Branded for Life

A criminal sentence is, by design, a number. Eighteen months. Five years. Life, on a bad day, which is at least a number of a kind. The number is the state’s promise about the debt: this much, and then it is paid.

The mark has no number. A conviction, an arrest, a line in a database, a score on a checklist: each of these is created at a moment and then simply persists, in a system that was built to remember and was never built to forget. Nobody sentenced you to it. It was not in the plea. It does not appear on the paperwork you signed, because it is not a punishment in the sense that the law uses the word, and that distinction, as we are going to see, is doing a surprising amount of load-bearing work.

A conviction, an arrest, a line in a database, a score on a checklist: each of these is created at a moment and then simply persists, in a system that was built to remember and was never built to forget.

By the mid-2000s there were nearly 81 million criminal records on file in the American states, 74 million of them in automated databases, with about 14 million more arrests recorded every year.¹ The National Institute of Justice, describing the national inventory of what happens to you after a conviction, says those consequences “rarely play a prominent role in sentencing or plea negotiations and are often not discussed with the accused.”² In 2018 the NIJ described that inventory as a database of more than 44,000 separate collateral consequences, and the National Reentry Resource Center now describes it as listing more than 40,000.³ The two figures differ because the inventory is revised, and I do not know which is the better description of the law today. Either way the order of magnitude is the point. Nobody has read all of it. Nobody could.

I want to ask whether a free country should be in the business of marks without numbers. I want to do it the way I do these things, which is by putting the best case for the mark on the table first, and letting it have its say, before I say what I think the evidence does to it.

The house rule

The same one as always, shorter this time.

An empirical claim says what happened or what is true: a rate, a count, a finding. I source those, date them, and say when they are contested. A moral premise says what matters: that a person is more than his worst act, that a community has a right to protect itself, that a punishment ought to end. I state those as premises so that you can reject them without having to find an error in my footnotes. Rhetorical force is the adjective and the dry joke. I will use some. When a paragraph is mostly force, I will say so.

And the boundary. This essay criticizes laws, agencies, databases, policies and courts. It says nothing about any people, neighborhood, nationality, or faith as such. Several of the things I describe fall unevenly on Black and Latino Americans, and I report that because the audits report it, and because the racial pattern is a property of the systems and not of the people in them. Several of the people described here have done terrible things. Several have been wrongly described as having done them. The argument is about what a state should do in the space between those two facts.

A man at forty

Two researchers at Carnegie Mellon, funded by the National Institute of Justice, once opened an article with a hypothetical that I cannot improve on. Take a man of forty who was convicted of burglary at eighteen and has committed no further crime in twenty-two years. Every time he applies for a job he must either say he has a felony conviction or wait for the employer to find it.⁴ The authors say that in either case “he probably will not get the job,” because many employers are unwilling to hire anyone with a conviction.⁵

I mean to keep that man in the room. He is not a typical former prisoner. He is the limiting case of the argument: the person for whom the mark is no longer evidence of anything except the date on which it was made. The strongest defenders of the mark are not defending the treatment of that man, and I will try not to pretend that they are. They are defending the system that cannot tell the difference between him and the man who was released last Tuesday. That is a better argument than it sounds. We will get to it.

What the mark does

Empirical. Start with the audit studies, because they are the cleanest evidence we have about the labor market’s response to a record. In the early 2000s the sociologist Devah Pager sent matched pairs of male testers, differing in race and in whether they presented a prior conviction, to apply for real entry-level jobs in Milwaukee. None of the testers actually had a record. White applicants without a record were called back 34 percent of the time. White applicants who presented a record were called back 17 percent of the time. Black applicants without a record: 14 percent. Black applicants with a record: 5 percent.⁶

Notice what that table says, because it carries two findings, and they are different in kind. The first is that a record roughly halves your chances: 34 to 17 among white applicants. The second is that race does at least as much work as the record. A Black applicant with a clean history was called back less often than a white applicant who disclosed a conviction. Both are from a single city and an experiment two decades old, and a careful reader will hold them loosely, and I am holding them loosely. But the direction has been replicated often enough, as we will see, that I am prepared to call it a finding and not a curiosity.

The aggregate picture is consistent with it. The Prison Policy Initiative, using the National Former Prisoner Survey, estimated that formerly incarcerated people in 2008 had an unemployment rate of 27.3 percent, compared with 5.8 percent for the general public, and noted that this exceeds the highest unemployment rate ever recorded in the United States, 24.9 percent in the Great Depression.⁷ I caveat this the way a pessimist should: it is a 2008 figure, from a survey with its own limits, published by an advocacy organization, and it measures unemployment in a population selected by having been to prison. But it points where the audit studies point.

