laqueur–Copus algorithm might well improve fairness under all three conceptions simultaneously.²⁵⁴ The Kleinberg et al. pretrial detention algorithm shows similar results.²⁵⁵ So too for Goel et al. They identify especially large errors in police risk assessments of nonwhite community members, which their algorithm reduces.²⁵⁶ So, by shifting from an absolutist perspective to a comparative one, we see that algorithms can make big improvements, no matter which theory of fairness one prefers.

This is a convenient shift. But is it normatively desirable? Bias is pernicious. Why should we accept a system that is biased at all, even if less so than the human alternative?

It might be better to ask why the comparative perspective shouldn’t supplant the absolutist one in the debate over algorithmic bias. The absolutist perspective has never been a useful tool for evaluating social policy, including as it relates to racial justice. Every great historical victory for civil rights, racial equality, and antidiscrimination has been a comparative one. None have been absolute—or even close to it. Emancipation did not end racial subordination in economic, social, or political life. The Civil Rights Act did not end discrimination. Affirmative action did not end unequal access to education or employment. But all of these were serious improvements over the alternatives, and worth implementing for that reason. Algorithmic policies should be judged no differently.

B. Bias, Levels, and Distribution

As just described, Algorithmic Abolitionist policies are likely to reduce bias, according to one or more major theories, compared with the alternative. But this is not a logical guarantee. It is in principle possible that some Algorithmic Abolitionist policies could leave bias unchanged according to some theories. Or, in the worst case, they might exacerbate it according to some. Should such policies be automatically rejected? This Section argues that they usually should not—at least not if they are genuinely Algorithmic Abolitionist.

The reasons have to do with levels. Even if an Algorithmic Abolitionist policy made the distribution of incarceration or policing more racially unfair, that would not be all it did. Algorithmic Abolitionist policies also substantially reduce the levels of policing or incarceration at which they are targeted.

Such reductions in levels give rise to new and important normative arguments—some of them surprising. This Subsection explains two, and the next Subsection adds a third. Each argument is designed to appeal to a different set of normative priorities.

For the sake of clarity, parsimony, and easy math, the arguments are made first using a thought experiment. After the thought experiment has made the arguments’ structure clear, the real-world Algorithmic Abolitionist policies described above are reintroduced and evaluated. The results, it will be argued, are the same.

Here is the thought experiment: Imagine a country that imprisons 1 million of its residents.²⁵⁷ Suppose 1 in 3 are Black (a disadvantaged group in the country) even though only 12% of the nation’s population is Black.²⁵⁸ Under the “equal outcomes” conception of bias, Black citizens are therefore incarcerated at nearly three times the fair rate. White citizens, by contrast, constitute just 30% of the prison population—half of their share of the general population.²⁵⁹ Now imagine a hypothetical Algorithmic Abolitionist sentencing policy that would, without increasing crime, reduce the standing number of incarcerated citizens from one million to just 100,000. Suppose, however, that under the policy, 40,000 of the 100,000 would be Black, and 30,000 would be white.

Here, then, is a policy that significantly increases bias—by 10 percentage points²⁶⁰—as evaluated using a traditional “equal outcomes” metric. That is, the share of incarcerated people who are white remains the same. But the share of incarcerated people who are Black increases—both as compared with whites and as compared with the share of the total population that is Black. For good measure, suppose that the policy would make bias worse by a similar amount according to the other theories—predictive parity and error-rate parity—as well.

1. Bias-Impact: A New, Levels-Linked, Measure of Fairness

Begin narrowly—with racial justice alone. Ignore all other considerations of potential normative import, including the total reduction in incarceration. Does our hypothetical policy make the imaginary country better or worse, just from the perspective of racial justice? According to traditional metrics of algorithmic bias, the answer is worse—by 10 percentage points. But the traditional metrics miss something crucial. They focus on “bias,” in the most literal sense—racial differences in the distribution of a set of costly outcomes. But they ignore the size of the set. They ignore the most important costs traditionally associated with discrimination—the harm individual people suffer when subject to a bad outcome because of one’s race. These two desiderata can vary independently. Thus, a policy can simultaneously increase the bias, as measured using existing tools, while also radically reducing the discriminatory harm suffered by individual people.

The thought experiment’s hypothetical policy does exactly that. True, within the set of those incarcerated, the policy substantially increases the percentage who are Black, worsening distributional bias. But because of the policy’s reduction in levels, the category of “incarcerated persons” now contains many fewer people. As a result, not only are many fewer Black people incarcerated under the new policy than under the status quo, but many fewer Black people are incarcerated unfairly because of their race.

Here are the numbers: Under the status quo, something like 210,000 Black citizens of our imaginary country are unfairly incarcerated each year, according to the “equal outcomes” conception of bias.²⁶¹ This, again, is not total Black incarceration. It is just discriminatory incarceration—the excess beyond what the fair rate would dictate. By contrast, under the hypothetical abolitionist policy—which is more biased, according to traditional measures—the number of Black Americans unfairly incarcerated nevertheless falls dramatically. Just 28,000 Black individuals are incarcerated above the figure the fair rate would dictate.²⁶² That is, nearly 200,000 fewer Black Americans suffer from racially discriminatory incarceration each year. The total harm from discriminatory incarceration falls by 87%. One might formalize these observations by saying that, although incarceration has become more racially skewed, the discriminatory impact—the number of people treated wrongfully—has radically decreased.

To capture this insight and encourage its application elsewhere, I propose a new, general-purpose quantitative measure of algorithmic fairness: “bias-impact.” Bias-impact is importantly distinct from extant measures like equal outcomes, predictive parity, and error-rate parity. It measures how a new policy changes the amount of discriminatory harm suffered by members of a disadvantaged group. It therefore shifts the focus away from the thing that every existing measure of algorithmic fairness treats as central: imbalances in the distribution of harmful outcomes. Such distributional imbalances are not categorically irrelevant to bias-impact. Reducing them can be one route to reducing discriminatory harm. But, as our hypothetical policy illustrates, bias-impact can also be reduced by reducing the total societal level of the harmful outcome. Locking many fewer people up implies locking many fewer people up unfairly.

