Sixteen Prison Beds to Prevent One Violent Crime. The Evidence Has a Better Ratio. AI Makes It Scale.

Sixteen. That is the number of people you must imprison to prevent one future violent crime through incapacitation alone, according to a quasi-experimental study that exploited the random assignment of defendants to harsher or more lenient judges to isolate prison's causal effect on violence. The study tracked defendants eligible for both prison and probation and found that sentencing someone to prison had no effect on their likelihood of a violent conviction within five years of release. The short-term incapacitation effect exists, but it is small: you lock up sixteen people, and one crime doesn't happen. The other fifteen did nothing except cost taxpayers roughly $40,000 each per year, sitting in cells that could have been left empty if anyone in the system had a reliable method for distinguishing the one who would reoffend from the fifteen who would not.

That ratio should haunt anyone who sets criminal justice policy, but the people who set criminal justice policy have never been shown this number in a budget meeting, and if they had been, the political incentives would still point toward severity over arithmetic.

The United States spends over $80 billion annually on incarceration. The National Academy of Sciences convened a committee to ask whether that investment paid off. Their 2014 conclusion: the quadrupling of America's prison population over four decades produced a "modest" effect on crime, accounting for perhaps 10 to 25 percent of the 1990s crime drop. Crime rates kept falling through the 2000s and 2010s while incarceration rates plateaued. Jeremy Travis, who chaired the NAS committee, summarized it bluntly: states and the federal government spend $80 billion a year on a system whose scientific payoff remains unclear.

Meanwhile, criminology has converged on findings that do replicate, findings backed by randomized controlled trials and systematic reviews spanning decades. The problem has never been evidence. It has always been execution at scale, in real cities, with real budgets and real elections and real political incentives pushing in the opposite direction.

What the Research Actually Shows

Daniel Nagin's review for the National Academy of Sciences, published across multiple papers and synthesized in a landmark 2013 article in Crime and Justice, established what is now the most replicated finding in deterrence research: the certainty of being caught deters crime; the severity of punishment, at the margins, does not. Longer sentences do not produce meaningful additional deterrent effects beyond the initial arrest and conviction, and the marginal deterrence from each additional year of imprisonment asymptotes toward zero in virtually every dataset that has been examined. Two days in jail, imposed swiftly and reliably, deters as effectively as six weeks imposed slowly and unpredictably.

Five strategies have accumulated the strongest experimental or quasi-experimental evidence for reducing violent crime. Each has been tested with randomized controlled trials or rigorous quasi-experiments, and each works through a different mechanism. None requires building a single new prison bed.

Focused deterrence has been subjected to more rigorous evaluation than almost any other crime prevention strategy. Braga and colleagues conducted a systematic review of all available RCTs and quasi-experiments, and the results showed significant crime reduction across every study examined. Group Violence Intervention programs targeting criminally active gangs produced the largest effects. McGarrell et al. (2006) found a 34.3% reduction in the monthly homicide rate in Indianapolis after implementing a "pulling levers" strategy that combined targeted enforcement with direct communication and social service offers to gang members. Critically, the meta-analysis found a 13.7% diffusion of crime control benefits to untreated nearby areas, meaning violence didn't just relocate to the next block over, a finding that distinguishes focused deterrence from most policing strategies where displacement is the perennial concern.

Hot spots policing exploits a remarkably consistent finding: roughly 5% of city addresses generate about 50% of crime calls, a concentration so extreme that targeting resources at those micro-locations produces consistent reductions without displacement to nearby blocks. Braga et al. (1999) demonstrated this in an RCT: problem-oriented policing at violent crime hot spots reduced violent crime, property crime, disorder, and drug activity with no evidence that criminals simply moved next door.

Place-based environmental remediation is the approach that sounds least like criminal justice and works among the best. Charles Branas and his team at Columbia ran a citywide cluster randomized controlled trial in Philadelphia, randomly assigning 541 vacant lots across 110 geographic clusters to greening, basic mowing and trash cleanup, or no intervention. Both treatments significantly reduced shootings resulting in injury or death: greening by 6.8%, mowing by 9.2%, with no evidence of displacement to adjacent areas. In neighborhoods below the poverty line, the effect was dramatic: gun violence fell nearly 29%. The cost per lot for basic mowing runs approximately $1,300 per year through Philadelphia's LandCare program, which means each lot's annual maintenance costs roughly one-thirtieth what it takes to imprison a single person for the same period. This is one of the cleanest RCTs in criminology, and it shows that changing the physical environment where violence concentrates can reduce it as effectively as any policing strategy.