I will add one thing that is not a number. The mark is not mainly delivered by the state. A person with a record does not meet the state at the job interview. He meets a private employer, a landlord, a licensing board, a rental-screening company, and each of them is making a rational local decision with a record that the state created, stored and sold or released. The state wrote the sentence and then outsourced the punishment to anyone with a keyboard. It is the cheapest form of punishment there is, because someone else pays for it.

The state wrote the sentence and then outsourced the punishment to anyone with a keyboard.

The strongest case for the mark

Here is the case in its best form, and it is stronger than the people who oppose the mark usually allow.

First, the base rate. People who have been to prison are re-arrested at high rates. The Bureau of Justice Statistics followed about 92,100 prisoners, representing roughly 408,300 released from state prisons in 34 states in 2012. Sixty-two percent were arrested within three years and 71 percent within five.⁸ That is not a rounding error. An employer who looks at a record is not indulging a prejudice. He is reading a signal that, on average, carries information.

Second, liability and duty of care. The NIJ authors themselves list “growing concerns about employer liability” as one of the forces driving the spread of background checks, and observe that employers serving vulnerable populations, such as children and the elderly, are rightly sensitive to a record involving violence, while a bank is rightly sensitive to a record of theft.⁹ An employer who hires someone he could have checked, and who then harms a customer, has a problem that no one in a reform conference is going to share with him.

Third, information beats guessing. This is the best of the arguments and I want to give it its full weight. In a 2006 study in the Journal of Law and Economics, Harry Holzer, Steven Raphael and Michael Stoll found that employers who run background checks were more likely to hire Black workers, and especially Black men, than employers who did not, and that the effect was larger among employers who said they were averse to hiring people with records. The authors read this as evidence that employers who cannot check will statistically discriminate, assuming a Black applicant is more likely to have a record.¹⁰ That cuts against the reformers: take away the check, the argument goes, and you take away the thing that was preventing the guess.

Fourth, the record is not a rumor. A conviction is a fact about a real act, usually with a real victim. A society that decides to forget it is making a moral claim about the victim’s claim on the public memory, and some victims do not want it forgotten, and are not obviously wrong.

Fifth, the law already pushes back. I should be fair about the other side of the employer’s ledger. The Equal Employment Opportunity Commission’s 2012 guidance says that a blanket policy of excluding people with records can create liability under Title VII where it has a disparate impact, and describes a safer course: a “targeted screen” that looks at the nature and gravity of the offense, the time that has passed, and the nature of the job, followed where possible by an individualized assessment.¹¹ Employers are not squeezed from one side only. They face liability for hiring and liability for refusing, and a cautious general counsel, faced with both, tends to choose the policy whose costs fall on someone else.

I think these are serious arguments and I will not pretend they are straw. Now let me say what the evidence does to them, one at a time.

What the 71 percent does and does not say

The base rate is real and it is the strongest card. But it is a card that gets played with the wrong face up.

Empirical. The 71 percent is a rearrest rate within five years, not a reconviction rate and not a re-offense rate. BJS counted arrests, which do not require a finding of guilt. Within the same study, 46 percent of the prisoners in the 21 states with the relevant data had a parole or probation violation or an arrest that led to imprisonment within five years. The annual arrest percentage fell from 37 percent in the first year after release to 26 percent in the fifth. Those released at 24 or younger were arrested at 81 percent within five years, compared with 74 percent for those aged 25 to 39 and 61 percent for those 40 or older. And across the 19 states in BJS’s 2005, 2008 and 2012 studies, the five-year arrest share declined, from 77 percent to 75 percent to 71 percent.¹²

Read that again as a shape and not as a headline. The risk is highest in the first year, falls each year after, is concentrated among the young, and has been drifting down across cohorts. A mark that does not fade treats a fifty-year-old in year ten as though he were a twenty-three-year-old in month three. The base-rate argument justifies caution for a period. It does not obviously justify forever.