Formally, the bias-impact of a proposed policy can be quantified using the following equation: B = ((A1 – F1) – (A2 – F2))/((A1– F1)) .²⁶³ B is bias-impact: the amount (expressed as a percentage) that a new policy reduces the number of people unfairly subjected to a harmful outcome. A₁ and A₂ are the number of people in the disadvantaged group (Black citizens, in our example) actually subjected to the harmful outcome under the status quo and the new policy, respectively. F₁and F₂are the number of people with the protected characteristic who would be subjected to the costly outcome if each respective policy were fair. Note that both the A and F terms are themselves functions. Each must be defined according to one’s preferred theory of algorithmic fairness. Under the equal outcomes conception, A is equal to the total number of people in the disadvantaged group times the incarceration rate for that subpopulation. F is equal to the number of people in the disadvantaged group times the incarceration rate of the whole population. Under the predictive parity conception of fairness, by contrast, A and F would be defined in terms of each policy’s positive predictive value for different subpopulations.²⁶⁴ For error-type parity, they would be defined in terms of sub-populations’ risks of false positives.²⁶⁵

Algorithmic abolitionist policies bear a special relationship to bias-impact: they reduce it almost mechanically. The whole idea of algorithmic abolitionism is to reduce levels of harm using more accurate decision tools. Reducing the level of policing and prison means reducing the number of people unfairly policed or incarcerated—a bias-impact reduction under the equal outcomes conception. And accomplishing that reduction in levels via improved accuracy means reducing errors—both total and false positives—and thus reducing bias-impact under the other conceptions. This relationship holds so long as the Algorithmic Abolitionist policy makes meaningful improvements in levels and accuracy for members of all racial groups.

Buy why should reductions in bias-impact trump increases in bias as measured using existing tools? Perhaps bias-impact is a poor measure of what we really care about. Maybe it is even a deceptive one. Skeptics may argue that bias-impact hides a kind of bait-and-switch, shifting the fairness debate to another topic entirely. After all, the existing measures of algorithmic fairness described above are all denominated in distributions, not absolute amounts of harm. Perhaps there is a reason for that. Perhaps the wrongfulness of invidious bias just is about distributions, not about the individual harms stemming from biased decisions.

This objection gets things mostly backward. Consider a paradigmatic example of wrongful discrimination: the child excluded from a high-quality, majority-white school because she is Black.²⁶⁶ Surely her exclusion is wrong, first and foremost, because of how it affects her, not because of how it affects some distribution. She is barred from receiving a quality education and, moreover, has suffered a serious dignitary harm. These are bad outcomes regardless of how her exclusion affects the distribution of Black and white students in the school. Indeed, suppose the girl’s exclusion somehow increased Black representation at the school. Perhaps her case triggers white flight, leaving a majority-minority school behind. This would have no bearing at all on the wrongfulness of her treatment. Thus, as sophisticated proponents of traditional fairness measures recognize,²⁶⁷ bias’s fundamental unit of normative analysis is its effect on people. We care foremost about the material harm of the costly outcome and about the moral harm of its being imposed because of race. These are exactly what bias-impact quantifies.

This is not to say that distributional measures of fairness are always normatively irrelevant. Often, they are relevant, but only for instrumental reasons. A skewed distribution can be good evidence that many individuals are suffering the kinds of harms just described.²⁶⁸ But as bias-impact shows, unskewing the distribution is not necessarily the only—or even best—way to eliminate those harms.

Skewed distributions are also normatively relevant if they generate additional harmful discriminatory acts. For example, unfair distributions of policing and incarceration might lead to inaccurate stereotypes about certain groups’ criminality. This could generate, for example, additional discriminatory exclusions from employment. Alternatively, unfair distributions of training and mentoring might lead to “statistical discrimination.” There, faced with two apparently identical candidates of different racial groups, an employer might assume that the member of the disadvantaged group has fewer unobservable skills.²⁶⁹ Statistical discrimination is especially pernicious because it can be actuarily sound—though unjust. The skewed distribution of training and mentoring might actually have changed the disadvantaged group’s average levels of unobservable skills.

But here again, making distributions fairer is not the only way to solve the problem. Reducing bias-impact may work just as well. For a stereotype or statistical discrimination to take hold, the outcome in question must be sufficiently common to be worth taking note of. Lightning is more likely to hit taller objects than shorter ones.²⁷⁰ Yet there is no stereotype that walking next to a tall friend in a thunderstorm is a risky proposition. Nor does it seem likely that life insurance companies statistically discriminate against the tall to account for their increased risk of death by lightning strike. The rarity of the outcome makes its skewed distribution irrelevant. This is an extreme example of a low probability event. But it illustrates how reductions in bias-impact could, just like distributional improvements, mitigate the harms of stereotyping and statistical discrimination.

There may also be reasons to care about unfair distributions intrinsically, rather than instrumentally. One such reason might be the expressive power of the law.²⁷¹ A skewed distribution—in policing, incarceration, or elsewhere—disfavoring historically oppressed groups communicates whom society values. Figuring out what, exactly, law expresses may often be complicated. For example, if the skewed distribution was imposed as part of an Algorithmic Abolitionist policy intentionally designed to reduce bias-impact, the message is ambiguous. Nevertheless, the point stands. In some situations, there will be reasons to assign freestanding normative value to fair distributions, as traditional algorithmic fairness measures do.

Thus, bias-impact captures the great majority of what matters most about bias better than traditional distributional measures. It captures the harms from discrimination to the real people who actually experience them. It also captures much of what has historically mattered about fair distributions, since distributional metrics were—in a pre-Algorithmic Abolitionist world—good proxies for individual discriminatory harm. Only in those limited contexts where fair distributions matter, qua fair distributions, might a tension arise between reducing bias-impact and improving traditional measures of algorithmic fairness. “Might” is important here. Usually, as already discussed, Algorithmic Abolitionist policies will improve according to both metrics simultaneously. But when they do not, it will be necessary to make trade-offs between reductions in bias-impact and improvements in distributional fairness. The next Section addresses how best to navigate such tradeoffs.

2. Trans-Theoretic Tradeoffs Between Levels and Distribution

Even now, some bias-wary progressives may remain unconvinced. Perhaps, despite the arguments above, they reject bias-impact as the sole measure of fairness because they significantly value pure distributional fairness, irrespective of how it affects individual people. Then, an Algorithmic Abolitionist policy that worsened things according to a traditional measure of bias might be objectionable on fairness grounds, regardless of its bias-impact.