Swift-certain-fair sanctions represent the purest test of Nagin's certainty-over-severity principle. Hawaii's HOPE program, evaluated in a 2009 NIJ-funded RCT by Angela Hawken and Mark Kleiman, assigned high-risk probationers to either standard supervision or a regime of frequent random drug testing backed by immediate, short jail stays for every violation. HOPE probationers were 55% less likely to be arrested for a new crime, 72% less likely to use drugs, and 53% less likely to have probation revoked. However, intellectual honesty demands a caveat: the multi-site HOPE Demonstration Field Experiment (Lattimore et al. 2018), a four-site RCT with 1,500 probationers, failed to replicate the original results. Recidivism outcomes were largely similar between HOPE and standard probation across all four replication sites, suggesting the approach may depend on specific judicial leadership and local implementation quality in ways that resist standardization through policy manuals alone.

Cognitive behavioral therapy rounds out the evidence base with a different mechanism entirely, targeting the distorted cognitions that sustain criminal behavior rather than the external environment. Lipsey and Landenberger's meta-analysis of 14 controlled studies found that CBT programs for offenders reduced the odds of recidivism to roughly 55% of control group levels. A 2022 Lancet Psychiatry meta-analysis of 29 RCTs confirmed the overall effect (pooled OR 0.72), though it attenuated when limited to larger trials. The NIJ's own review found CBT more effective at reducing criminal behavior than punishment-based or deterrence-based interventions.

The Execution Gap

If this evidence is so strong, why aren't these approaches universal? Because they are all operationally brutal to implement at scale. Focused deterrence requires identifying the specific 200 people in a city of 500,000 who are driving most of the gun violence, then coordinating law enforcement, social services, and community organizations to deliver a unified message to each one. Hot spots policing requires shift-by-shift reallocation of patrol resources to micro-locations that change over weeks, not the static beat assignments most departments use. Environmental remediation requires identifying, prioritizing, and maintaining thousands of vacant lots across a city year after year, decade after decade, without the political glamour that comes with announcing arrests or the media coverage that follows a homicide spike. Swift-certain-fair requires detecting every probation violation within hours, scheduling a hearing within days, and executing a sanction within the week.

Every one of these bottlenecks is a logistics problem, and logistics is where AI excels, not because the algorithms are sophisticated but because the operational demands of evidence-based crime prevention are precisely the kind of multi-source data integration, real-time pattern recognition, and resource optimization problems that machine learning handles better than any spreadsheet or shift sergeant ever could.

Where AI Actually Fits

The first distinction to make is critical, and it is the distinction that most media coverage of AI in criminal justice gets catastrophically wrong, conflating two completely different applications with completely different evidence bases and completely different civil liberties implications. The question is not whether AI can predict who will commit a crime. That path leads to COMPAS, which Dressel and Farid (2018) demonstrated is no more accurate than random Mechanical Turk workers given a defendant's age and prior record. Sixty-five percent accuracy, and a two-feature linear classifier matches the 137-feature proprietary algorithm. Predicting individual human behavior remains, empirically, barely better than a coin flip.

The productive question is different: can AI make the evidence-based strategies that already work operationally feasible at scale?

Geographic prediction, not individual prediction. Chattopadhyay and Evans at the University of Chicago built a spatiotemporal model that predicts where crime will occur, not who will commit it, using publicly available crime report data divided into spatial tiles roughly 1,000 feet across. Published in Nature Human Behaviour in 2022, the model predicts crime one week in advance with approximately 90% accuracy across eight major U.S. cities. Crucially, the same research team used a companion model to detect enforcement bias: they found that crime in wealthier neighborhoods led to more arrests, while equivalent crime in poorer neighborhoods did not. The tool simultaneously enables better resource allocation for police departments and exposes the systemic bias in how those resources are currently distributed, a combination that no previous crime analysis framework has delivered from a single model.