How long is the period? Alfred Blumstein and Kiminori Nakamura tried to answer that. Using the criminal-history records of about 88,000 people first arrested in New York State in 1980, they asked when the risk of a new arrest for people who had stayed clean fell to the level of same-age people in the general population. For people first arrested for robbery at 18 the answer was 7.7 years. For burglary at 18, 3.8 years. For aggravated assault at 18, 4.3 years. The age of the first arrest mattered: for robbery, 8.5 years for those first arrested at 16 and 4.4 years for those first arrested at 20.¹³

The caveats are the authors’ own, and they are not small. The results were described as preliminary. They were from a single state and a single cohort that began in a peak crime year. They measured arrests, not convictions. And “redeemed” in their sense does not mean the person’s risk becomes the same as someone never arrested. By their own account the curves for someone who had been arrested “would never be the same” as for people who had never been.¹⁴ So what the study establishes is narrower and, I think, more useful: for the offenses they tested, the point at which a clean record in the past tells you no more about an adult’s chances of a new arrest than his age does is measured in years, not decades.

Next to this sits the expungement evidence. J. J. Prescott and Sonja Starr studied expungement in Michigan. They found that among those legally eligible, only 6.5 percent obtained expungement within five years, a serious “uptake gap,” that recipients had “extremely low subsequent crime rates, comparing favorably to the general population,” and that within a year of expungement wages rose by over 22 percent against the prior trajectory.¹⁵ I should flag the obvious worry about studies of people who opt in: those who get an expungement are those who qualified and could navigate a process, and a reader should ask who is in the comparison group. The authors compared recipients with comparable nonrecipients, and I have read the abstract and not the full identification strategy, so I mark that as something a skeptic should check.

I think the honest summary is this. The base-rate argument is correct about the first few years and wrong about the whole of a life. The state’s record system treats those as the same thing. A defender of the mark could accept every number above and reply, correctly, that nothing in them requires the mark to be abolished, only dated. I agree. Dated, in my view, is the position.

Ban the box: the policy everyone likes, and the evidence that complicates it

If background checks are the problem, then removing the felony box from the job application is the obvious reform, and it has been enacted in a great many places on the strength of a moral intuition that I share. It is also the part of this subject where the honest evidence is least comfortable for the people who agree with me.

Empirical. The key studies point in different directions, and I will set them out as they are.

In a field experiment published in the Quarterly Journal of Economics in 2018, Amanda Agan and Sonja Starr submitted about 15,000 online applications from fictitious young male applicants, with distinctly Black or distinctly white names, to employers in New Jersey and New York City, before and after ban-the-box rules took effect. Before the rules, white applicants received 7 percent more callbacks than Black applicants. After, they received 43 percent more.¹⁶ Their own interpretation, as reported, was that “employers are relying on exaggerated impressions of real-world racial differences in felony conviction rates.”¹⁷

In a 2020 paper in the Journal of Labor Economics, Jennifer Doleac and Benjamin Hansen, using Current Population Survey data from 2004 to 2014, found that ban-the-box policies reduced employment among young, low-skilled Black men by 3.4 percentage points, a decline of 5.1 percent relative to baseline, and by 2.9 percent among comparable Hispanic men.¹⁸ Their term for the mechanism is statistical discrimination: when the signal is hidden, employers fall back on the group average.

The Holzer, Raphael and Stoll result I described above points the same way from an earlier decade.¹⁹

In the other direction, there is Terry-Ann Craigie’s study in Economic Inquiry, which found that public-sector ban-the-box policies raised the probability of public-sector employment for people with convictions by about 30 percent on average, with no evidence of increased statistical discrimination against young, low-skilled minority men.²⁰ I am relying on the paper’s abstract here, not its full tables. The public sector is also a special employer: it tends to have formal procedures, accountable hiring and anti-discrimination law that actually bites, which may be exactly why the policy behaves differently there.

There is a fourth set of results that cuts the other way for the policy’s supporters. A National Bureau of Economic Research working paper by Joseph Sabia and colleagues associated ban-the-box laws with a 10 percent increase in criminal incidents involving Hispanic male arrestees, and found no evidence that the laws increased property crime among Black men.²¹ It began as a working paper, the authors attribute the pattern to employer statistical discrimination, and a result of that kind deserves replication before anyone leans on it. I include it because an essay that cites only the findings it likes is not an argument, it is a mailing list.

And in a sign that this is not a story with a single villain, a study of Hawaii’s 1998 law found that felony defendants prosecuted in Honolulu were 57 percent less likely to have a prior conviction after the law took effect.²² That is a suggestive result, not a causal one, and it points toward the policy’s supporters.