But even this would not be reason to object to the policy, full stop. Fairness is not the only thing that matters. And Algorithmic Abolitionist policies are not mostly about fairness. Instead, they are about massively reducing carceral harm by eliminating large shares of policing and prison.

Here, we widen the scope of the normative inquiry to consider desiderata beyond racial justice alone. The question now is whether and when a policy’s total benefits might be sufficiently large to overcome some worsening of distributional fairness. The possibility that algorithmically driven policies could present such tradeoffs is not new.²⁷² But in contrast with the existing literature, this Section eschews pure cost-benefit-analysis. Instead, it fashions a test under which trade-offs should be deemed acceptable even to individuals placing a very high premium on pure distributional justice.

Suppose one holds a mixed normative theory that assigns value both to the magnitude of a policy’s net benefits and to who gets them—that is, to distribution. Prioritarianism is one example of such a theory. Prioritarians insist that, in any analysis of a policy’s costs and benefits, greater weight must be assigned to those who were worse off under the status quo.²⁷³ Rawls’s Difference Principle is another example. It is even stricter—a kind of prioritarian-plus rule. It assigns lexical priority to the wellbeing of the worst-off members of society. ²⁷⁴ The Difference Principle holds that an increase in distributional unfairness is justified only if accomplished via whatever policy sets the highest “floor” of outcomes for the worst-off individuals.²⁷⁵

But one could go stricter, still, fashioning a test for policy trade-offs that stacked the deck even more strongly in favor of distributional justice. One could insist, though no mainstream normative theory requires it,²⁷⁶ on a “Super Difference Principle.” Under it, the ordinary Difference Principle must be satisfied. The inequality-increasing policy must have a higher floor of outcomes than alternatives. But that is not enough. Under the Super Difference Principle, the group that was the worst off under the status quo must get the biggest benefit from the new policy. This is a kind of hyper-prioritarianism, wherein the needs of those at the bottom of the distribution are given as much weight as necessary to outweigh all other groups’. The Super Difference Principle is so demanding that it may often be impossible, as a practical matter, to satisfy. This will be the case if the policy that most raises the floor on bad outcomes may give somewhat larger benefits to groups other than those most disadvantaged ex ante.²⁷⁷

Nonetheless, Algorithmic Abolitionist policies would almost always satisfy all of these constraints, up to and including the Super Difference Principle. Our thought experiment again illustrates why. Under the status quo, Black citizens are so overrepresented in the imaginary country’s prison population that, despite being a minority group, they outnumber incarcerated white citizens.²⁷⁸ Thus, the hypothetical Algorithmic Abolitionist policy would free thousands more Black people than white ones.

The story is the same—indeed, starker—viewed from the perspective of the risk of incarceration faced by different groups. Assume our imaginary country has roughly the population of the United States. Then, under the status quo, 0.92% of the country’s Black population is incarcerated at any given time.²⁷⁹ By contrast, only 0.17% of the white population is incarcerated—about five times fewer.²⁸⁰ But the hypothetical Algorithmic Abolitionist policy would bring Black incarceration risk down by 0.81 percentage points.²⁸¹ This is a much greater benefit than for white citizens, for whom risk would go down by only 0.15 percentage points.²⁸² The absolute benefit to Black citizens, in terms of reduced risk of incarceration, would thus be over five times greater than that to whites.

The hypothetical sentencing policy is thus prioritarian. The needs of the worst-off group under the status quo—Black Americans—get special attention. It complies with the ordinary version of the Difference Principle.²⁸³ As compared with the status quo, the worst-off group sees both its total incarceration and its per-capita risk of incarceration decline. Nor, as argued below, could any feasible alternative policy raise the floor on incarceration risk more.²⁸⁴ The sentencing policy even satisfies the Super Difference Principle. Black Americans see a larger reduction in incarceration—both total and per capita—than any other group.

These are nearly mechanical results of abolitionist—including Algorithmic Abolitionist—policies. Dramatically reducing the system’s burdens across the board automatically raises the floor for bad outcomes, and in doing so gives special priority to the worst off. And so long as the reductions are relatively evenly distributed the largest benefits will accrue to those whom the status quo harmed most disproportionately.²⁸⁵ Anyone holding a plausible theory that assigns value both to total net benefits and to their fair distribution should be satisfied.²⁸⁶

3. How Real Policies Fare

The previous Subsections used a thought experiment, with easy math, to demonstrate in principle how an Algorithmic Abolitionist thinking upends traditional debates over algorithmic fairness. They showed that even a policy that worsens racial bias according to traditional measures may nevertheless decrease it according to an alternative, richer measure. And they showed that any lingering concerns about fairness could be outweighed by non-distributional benefits, even when the weighing is severely constrained.

But this Article is not about thought experiments. It is about real Algorithmic Abolitionist policies that could be implemented using existing technology. The thought experiment was useful because it showed that these results all flow from a single source: large reductions in levels of carceral harm. The question, then, is whether the actual policies discussed herein reduce levels enough to make them normatively comparable to the hypothetical one.

None of the actual policies advocated herein could, like the hypothetical one, singlehandedly bring the American prison population down to 100,000. Nevertheless, some of them are very nearly as ambitious as the hypothetical one, at least within their own domains. Algorithmic Abolitionism for Terry stops could eliminate, at the high end, nearly 90% of them.²⁸⁷ And Algorithmic Abolitionism for traffic enforcement could replace an arbitrarily large proportion of traffic stops.²⁸⁸ Thus, even if these policies had no impact on bias, as quantified using traditional metrics, they could reduce bias-impact 90% or more. And even if bias in street stops doubled, as measured using existing metrics, bias-impact would still be reduced by about 80%.²⁸⁹

Likewise, non-fairness related benefits could be huge. Each year, these Algorithmic Abolitionist policies might spare tens of millions of Americans from unwanted—and often dangerous—traffic stops. That is hundreds of millions in the policy’s first decade.²⁹⁰ Over 2.5 million people could be reprieved from invasive and intimidating street stops.²⁹¹ As with the hypothetical policy, the benefits from these reductions would flow disproportionately to Black and Hispanic Americans, who make up more than 80% of stops under the status quo.²⁹² Improved accuracy in determining who to stop would skew benefits in further favor of Black and Hispanic Americans.²⁹³ That is because human decisionmakers systematically overrate these groups’ likelihood of carrying a weapon, as compared with whites.²⁹⁴ These improvements satisfy all of the above-discussed constraints on tradeoffs, including the Super Difference Principle.