Computer vision for environmental monitoring. The Branas RCT proved that remediating vacant lots reduces violence, but the operational barrier is identifying which of a city's tens of thousands of vacancies need attention first. is identifying which lots need remediation across a city with 44,000 vacancies, a task that currently requires teams of inspectors driving every block and recording conditions by hand, a process so expensive and slow that most cities do it once every five to ten years, if they do it at all. Deep learning models using street-view imagery can now detect abandoned houses and urban blight automatically. Researchers at the University of Missouri-Kansas City built an ensemble of deep learning classifiers, including ResNet-50, that identifies abandoned structures from Google Street View images. A separate team using YOLOv5 ran inference on 114,000 street-view images across San Francisco, Mexico City, and South Bend to generate neighborhood-level urban decay indices and track change over time. Mao et al. used semantic segmentation of high-resolution satellite imagery to automatically identify vacant land across 36 Chinese cities. These systems can scan an entire city in hours, producing a prioritized remediation queue that would have taken a team of field surveyors months to compile and updating that queue continuously as conditions change.

Network analysis for focused deterrence. The hardest part of focused deterrence is identifying the small number of individuals and group structures driving violence in a city. Social network analysis using arrest records, co-offending data, and field intelligence contact cards can map these structures computationally. Andrew Papachristos at Northwestern has demonstrated that gunshot victimization clusters in social networks with the same mathematical properties as infectious disease transmission: knowing who someone associates with predicts their risk of being shot far better than demographic characteristics alone. AI applied to co-arrest and co-location data can identify the 200 people who need a focused deterrence intervention out of a population of hundreds of thousands, without the months of manual intelligence work that currently bottlenecks these programs.

Automated compliance monitoring for swift-certain-fair. The HOPE program's original success rested on a simple operational achievement: every violation detected, every detection producing a hearing, every hearing producing a sanction, all within days. That requires a system that never drops a case, never loses a court date, never lets a failed drug test sit unprocessed for three weeks because a probation officer has 150 files on their desk. Case management systems with automated flagging, scheduling, and escalation can operationalize the swiftness and certainty that Kleiman identified as the active ingredients, turning what was a judicial personality trait in one Hawaii courtroom into a systemic capability.

The Original Contribution: A Cost-per-Crime-Prevented Comparison

By combining the effect sizes from the primary research with published cost data, we can estimate the cost per violent crime prevented for each approach. These numbers are approximate, built from the midpoints of reported ranges, and should be treated as order-of-magnitude comparisons rather than precise figures.

Strategy	Effect Size (Source)	Estimated Annual Cost	Est. Cost per Violent Crime Prevented
Incarceration (incapacitation)	16:1 ratio (judge randomization study)	$640,000 (16 × $40,000/prisoner/yr, BJS 2020)	~$640,000
Focused deterrence (GVI)	34% homicide reduction (McGarrell 2006)	$1M–$3M program cost/city	~$20,000–$60,000
Hot spots policing	10–25% crime reduction (Braga meta-analysis)	Near-zero marginal (reallocation)	~$2,000–$10,000
Vacant lot remediation	29% gun violence reduction in poverty areas (Branas RCT)	~$1,300/lot/year (PHS LandCare)	~$5,000–$15,000
CBT for offenders	OR 0.72, ~28% recidivism reduction (Lancet Psychiatry 2022)	$2,000–$5,000/participant	~$15,000–$40,000

The ratios diverge by one to two orders of magnitude. Incarceration-based incapacitation costs roughly 30 to 300 times more per crime prevented than the alternatives with the strongest evidence. AI's contribution is not to change these ratios but to remove the operational barriers that prevent cities from capturing them. A city that can identify its hot spots in real time, detect vacant lot blight from satellite imagery, map its violence networks computationally, and automate its probation compliance monitoring has the infrastructure to run every evidence-based strategy simultaneously, at scale, for a fraction of what it currently spends warehousing people at $40,000 a year.

What Could Go Wrong

Everything that has already gone wrong in this field serves as a warning. PredPol, the Los Angeles Police Department's predictive policing tool, was retired in 2021 after audits showed it directed officers disproportionately to Black and Latino neighborhoods, creating a feedback loop where more policing generated more arrests which generated more data which directed even more policing to the same communities. Chicago's "Strategic Subject List" placed 56% of the city's Black men aged 20 to 29 on a risk list. COMPAS's false positive rate for Black defendants was 40.4%, compared to 25.4% for white defendants, a disparity that persists whether humans or algorithms make the prediction.