So what does the pessimist conclude? Not that ban-the-box is a failure. The evidence is mixed and the heterogeneity is the story. What I conclude is that the policy treats the symptom, which is the question on the form, and leaves the underlying thing untouched, which is the belief of an employer about what kind of person has a record. Hide the information and the belief does not vanish. It goes looking for another source, and in a country with racialized assumptions about crime, the other source can be a name. Pager’s table, with its fourteen percent next to its seventeen, was already telling us that.

The lesson I take is not “go back to the box.” It is that the box was never the disease. If you want an employer to stop treating the record as a verdict, you have to change what the record says and how long it says it. Which is a different reform, with a different evidence base, and I want to turn to it now.

What actually moves the number

Empirical. Three kinds of intervention have a better evidence base than the box.

The first is making the record go away on a schedule. The Clean Slate Initiative, an advocacy organization, counts 14 states and the District of Columbia that have passed laws meeting its criteria for automatic record sealing, beginning with Pennsylvania in 2018 and most recently Missouri in 2026.²³ I note that it is a count of laws that meet one organization’s standards, and that the details of what is sealed, and when, vary a good deal from state to state. The reason “automatic” matters is the 6.5 percent in Michigan. A right to clear your record that you have to know about, afford and file for will be used by those who need it least.²⁴

The second is giving people work and watching what happens. Sara Heller’s randomized evaluation of Chicago’s One Summer Plus program, a youth jobs and mentoring program, found a 43 percent reduction in violent-crime arrests over 16 months for the 2012 cohort and, across both cohorts studied, a reduction of about 35 percent in violent-crime arrests, 6.4 per 100 youths against a baseline of 18.3. The 2015 cohort’s apparent 52 percent decline in arrests was not statistically significant.²⁵ This is not a reentry program, and it is a program for youth, so I treat it as evidence of one thing only: that employment, in the right form, appears to reduce violent offending in a population at risk of it.

The third is just not recording what you cannot defend, which brings me to the databases.

The ballot

The oldest of the marks is the one that takes away the vote.

Empirical. The Sentencing Project’s Locked Out 2024 estimates that about 4 million Americans, 1.7 percent of the voting-eligible population, could not vote in the 2024 election because of a felony conviction, in the 48 states that restrict voting on that basis. Maine and Vermont are the exceptions, and also allow people to vote from prison. The report estimates that about 40 percent of the disenfranchised have completed their sentences entirely, that ten states disenfranchise some or all people after sentence, and that the rate for Black adults, 4.5 percent, is more than three times the 1.3 percent rate for non-Black adults. Florida alone accounts for over 961,000 of the total, with an estimated 730,000 of them having finished their sentences. After the state’s voters passed Amendment 4 in 2018 to restore the vote to most people who completed their sentences, the legislature passed a law in 2019 conditioning restoration on payment of outstanding fines, fees and restitution. The report says the law effectively “re-disenfranchised” most of those the amendment was meant to restore.²⁶

Two cautions. The Sentencing Project is an advocacy organization, and its numbers are modeled estimates built from life tables and state correctional data, not head counts. The authors say so, and that “these and other estimates must be interpreted with caution.”²⁷ And the characterization of the Florida law is theirs. I can confirm the numbers as the report gives them. I cannot independently confirm every count.

Now the other side. The Supreme Court held in Richardson v. Ramirez in 1974 that a state does not violate the Equal Protection Clause by excluding felons who have completed their sentences and paroles from the franchise. The Court’s reasoning rested on Section 2 of the Fourteenth Amendment, which exempts disenfranchisement “for participation in rebellion, or other crime” from the penalty of reduced representation.²⁸ Justice Rehnquist, writing for the majority, said that the case for restoring the vote to ex-offenders, as a matter of rehabilitation, might be a good one if addressed to the legislature, and added that if the people of a state did not come around to it, “their failure is some evidence, at least, of the fact that there are two sides to the argument.”²⁹

There are two sides to the argument, and the other one is not stupid. A political community may reasonably set conditions on those who have made a serious breach of its terms. A person who has defrauded voters, or robbed their neighbors, or killed someone, has made a claim on the rest of us, and the rest of us may have a say about whether he gets to help choose the next set of rules. The question is not whether some version of that is defensible. It is.