The same goes for Algorithmic Abolitionist decarceration policies. Here, the magnitude of abolition is somewhat smaller. As discussed above, the highest-end estimate, given some assumptions, is that existing algorithms could be used to eliminate 80% of incarceration.²⁹⁵ At the risk of becoming repetitive, an 80% reduction in incarceration with no change in bias, would reduce bias-impact—the number of people discriminatorily incarcerated—by 80%. Pessimism about the capabilities of Algorithmic Abolitionist policies does not change the picture much. Even if a policy doubled the strength of bias—an implausible figure—bias-impact impact would still be reduced by about 60%.²⁹⁶ Likewise, even if only 40% of incarceration could be eliminated—the low-end estimate—and bias was held constant, bias-impact would fall 40%.²⁹⁷

And for the reasons discussed above in relation to the model policy, broad-based reductions in incarceration satisfy even an extremely strict fairness constraint on tradeoffs, like the Super Difference Principle. Black Americans are so systematically disadvantaged by status-quo policies that any broad-based reduction in incarceration benefits them more than any other group.

Thus, actual Algorithmic Abolitionist policies are not so different from the policy in our thought experiment. Analysis of the hypothetical policy simply made clear that policies implementing sufficiently large reductions in levels upend traditional distribution-focused objections. And though the estimates are wide as to how much policing and imprisonment the Algorithmic Abolitionist policy discussed here would eliminate, even the low-end figures are quite large.

C. Structural Inequality

Perhaps the real justice-oriented objection to algorithms is not to bias in predictions or outcomes, per se, but rather to the long-run effects of those initial inequalities. Abolitionist and progressive opponents of algorithms regularly charge that such tools “exacerbate” inequality and “reenforce” certain “structures” of disadvantage.²⁹⁸ This is a version of the algorithmic bias argument, but extended through time. And it is true—disadvantage can breed further disadvantage.²⁹⁹ Incarceration reduces one’s ability to legally earn a living, which in turn increases the probability that one will resort in the future to illegal alternatives, producing more incarceration, and so on. ³⁰⁰ Such cycles of amplification could possibly result in group disparities at the societal level.

Algorithmic Abolitionism again has a distinctive response to such concerns—one not available in the existing literature on algorithmic fairness. This is again because Algorithmic Abolitionist policies are, first and foremost, designed to reduce levels of penal harm. And yet again, reducing levels has unexpected and normatively relevant implications.

The crucial question here is how modest initial disadvantages imposed by the criminal justice system might cyclically reenforce themselves into large disadvantages for entire social or racial communities. Such compound harm does not always occur. If it did, then every group affected at all by the criminal justice system—i.e., every group in American life—would have already experienced a similar upward spiral. Rates of arrest, charging, conviction, and incarceration would then be very high—and roughly equally so—for all groups. But they are not. On the contrary, Hispanic Americans are incarcerated at rates over two and a half times higher than whites, and Black Americans are incarcerated at a rate over five times higher than whites.³⁰¹ Thus, the “structural inequality” critique of algorithms depends on a model under which initial harms compound for some groups, but not others.

Here is a simple, plausible, model on which such differential compounding could occur. Suppose that harmful interactions with the criminal justice system have some rate of reproduction. The COVID-19 pandemic has made this “R₀” concept familiar.³⁰² An R₀ of 1 means that a given harmful incident—in our case, a street stop, arrest, prosecution, term of incarceration, etc.—will, in expectation, cause one additional interaction of roughly equal harm. In that case, total harm will remain level, with each instance reproducing itself, but no more. As R₀rises above 1, the total number of instances begins to grow—perhaps exponentially.³⁰³ But as it drops below 1, the total number decays in the same way.³⁰⁴

What could modulate the R₀ of interactions with the criminal justice system, such that some, but not all, groups might experience runaway feedback loops? Differences in the ambient levels of criminal enforcement endured by different groups probably play a role. There are many plausible mechanisms by which such levels could influence R₀ rates. Consider, for example, the direct effect of an arrest on the arrestee’s employment prospects. If a given individual is arrested once and briefly detained, there is some probability that they can retain their job. Perhaps they can use sick days to cover the absence, or perhaps their employer will extend some lenience for a one-time occurrence. But if—due to higher ambient levels of enforcement—an individual is arrested twice in relatively quick succession, the odds of remaining employed may decrease dramatically. Then, a job-loss/arrest spiral may result.

There are other plausible mechanisms by which levels could influence R₀, as well. As discussed above, longer-term incarceration may be criminogenic for those exposed to it.³⁰⁵ This is partly because of peer effects. Spending lots of time with people who commit crimes may make you more likely to commit them.³⁰⁶ Such peer effects matter outside of prison, too.³⁰⁷ A young person whose friends have mostly been arrested for crimes will be more likely to be arrested than a young person with just one such friend.³⁰⁸ Here, because connections within social networks grow exponentially, linear increases in ambient enforcement will quickly push R₀upward.

Critics of algorithms in criminal enforcement raise another possible mechanism: High initial levels of enforcement in neighborhoods of color lead to the detection of a higher proportion of crimes there than elsewhere.³⁰⁹ This generates a false perception of the neighborhoods as dangerous, which may be used to justify unusually large police responses to additional reported crimes. Those larger responses push the share of detected crimes up even further, and so on.

Observe that all of these mechanisms by which R₀ is modulated are, first and foremost, functions of levels of policing and imprisonment, not of disparities in them. This means that, for certain groups, but not others, to experience spirals of carceral harms, disparities are necessary. But they are not sufficient.

To see why, observe that, for any two groups with an R₀ of less than one, the long-run result is convergence, not divergence. In both populations, carceral interactions decay and run to zero.³¹⁰ This is true even if there are major disparities in how the two groups are policed or incarcerated. Since levels drive R_0,disparities in the ambient level of enforcement only matter to the long-run trend when they push one—and only one—group’s R₀above one. If not—if both R₀s are either above or below one—the long-run result is equality.

Here is an example to make the point concrete. Consider again the employment-linked mechanism for modulating R₀. Suppose that initial enforcement levels are such that members of Black communities are arrested, on average, one time over the course of their holding a given job, and members of white communities are arrested on average 0.5 times. This is a huge disparity in enforcement. But supposing that it takes at least two arrests to lose a job—thus generating more crime, and then more arrests—the R₀ in both communities is below one. Then, assuming prior arrests are the only thing that generates additional crime and future arrests, arrest rates in both communities would run to the minimum. Under this model, it is levels, not disparities, that primarily drive long-run social outcomes.

Now we can see why Algorithmic Abolitionist policies could turn the standard “structural inequality” critique of algorithms on its head. Again, the point of Algorithmic Abolitionism is to use algorithms to significantly reduce levels of prison and policing. If, as discussed above, levels of criminal enforcement are what generate (or not) structural inequality, then Algorithmic Abolitionist policies are not likely to exacerbate it. On the contrary, Algorithmic Abolitionism could be a powerful force for eliminating structural inequality. Algorithmic Abolitionist policies would help push the R₀ of prison and policing in vulnerable communities below one. This could, in turn, “crush the curve” of spiraling disadvantage and promote long run racial equality.³¹¹

Of course, there is no guarantee that Algorithmic Abolitionism would, on its own, cure America’s racial or other structural injustices. Indeed, one must ask why, under the above-described account of inequality’s genesis, ambient levels of policing and prison are unequal in the first place. Racism, of course, is a major part of the answer. For centuries, race-based disadvantage—including in criminal enforcement—was written explicitly into the law.³¹² And even though most such laws have now been overturned, racially discriminatory decision making continues to pervade society.³¹³ Thus, even under Algorithmic Abolitionism, such baked-in factors would continue to push levels of policing and prison upward for members of racial minority groups.

The point, however, is that if high levels of policing and prison are what drive structural inequality, then policy approaches that reduce those levels will help fight it. Algorithmic Abolitionism is such a policy approach. Even if it is not a panacea, it is likely to be a force in the right direction. At a minimum, Algorithmic Abolitionist policies would help to “slow the spread”³¹⁴ of structural inequality. This would buy time for additional interventions—education, healthcare, housing, and other creative solutions³¹⁵—to be brought to bear. Algorithms could be a powerful tool here, too. They could, for example, help to identify the people who would benefit the most from additional help.³¹⁶

IV. Remarks on Other Normative Questions

The previous Part had much to say about one important and longstanding normative objection to algorithms—discrimination. It had much to say there because Algorithmic Abolitionist thinking offers several genuinely new insights into the problem of algorithmic bias. It also had much to say there because the possibility of racial injustice is perhaps the most common reason for opposition to algorithms, both in criminal justice and elsewhere.³¹⁷ The hope, therefore, is that Part III represents a comprehensive response to the most important normative concerns about Algorithmic Abolitionism.

There are, of course, other longstanding normative concerns about algorithms, some of which apply to Algorithmic Abolitionist policy. This Part raises several, but treats them much more briefly, and thus less satisfyingly, than Part III treated the bias-related objections. That is in part because one Article can only do so much. But it is also in part because Algorithmic Abolitionist thinking introduces fewer genuinely novel arguments into these other debates. The Part thus seeks principally to situate Algorithmic Abolitionism within each normative debate, allowing the principal debaters to compare it to the alternatives.

There are, however, a few unique Algorithmic Abolitionist insights here, too. The first has to do, again, with levels. For certain normative concerns—like privacy and surveillance—Algorithmic Abolitionist policies could greatly reduce total levels of costly interventions. Such reductions might produce large net normative improvements, even if the remaining interventions increased their intensity. The second Algorithmic Abolitionism-specific insight explored here is positive-sum bargaining. Algorithmic Abolitionist policies are nearly unique among criminal justice reforms in that they could be used to simultaneously reduce both crime and policing and prison. This could allow for agreements between political factions that are usually bitterly divided. These themes will recur throughout the Part.

A. The Political Economy of the Carceral State

The core idea of Algorithmic Abolitionism is that algorithms should be used to maximally reduce prison and policing without increasing crime. But despite abolitionists’ success at convincing many liberal voters of the value of that goal, not everyone agrees. Some would prefer to use algorithms in the opposite way, to further reduce crime while holding prison and policing constant—or even increasing them. ³¹⁸ Insofar as this “tough-on-crime” sentiment is a powerful political force, perhaps algorithms in policing should be vigilantly opposed, irrespective of their abolitionist potential. The worry is that, if they are not opposed, they will inevitably be coopted for pro-carceral, rather than anticarceral, purposes.

Certainly, the desire to reduce crime is a powerful political force. And often for good reason. Crime is extremely harmful, with the harm falling disproportionately on the poor and communities of color.³¹⁹ There are also, of course, bad reasons to support increased crime-prevention efforts—for example if one was motivated by racial animus to impose more harms from policing on these same communities. But even for those who prioritize reducing carceral harm over reducing crime, the threat of tough-on-crime politics should not be sufficient reason to uniformly oppose algorithms in criminal enforcement. At least not when Algorithmic Abolitionism is an option.

The reason is that, unlike almost any other criminal justice reform policy, Algorithmic Abolitionism can improve on both fronts simultaneously. Laqueur and Copus provide a useful illustration. Recall their estimate that, holding crime constant, their algorithm could be used to release as many as 80% of parole-eligible prisoners without increasing crime.³²⁰ They also give an estimate of the inverse: the amount of crime their improved risk assessments could eliminate without increasing incarceration. Under the status quo, 33% of parolees were rearrested within three years.³²¹ An algorithmically driven policy could reduce that figure to as few as 10%, a nearly 70% reduction in crime.³²²

This, then, represents the “deal space.” Both groups could agree to some large, but sub-maximal policy that made both sides better off simultaneously. Suppose, lacking the political power to act unilaterally, abolitionists had to agree to an Algorithmic Abolitionist policy that reduced incarceration by “only” 60, 50, or 30%. Any of these would still be huge, even by the heady standards of abolitionist goals circa 2020.³²³ And suppose, given those reductions in incarceration, crime among those released was reduced by “only” 40, 30, or 20%.³²⁴ This would have a massive impact, on the order of the unprecedented nationwide reductions in violent crime that occurred in the 1990s.³²⁵