My objection is narrower. It is about what the mark is for once the sentence is over. If a loss of the vote is a punishment, then it belongs in the sentence, with a number attached, and the court should have to say how much of it it is imposing. If it is not a punishment, then there has to be some forward-looking public-safety reason a person cannot vote, and I have not found one that survives the question “why does it protect anyone to keep a man with a completed sentence from a ballot?” The most defensible form of the argument, I think, is the one that applies only while a person is in prison, which is the rule in 23 states.³⁰ What is hardest to defend is the system that keeps the mark on for life, or makes the restoration conditional on whether a person can pay.

The registry

I am going to be most careful here, because this is the subject on which the strongest defense of the mark is the best, and the most sympathetic objection to it is the hardest to say out loud.

The best case for the registry. In 2019 the Bureau of Justice Statistics published a nine-year follow-up of prisoners released in 2005. Among those released after a conviction for rape or sexual assault, 7.7 percent were arrested for rape or sexual assault within nine years, more than three times the 2.3 percent rate among other released prisoners. Released sex offenders accounted for 5 percent of releases and 16 percent of arrests for rape or sexual assault in the follow-up period.³¹ Every one of those arrests represents a person who was harmed, or alleged to have been, and it is a finding that cannot be waved away by anybody who is serious. An argument that the public has a legitimate interest in knowing is an argument that must be answered, and not by pretending that the risk is zero.

That is the empirical high ground for the registry, and I concede all of it. Now here is what else is in the same report. The same study found that about two-thirds, 67 percent, of released sex offenders were arrested for any crime in the nine years, compared with about 84 percent of other released prisoners. Half had a subsequent arrest that led to a conviction. And BJS measured arrests, which means it both overstates some things, since an arrest is not a conviction, and understates others, since many sexual offenses are not reported.³² I note the second point because it cuts against my side and I want it on the table.

So the registry’s defenders have a real finding. The question is whether the design of the registry follows from it, and for that we have a court record. In Smith v. Doe in 2003 the Supreme Court held that Alaska’s registration and public-notification law was a civil, regulatory scheme and not punishment, so applying it retroactively did not violate the Ex Post Facto Clause.³³ The Court’s reasoning, as the Sixth Circuit later summarized it, leaned on the premise that the risk of recidivism among sex offenders is “frightening and high.”³⁴

Thirteen years later, in Does v. Snyder, the Sixth Circuit looked at Michigan’s far more elaborate law, with its residence, work and “loitering” limits around schools, its tiers based on the offense of conviction and not on individual assessment, and its requirement of in-person reporting, and held that its retroactive application was punishment. The court said the record gave “scant support” for the claim that the law accomplished its goal, noted evidence that offense-based registration had “at best, no impact on recidivism,” and observed that one analysis in the record concluded laws of this kind may actually increase the risk, “probably because they exacerbate risk factors for recidivism by making it hard for registrants to get and keep a job, find housing, and reintegrate into their communities.”³⁵ That last claim traces to the Prescott and Rockoff study, whose findings are more mixed than the quote suggests: as I read the abstract, registration itself, by giving police information, appeared to reduce some reported offending, while public notification could, at the margin, increase recidivism among registrants.³⁶ (I have read the abstract and the court’s use of the paper, not the whole study.)

I want to quote one more sentence from the Sixth Circuit, because it does what an essay like this one is supposed to do, which is to hold both truths at once: “while many (certainly not all) sex offenses involve abominable, almost unspeakable, conduct that deserves severe legal penalties, punishment may never be retroactively imposed or increased.”³⁷ The court also described Michigan’s registry as one that “brands registrants as moral lepers solely on the basis of a prior conviction” and that “consigns them to years, if not a lifetime, of existence on the margins.”³⁸ The same opinion noted that one plaintiff was on the registry for a nonsexual kidnapping offense that arose from a 1990 robbery.³⁹

I draw two limits around what I can claim here. Does v. Snyder is one federal circuit’s holding about one state’s law, and it applies the Smith framework, not overrules it. And I am not arguing that registries are useless or that the people on them are owed nothing. My claim is the narrow one the court made: that a mark which is public, address-based, tiered by offense and not by present risk, and without any exit, is a different object from a record that law enforcement keeps, and the case for it needs evidence that the extra does something, and not just that the underlying risk is real.

The gang file

Empirical. A gang database is a file in which police record that a person is, or is believed to be, a member or associate of a criminal gang. Unlike a conviction there is no court, no verdict and no sentence. What puts a person in is a set of criteria applied by an officer. What takes a person out, in many systems, is time, or nothing at all.