Thus, endorsing algorithms for use in Algorithmic Abolitionist policies does not necessarily risk their co-option for use in pro-carceral policies. Much more likely, a deal could be struck that would simultaneously reduce the harm from crime, prison, and policing. Moreover, in the large urban jurisdictions most likely to implement algorithmic policies, the deal would likely favor abolitionist preferences. As described above, Democratic voters appear open to abolitionist arguments for reducing prison and policing. They do not want to see crime go up; but there is little evidence of an appetite to drive it down at all costs.³²⁶

B. Humane Replacements for Police

The political argument for Algorithmic Abolitionism posits that there is pent up demand among left-leaning voters for policies that could significantly reduce prison and policing without increasing crime. That is, abolitionists succeeded in convincing the voters about the goal, but their concrete proposals were insufficiently attentive to the voters’ fears. This theory would be falsified if previous abolitionist proposals would in fact have avoided increases in crime but lost popularity anyway. Many abolitionists might argue exactly that. Often, their proposals called not just for eliminating policing and prison, but for replacing them with more humane alternatives.

If such alternatives existed, they might be normatively preferable even to Algorithmic Abolitionism. After all, Algorithmic Abolitionism still relies on the costly tools of policing and prison to control crime. It simply eliminates their excess use. But if other, more humane, tools could control crime just as well, then they might be used to reduce carceral harm even more.

For better or worse, however, the existing empirical evidence does not suggest that abolitionists’ proposed alternatives could replace very much prison and policing. Many of them have independent merits. And all of them would benefit from additional high-quality causal research. But, today, none seems likely to credibly rival the abolitionist potential of an algorithmic approach, at least not without increasing crime.

Here are a few of the most popular proposals, along with brief summaries of the evidence, as it stands.

Replacing police with mental health workers:³²⁷ Eugene, Oregon has had such a program—called CAHOOTS—since 1989.³²⁸ Despite its maturity, CAHOOTS workers respond to only 17% of 911 calls.³²⁹ Of those, 75% are welfare checks, requests to provide transportation, or non-emergency service requests from the public.³³⁰ Less than 2% of calls to which CAHOOTS is dispatched involve an identifiable report of a crime, and even those are simple trespass.³³¹ Thus, CAHOOTS almost never responds on its own to the kinds of adversarial calls most likely to escalate into violence. CAHOOTS may be effective within its domain, but that domain is quite small.³³²

Violence Interrupters: Violence interruption³³³ is an intervention where members work to “dissuade specific individuals and neighborhood residents in general from engaging in violence.”³³⁴ At the height of the “defund” movement, the approach was widely endorsed, even by non-abolitionists like New York Mayor Bill de Blasio and President Joe Biden.³³⁵ Dozens of cities have already deployed this approach.³³⁶ But the best empirical evidence so far does not show any clear crime controlling effect. A 2015 review of the empirical literature on violence interruption finds that the “evidence in support of the . . . model to date is mixed at best.”³³⁷ In the most favorable studies it had zero measured effect on crime at nearly half of deployment sites.³³⁸ The effects were even smaller on gun injuries and deaths.³³⁹ And since no studies so far are randomized or even quasi-randomized, they cannot “clearly disentangle the results from national and regional trends in violent crime.”³⁴⁰

Restorative Justice: “Restorative justice” practices are a proposed alternative to punishment via imprisonment. There, where wrongdoers are invited to “encounter” those whom they have wronged and try to “repair” those wrongs.³⁴¹ Again, the empirical results are mixed at best. One high-quality, randomized trial found no statistically significant effect of a restorative program on future commissions of domestic violence.³⁴² A recent review the literature likewise finds modest, at most, effects.³⁴³

Decriminalization: As discussed above, untargeted, across-the-board reductions in prison would cause crime to go up. But maybe some crimes should not be criminalized, or at least should not carry the penalty of incarceration, in the first place. ³⁴⁴ Could large amounts of prison be eliminated simply by updating our view about what is and isn’t worthy of punishment? Possibly, but there is little low-hanging fruit in the current system. Simple drug possession seems a likely place to start.³⁴⁵ But only 3.8% of Americans in state prisons were convicted of drug possession as their most serious offense.³⁴⁶ Likewise for sex work and other “quality of life” crimes. Only 6.2% of state prisoners’ most serious offenses were, “vice,” “decency,” or similar offenses.³⁴⁷ The vast majority of incarcerated persons in the U.S. are thus imprisoned for what are commonly considered serious crimes—violent crimes (58.2%),³⁴⁸ felony property offenses (15.3%),³⁴⁹ and major drug crimes like trafficking (10.2%).³⁵⁰ These seem like the kind of crimes about which even abolitionism-curious voters continue to worry a great deal.

Thus, abolition via decriminalization would require freeing people who have committed serious crimes and are likely to do so again. Ben Grunwald has recently published the best analysis of this approach.³⁵¹ Grunwald evaluates various alternative programs for eliminating 25, 50, or 75% of incarceration.³⁵² One thing remains constant across all strategies analyzed: each would increase crime, including serious crime, albeit to varying degrees.³⁵³ Algorithmic Abolitionism is favorable, then, because decriminalization does not seem like an alternative strategy that could satisfy voters’ twin aims of abolishing incarceration without increasing crime.

C. Constitutional Constraints

Algorithmic Abolitionism raises multiple constitutional questions. Many of these are well-canvassed in the existing literature on algorithms. What follows are a few brief thoughts on how Algorithmic Abolitionism, in particular, fits into these extensive debates.

The first possible constitutional speedbump is Due Process. In 2016, the Supreme Court of Wisconsin upheld the state’s use of the COMPAS tool in criminal sentencing.³⁵⁴ The defendant argued that, because that algorithm relied on statistical, rather than individual, data, its use violated his Due Process right to an individualized sentence.³⁵⁵ The Court disagreed, holding that algorithmic risk scoring was a way of arriving at accurate individualized determinations of relevant factors, including future crime risk.³⁵⁶ Nevertheless, the Court thought that, for sentences to be sufficiently constitutionally individualized, judges needed to be able to “disregard risk scores that are inconsistent with other factors.”³⁵⁷ It thus suggested that, if Wisconsin stripped its judges of discretion to ignore algorithmic recommendation, then that would violate the Due Process Clause.³⁵⁸

As described above, Algorithmic Abolitionist policies would need to give algorithmic recommendations bite. Sometimes, but not always, this might mean making them mandatory. The Wisconsin Supreme Court’s decision casts doubt on the constitutionality of taking humans out of the loop entirely. But it does not call less aggressive approaches into question. Moreover, this is just one state supreme court decision, and there are good reasons to think that it is wrong. Aziz Huq’s article-length argument that there is no constitutional right to a human decision is the best treatment of the topic.³⁵⁹ At most, Huq contends, there might be a constitutional right to a well-calibrated algorithmic decision.³⁶⁰

In lieu of a full recapitulation of the arguments, here are a few quick highlights. The Wisconsin Supreme Court argued that, to satisfy Due Process, humans need to be able to step in and override algorithmic decisions they deemed erroneous, given “other factors.” But the sophisticated machine-learning algorithms that would power Algorithmic Abolitionist policies already ignore “other” irrelevant factors, on a case-by-case basis. Unlike, say, an older-fashioned regression model, they do not assign the same weight to each fact under every circumstance. Rather, their complex decision procedures are highly context sensitive to interactions among all of the available information. Moreover, humans regularly allow such “other” irrelevant factors—race, charisma, attractiveness, and more—to drive their decisions. It therefore seems likely that keeping humans in the loop would sometimes worsen things, even by the Wisconsin Supreme Court’s own lights.

Another potential constitutional hurdle for Algorithmic Abolitionism comes from the line of U.S. Supreme Court cases that invalidated the mandatory U.S. Sentencing Guidelines. In United States v. Booker, the Supreme Court held that the Sixth Amendment forbade mandatory increases to sentences “beyond the prescribed statutory maximum” based on facts found by a judge, not a jury.³⁶¹ This dictate might again conflict with mandatory algorithmic sentencing—though not with other approaches to giving Algorithmic Abolitionism bite.

Moreover, even mandatory algorithmic sentences might be compatible with Booker, if carefully designed. First, Booker forbids only upward mandatory departures based on facts not found by a jury. An Algorithmic Abolitionist sentencing scheme could thus avoid the Booker problem by structuring itself around downward departures. The legislature might set a high default sentence for each crime and then mandate downward departures—up to and including no incarceration—for defendants with low recidivism risk. The Booker Court itself suggested that such a scheme would be constitutional. It declined to impose the scheme as a remedy in that case only because the Court thought it incompatible with legislative intent.³⁶² A legislatively enacted Algorithmic Abolitionist program would, obviously, not face that problem.

Moreover, Booker permits even upward sentence adjustments based on a defendant’s criminal history.³⁶³ The Kleinberg et al. algorithm, for example, uses criminal history to make its predictions.³⁶⁴ Other demographic information, like sex, number of dependents, or marital status, are matters of public record and will thus generally be easy to prove beyond a reasonable doubt, insofar as the defendant disputes them. Indeed, defendants could simply be asked to admit such facts, since admissions also raise no Sixth Amendment issue.³⁶⁵ Such demographic admissions could be made a precondition of eligibility for an algorithmic downward variance described above. Essentially all defendants would oblige, since there would be large potential benefits and no downsides.

The prior constitutional questions were mostly relevant to Algorithmic Abolitionist sentencing policies. What would the Constitution say about Algorithmic Abolitionist policing? The most obvious question here is whether the Fourth Amendment would prohibit algorithmically directed searches and seizures. Street stops—whether initiated on the basis of human intuition or algorithmic prediction—may be constitutionally carried out only on the basis of reasonable suspicion.³⁶⁶ Insofar as reasonable suspicion is a measure of probability, the result is straightforward. Empirical analyses suggest that the vast majority of such searches—initiated by humans—have a vanishingly small probability of turning up any contraband.³⁶⁷ Under Algorithmic Abolitionism, the probabilities triggering a search would be much, much higher.³⁶⁸ Thus, if any non-trivial share of human-initiated searches are rightly upheld as constitutional, Algorithmic Abolitionist searches are, a fortiori, legal too.

Both Kiel Brennan-Marquez and Emily Berman have argued that Fourth Amendment tests for reasonableness ought not rely on probability alone.³⁶⁹ They should instead require, they argue, a plausible explanatory narrative. Brennan-Marquez argues that positive law is ambiguous on this point—plausibly endorsing a probabilities-only view.³⁷⁰ But he contends that normative considerations like antidiscrimination, fair notice, and oversight cut against that reading and in favor of a more holistic one.³⁷¹

Much is correct in this account. But here are a few points in defense of Algorithmic Abolitionist policing—if not of all policing algorithms. First, Algorithmic Abolitionism brings even more normative considerations into the mix. Algorithmic Abolitionist policing would greatly reduce the burdens of policing, especially for those who currently suffer them the most. These serious benefits must be weighed against the putative normative costs Brennan-Marquez identifies.³⁷² Second, it is not even clear there is a normative tradeoff at issue here. Brennan-Marquez’s enumerated values might be well-served by certain well-designed Algorithmic Abolitionist policing policies. For example, in making street stop recommendations, the Goel et al. algorithm assigns significant weight to factors—the presence of a suspicious object, the sight of criminal activity—that look like explanatory reasons.³⁷³ This directly serves the value of fair notice. Antidiscrimination values are served when—as described above—algorithms reduce bias as compared with the human status quo. Finally, algorithmic designers can be subjected to democratic oversight just as police officers and their bosses can. The former is arguably easier. Algorithmic training data and design strategies can be audited for compliance with best practices, allowing detection of and consequences for gross deviations. By contrast, if a police officer lies to fabricate reasonable suspicion, that misbehavior is comparatively difficult to detect, prove, and punish.

Finally, a note on Equal Protection. As noted above, it is possible—though unlikely—that some Algorithmic Abolitionist policies could worsen bias in prison or policing, as measured using existing tools. Such statistical disparities would not be sufficient to raise an Equal Protection claim. For better or worse, proving an Equal Protection violation requires proving that a government policy was enacted with discriminatory intent.³⁷⁴ And if the policy was enacted with the intent of reducing incarceration, including for vulnerable racial groups, that would not qualify.