Start with the audit. In 2016 the California State Auditor examined CalGang, the state’s shared gang database. At the four agencies it reviewed, it found that all four lacked adequate support for including 13 of the 100 individuals it examined, and that of more than 560 criteria it reviewed, the agencies lacked adequate support for 131, or 23 percent. It found 42 people in the database who were supposedly younger than one year of age at the time of entry, 28 of them listed for “admitting to being gang members.” It found more than 600 people with purge dates beyond the five-year limit, and more than 250 with purge dates more than 100 years in the future. The database’s own audits reviewed less than 1 percent of records. Two of the agencies had failed to properly notify more than 70 percent of the 129 juveniles whose records the auditor reviewed, and three agencies admitted using CalGang for employment or military screening, which the database’s own policy prohibited.⁴⁰

I want to put the state’s response on the table next, because it is a rare instance of the system doing what a critic asks. California passed the Fair and Accurate Gang Database Act of 2017 (AB 90), and the Department of Justice adopted regulations in October 2020. Under them, adults are removed from CalGang after five years and minors after three, unless there is new qualifying contact, and the records may not be disclosed to assess eligibility for education, employment or housing, to enforce federal immigration law, or for military screening, unless a statute or regulation requires it. The database is not accessible through the system police routinely check when they stop someone.⁴¹ The Attorney General’s 2025 report counts 9,948 records as of September 30, 2025, and says that the figure may overcount individuals, since the same person can be entered by more than one agency. By race, it lists 7,091 Hispanic, 2,077 Black, 599 white, 98 unreported and the remainder in small categories.⁴² Of the total, 273 entries are for people aged 13 to 17.⁴³

So in one state, the 2016 findings produced a statute, a regulation and an annual public report. I count that as a system doing what a critic asked of it, and I do not want to be the kind of critic who cannot say so.

The other cities tell a less comfortable story. In April 2019 Chicago’s Office of Inspector General reported that over the previous twenty years, 91.3 percent of the 134,242 people designated as gang members on the police department’s gang arrest cards were Black or Latino men, and that 13 of the city’s 77 community areas accounted for more than half the arrest cards. The department’s data, the Inspector General said, was shared with more than 500 external agencies, including immigration agencies, with no intergovernmental agreements setting out controls, and found that the department’s gang information practices lacked procedural fairness protections, recommending among other things that people be notified and given a way to contest a designation. The department concurred with the findings but offered no timeline for the new system it promised.⁴⁴

In April 2023 the New York City Department of Investigation’s inspector general for the NYPD reported on the department’s Criminal Group Database. Of its active entries, 11,221 were Black people and 4,729 Hispanic, together about 99 percent. The report found that the department did not give sufficient guidance on entry criteria, was in some cases not following its own procedures, did not notify minors or their parents, had no appeals process for minors, and, in a sample of 494 entries from 2018 to 2022, had about 10 percent where one person served in multiple roles in the approval chain.⁴⁵

And then, because it is in the same report and it is what a fair summary must include: the investigators “did not identify a relationship between inclusion in the CGD and any individual adverse outcomes.” They cautioned that this may be because gang affiliation information exists in other forms and systems across the department, so that harm “is not attributable directly, or solely, to CGD,” and that harm may be hard to determine since people were not notified of inclusion. What they did find, mostly from advocacy groups, was “uncertainty and fear about CGD inclusion and its consequences in certain communities.”⁴⁶ I take that as a limit on what the audit shows, and I note that the people who say a gang database can hurt someone need to show the hurt, not just the flaw.

The best case for the gang file

Because the case for gang intelligence is not weak, I should state it properly.

Gangs are real, and a small number of the young people in them account for a lot of the violence. In the Boston Gun Project’s 1990s study, researchers working with police and street workers identified 61 gang crews with some 1,300 members. Gang members were less than 1 percent of the city’s young people, and at least 60 percent of the 155 killings of people aged 21 and under, by gun or knife, in Boston from 1990 to 1994 were gang related.⁴⁷ The data are thirty years old and from one city, which is a limit on how far they travel. The criminologist Terence Thornberry and colleagues, studying youth in Rochester, found that joining a gang appears to raise delinquent behavior, including violence, drug use and drug selling, beyond what the individuals showed before joining, and that offending falls after they leave.⁴⁸ (I have that conclusion from the book’s published summaries, not from a full reading.) A police department that did not track this would be negligent. A prosecutor who could not show a jury a pattern would lose cases that deserve to be won. And a neighborhood that lives with a feud has a legitimate demand that the state do something about it.