If the algorithm powering such a policy used race among its many input variables, strict scrutiny would apply.³⁷⁵ Then, the algorithm would have to serve a compelling interest and be narrowly tailored to that interest. Here is one way in which it might. First, excluding race as an input variable can often have the unexpected effect of increasing algorithmic bias against minority groups.³⁷⁶ Thus, including race might be the only way to serve the compelling interest of avoiding discrimination.³⁷⁷ And in situations where including race did not improve accuracy and thereby reduce unfair disparities, it could simply be dropped from the model.

D. Privacy and Surveillance

The literature on law, technology, privacy, and government surveillance is massive. As with the prior thorny normative issues, this Section does not seek to resolve higher-order debates. It does, however, situate Algorithmic Abolitionism within them. It suggests that, as compared with other potential uses of algorithms in public policy, the privacy and surveillance concerns here are less dire.

Facial recognition technology powered by machine learning is, for example, now being used by the Chinese government. The goal appears to be building a surveillance system to track and log all citizens’ public movements at all times.³⁷⁸ Alternatively, machine learning can be used to infer private facts, even when their subjects actively try to conceal them.³⁷⁹ Consider, e.g., ad-serve algorithms that infer a woman’s pregnancy even before she has told her own family.³⁸⁰

These two worrisome examples illustrate the two ways in which machine learning can help increase surveillance and reduce privacy. They can automate the collection of data that, while public in some sense, could not previously be aggregated at scale. They can also use data—including the data they help to aggregate—to discover further, private facts that could not previously be discovered.

How do algorithmic abolitionist policies fit into these twin concerns? Begin with data collection. Nearly all of the algorithms described above would operate only on data that the government already routinely collects. Indeed, that is why they exist in the first place. To train algorithms, you need existing data. The algorithms discussed above consisted of public administrative records. Sentencing algorithms rely on the ordinary contents of a criminal record: sentence, arrest history, offense, age, and the like.³⁸¹ And the algorithm for Terry stops uses exactly the same variables that police currently record as their reasons for making stops today.³⁸²

Thus, these algorithms do not surveil at all, in the sense of automating the collection of new raw data. But they do make inferences about facts, like whether a person is likely to commit a crime or a geographic area is likely to suffer from elevated crime. These facts might be considered private, in the sense that the individuals they are about would prefer them to remain unknown. But this is unlike an algorithm’s exposure of an early stage pregnancy.

The reason is that our law does not usually treat facts about future crimes as private. Consider, for example, the structure of the Fourth Amendment, a privacy-protecting constitutional rule.³⁸³ The Fourth Amendment protects the privacy of “persons, houses, papers, and effects,” except if there is “probable cause” to think that violating that privacy will reveal facts about a crime.³⁸⁴ The fact about the crime is not only non-private, but law affirmatively abrogates some legitimate privacy interests to reveal it.

The exception to all of this, at least potentially, is Algorithmic Abolitionist traffic enforcement. That policy relies on a new network of algorithmically enabled cameras. How does this compare, for example, to China’s facial recognition network? It is entirely up to us. Traffic enforcement cameras would not need facial recognition and certainly would not need to track everyone’s location at all times. They would not need to make any record of their feeds at all, saving only those brief snippets on which citations would be based. This would entail some surveillance. But arguably less than the traffic officer on the side of the road, who observes and can remember anything he sees during his shift, not just the violations.

Perhaps merely building the camera network would inevitably lead to its use in mass surveillance. After all, ubiquitous location tracking would surely, for example, help to increase America’s shockingly low murder clearance rate.³⁸⁵ Such uses might be tempting. But they are not inevitable—or even necessarily constitutional. In 2018, the Supreme Court held that the Fourth Amendment forbade law enforcement to use cell phone geolocation data as such an on-demand location tracking system.³⁸⁶ This holding does not guarantee a surveillance-minimalist implementation of Algorithmic Abolitionist traffic enforcement. But it cuts against the possibility of a surveillance-maximalist approach.

Finally, it is important to emphasize that abolishing prison and police also means abolishing status quo surveillance and privacy invasions. Prisoners are subject to round-the-clock surveillance, with no privacy, and much of what police do imposes the same costs. Significantly reducing both would be a huge victory for privacy and civil liberties. True, Algorithmic Abolitionism might intensify surveillance for some individuals—for example, habitual carriers of illegal weapons. But because of Algorithmic Abolition’s reductions in levels, the net result would be to lift the burden of surveillance from many more people. Likewise, insofar as law enforcement would also wish to use these technologies to reduce crime somewhat, there is ample space for a positive-sum deal.

Conclusion

This Article has proposed Algorithmic Abolitionism as a solution to an extremely difficult problem. Really, it is two problems—the evils of crime and the evils of policing and prison. The recent rise of prison and police abolitionism has made the latter problem more publicly salient than ever before. But it has not made the dilemma more tractable. Until now, most available approaches have been distinctly unappetizing: One could—it seemed—either eliminate police and prison and accept high costs from increased crime or instead preserve police and prison and accept the high costs that come with them. Algorithmic Abolitionism is a third, more appealing, approach. It is a way to radically reduce the high costs of prison and policing without generating additional costs from serious crime. Algorithmic Abolitionism is possible today. And its normative risks are small, as compared with the upside. Thus, this Article contends, we should all be Algorithmic Abolitionists.

* Assistant Professor of Law, The University of Houston Law Center; Law and Policy Advisor, Center for AI Safety. Thanks to Nicholas Almendares, Omri Ben-Shahar, Joseph Blocher, Jacob Bronsther, Kiel Brennan-Marquez, Seth Chandler, Ryan Copus, Nikolas Guggenberger, William Hubbard, Orin Kerr, Guha Krishnamurthi, David Kwok, Jonathan Masur, Alex Platt, Nicholson Price, Daniel Rauch, James Tierney, Justin Vanderschuren, Carleen Zubrzycki, and the participants in the University of Chicago International Junior Scholars Conference; the University of Michigan Junior Scholars Conference, The Emory AI Scholars Roundtable, and the University of Houston Law Center Works in Progress Workshop. Special thanks to Madeline McCune and Nathan Halaney for exceptional research assistance.