I accept all of that. And I notice what it implies, which is a database that is accurate, narrow, reviewed, time-limited and contestable. That is exactly the standard the audits say the databases in California, Chicago and New York were found not to meet. The Boston result is an argument for finding the 1,300, not for adding a hundred thousand names, and the more a database grows past the people who are in fact driving the violence, the less value it has to the police and the more harm it does to everyone else.

The risk factors point the same direction. The National Gang Center’s review of the research says that no single risk factor explains gang joining, that risk factors accumulate across domains, and that among those that matter are family poverty, poor parental supervision, school failure, disadvantaged neighborhoods and a living gang member in the home. In the Seattle study it cites, children with seven or more risk indicators were 13 times more likely to join a gang than children with none or one, and yet only 32 percent of those children joined.⁴⁹ That last number is the one a file like this cannot hold. A database that records “gang member” as a state of being, and keeps it for years, is describing a process, a pile-up of disadvantage at a particular age, as though it were an identity.

From poverty’s product to alien enemy

Nothing I have said so far turns on immigration. The last piece does.

Empirical, and recent. In March 2025 the administration invoked the Alien Enemies Act of 1798, a wartime statute last used in the Second World War, to remove Venezuelan nationals it identified as members of the Tren de Aragua gang, without ordinary removal proceedings.⁵⁰ A document filed in federal court a few weeks later, “Alien Enemy Validation Guide,” shows how membership was assessed. The guide applies to non-citizens aged 14 or older who are Venezuelan citizens. It is a scorecard of 20 items in six categories. A conviction under a law criminalizing activity related to the gang scores 10, a self-admission 10, a tattoo “denoting membership/loyalty” 4, a social media post displaying the group’s symbols 2, graffiti or hand signs 2, insignia, logos, drawings or dress “known to indicate allegiance” 4, and associating with or living with known members 2 each. Eight points or more validates a person as a member. Six or seven requires review by a supervisor and the agency’s legal office. Five or fewer does not validate.⁵¹

I want to be exact, because there is a lazy version of this story and it is not true. The guide contains safeguards. If all of the points come from the “symbolism” and “association” categories, the officer must consult a supervisor and agency counsel before validating. And at least one scoring category must involve conduct or information from the past five years.⁵² So a person with a tattoo and a piece of clothing, four points each, reaches eight points, but on the guide’s own terms that tally must go to a supervisor and agency counsel before anyone validates it. NBC reported the government’s position: a spokesperson for DHS said its “intelligence assessments” determine membership, a Trump administration official told NBC the assessments go “well beyond single tattoos” (NBC’s paraphrase), and DHS described the premise that the people are not gang members as based on “faulty assumptions.”⁵³

NBC also reported what the families and lawyers said. They described men with tattoos such as “Mom” and “Dad” with crowns, a rosary, and a Real Madrid crest, and gang experts who said that the Venezuelan gang, unlike some others, is not known for using tattoos to identify its members.⁵⁴ These are claims by lawyers, families and experts on the record in litigation. I cannot verify any individual’s case. I can say that the instrument the government used gives a tattoo four of the eight points needed, and that this, on its face, is a lot of weight to put on a thing that people get for all kinds of reasons.

The legal status is this. In May 2025 the Supreme Court held that the administration had not given the Venezuelans enough time or information to contest their removal under the Act and extended its injunction. A three-judge Fifth Circuit panel in September 2025 agreed with those who argued the Act was not meant to be used against gangs like Tren de Aragua. The full court vacated that and heard the case en banc. On August 13, 2026 it dismissed the case as moot, because the three named petitioners had by then been removed under ordinary immigration law, and it did not decide whether the invocation was lawful.⁵⁵ The ACLU’s lead lawyer said the injunction remains in place and that courts have “soundly rejected the use of this wartime authority during peacetime,” which is his characterization.⁵⁶ So this case ended without a ruling on the merits. I have not surveyed every other challenge to the Act, and I do not claim to know where all of them stand.

The argument. I am not claiming that the Tren de Aragua does not exist or that some of its members have not done terrible things. A government has a duty to protect people from violent organizations, and the guide has real safeguards that a more casual process would not. My claim is about the category.

A gang member in an American city is, in the research literature, a person at the end of a long list of risk factors that the state could have affected, and whose offending, on the Rochester findings, appears to fall once he leaves. That is a problem category: you work on it, you reduce it, you intervene at 13 and not at 23. An “alien enemy” is a different category. It is a category of war. It does not ask what put a man there. It asks only whether he is one of them, and it answers by scorecard, and the remedy is not intervention but removal. The shift from “product of a poverty and a neighborhood” to “enemy of the nation” is a shift from a social problem to a security threat, and the most reliable property of a security threat is that it is not owed an explanation.

You can see the same shift in Chicago, where the Inspector General recommended amending the city’s Welcoming City ordinance, whose protections excluded “gang-associated individuals,” because a database the Inspector General found to be flawed was being shared with immigration agencies.⁵⁷ In California the regulations expressly forbid using the database for immigration enforcement,⁵⁸ which suggests the risk was understood.

The pessimist’s one-line version of this: the database is where we put a person’s past. The scorecard is where we decide he has no future here. Both are made of the same material, which is suspicion with a number on it.

What a limit would look like

I said at the beginning that I wanted an argument for a limit, not for abolition. Here is what the evidence in this essay supports, and what it does not.

Date the mark. The Blumstein and Nakamura finding, if it generalizes, says that for common offenses the risk of a new arrest for someone who has stayed clean converges with that of same-age peers within roughly four to eight years. That is an argument for a presumptive sunset on the visibility of old records, with the length depending on the offense, and with exceptions for roles where a specific, severe risk persists, which is the position the EEOC guidance already takes in asking about the nature of the offense, the time that has passed and the nature of the job.⁵⁹

Make it automatic. Michigan’s 6.5 percent uptake figure is the strongest argument for automatic clearing. A right that you must know about, afford and file for is a right for the people who need it least.⁶⁰ Fourteen states and the District of Columbia have gone that way, according to the organization that tracks it.⁶¹

Make every mark contestable. The gang-database audits converge on weak or absent notice, weak or absent appeal and weak purging. The cheapest fix in this whole essay is to tell a person he has been entered and to give him a way to say that the entry is wrong. California’s regime under AB 90, with its automatic purges, its limits on disclosure and its annual public report, is the nearest thing to a live test, and the database is still in operation.⁶²

Change the record, not just the box. The evidence on ban-the-box is that hiding the information does not remove the belief behind it. The reforms that move the number are the ones that change what the record says and for how long: sealing, expungement, and employment itself.⁶³

Be honest about the cost. A dated mark will, occasionally, hide a record from an employer who would have wanted to see it. Some of the people whose records disappear will offend again. I think that cost is worth paying, and I want the people who disagree to be able to say that I admitted it.

What the evidence does not support is the claim that records should simply disappear, or that employers and communities have no legitimate interest in knowing. Nor does it support the claim that the systems I have criticized were designed by anyone to be cruel. They were built, mostly, by people trying to prevent something bad, with tools that don’t forget.

Where I could be wrong

A pessimist should audit his own pessimism.

I could be wrong that the risk fades as fast as the Blumstein and Nakamura curves suggest. Those are arrest data from one state and one cohort, and the authors themselves said so. If risk declines more slowly for some offenses, my “four to eight years” is too short, and the right design is a longer, offense-specific schedule.

I could be wrong about ban-the-box. The evidence is mixed, and the largest effects come from studies with different designs. The public-sector result and the Hawaii result point the other way, and one honest reading is that the policy works where hiring is formal and fails where it is not.

I could be wrong about the gang databases, where the best evidence I have comes from audits of three jurisdictions. The New York finding that no adverse outcome could be tied to inclusion is a reminder that I am describing a risk and a procedural failure, not a measured harm.

And I could be wrong about the whole frame. A society that is bad at rehabilitation may do better to be honest about it and keep the records, than to pretend and hide them. That is a respectable position. I do not hold it, because I think the evidence that the mark does not fade is thinner than the evidence that the person does. But I hold that with less certainty than I would like.

The pessimist’s coda

The mark persists because it is cheap. A sentence costs a prison bed. A mark costs nothing: it is stored, searched and sold at the margin by people who bear none of its consequences, and the consequences land on someone who is not in the room when the decision is made.

I do not think the people who built these systems were cruel. I think they were afraid, and they had a machine that could remember, and nobody asked the machine what it was for. The thing that distinguishes a punishment from a mark is that a punishment can end. If we want to keep the second, we should at least have the honesty to put a date on it.

The thing that distinguishes a punishment from a mark is that a punishment can end.