The lesson is specific and worth stating clearly: person-level prediction is the failure mode, the place where every system that has crashed and burned made its fatal error. Every system that drew public outrage and justified criticism tried to predict who would commit a crime, and every one either failed at accuracy or succeeded at amplifying bias, or both. The systems with the strongest evidence predict where resources should go (hot spots), what environments breed violence (blight detection), and whether institutional processes are running on time (compliance monitoring). The distinction matters enormously, because place-based and process-based predictions do not carry the civil liberties freight of person-based predictions, and no one's liberty is threatened when an algorithm recommends mowing a vacant lot or flags a probation violation report that has been sitting unprocessed on someone's desk for eleven days.

Bias auditing is non-negotiable. The Chattopadhyay model's most valuable output was not its crime predictions but its detection of enforcement bias across socioeconomic lines. Any AI system deployed in criminal justice should be required to audit the system it is embedded in, not just optimize it. If the algorithm surfaces the fact that arrests drop in poor neighborhoods when crime rises, that finding should reach city council, not just the police chief.

Limitations

The cost-per-crime-prevented estimates in our comparison table rely on published effect sizes from specific studies conducted in specific cities. The McGarrell focused deterrence result comes from Indianapolis; the Branas lot remediation result from Philadelphia; the HOPE result from Honolulu. Generalizability across jurisdictions is uncertain. The HOPE replication failure is the strongest evidence that context-dependent effects are real and large.

The $40,000 per-prisoner annual cost used in our table is a national average from BJS, and actual costs vary enormously by state, from roughly $25,000 in some southern states to over $69,000 in California and past $100,000 in New York. Using California's figure would make incarceration's cost-per-crime-prevented ratio even worse; using Alabama's would improve it, though it would still be an order of magnitude above the alternatives.

We do not have published cost data for AI-augmented versions of these programs because they have not been deployed at scale. The claim that AI makes evidence-based strategies operationally scalable is architecturally sound but empirically untested at the city level, which means it remains an engineering argument about what should work rather than a proven result about what does.

The Strongest Counterargument

The best case against this entire framework is that violent crime has already dropped roughly 50% from its 1990s peak, and the current system, whatever its inefficiencies, presided over that decline. Perhaps the $80 billion buys more than the NAS committee's "modest" label suggests. Perhaps incapacitation effects compound in ways that quasi-experimental designs, which measure marginal effects, cannot capture. William Spelman's research attributes roughly 25% of the 1990s crime drop to increased incarceration, which, at the peak, represented hundreds of thousands of fewer violent victimizations per year. If you are one of those potential victims, the cost-per-crime ratio matters less than the absolute number of crimes prevented. And the political economy is real: evidence-based programs require sustained multi-agency coordination that few city governments can maintain across election cycles. Prisons, whatever their inefficiency, are operationally simple. You build them. You fill them. The incapacitation effect is automatic. Sophistication is the enemy of durability in government.

What You Can Do

If you work in city government: Request a hot-spots analysis from your police department. Most have the data; few have run the analysis. The UChicago model's code and methodology are published. A competent data analyst can replicate the geographic prediction framework using your city's existing crime report data in weeks.

If you work in criminal justice: Audit your probation violation response times. Measure the gap between violation detection and sanction. If it exceeds two weeks, the swift-certain-fair mechanism is broken, and no amount of severity will compensate. Modern case management systems can close this gap without new legislation.

If you fund research: The field needs AI-augmented focused deterrence RCTs. The underlying strategy has strong evidence. The AI components for network identification and resource allocation exist. Nobody has run a trial combining them. That trial would be the most important contribution to evidence-based crime prevention in a decade.

If you are a resident: Vacant lot remediation is the most directly actionable intervention on this list. Organize a cleanup. Push your city council to fund systematic lot maintenance in high-violence neighborhoods. The Branas RCT demonstrated that even basic mowing and trash removal, without full greening, significantly reduced shootings. The barrier is not cost or technology. It is political will.

The Bottom Line

Criminology has spent 50 years converging on a simple finding: certainty beats severity, place matters more than punishment length, and focused interventions outperform broad ones by an order of magnitude on cost-per-crime-prevented. The reason these findings gather dust in journals while prison budgets grow is not ideological resistance. It is operational complexity. Focused deterrence requires identifying 200 people in a city of 500,000. Hot spots policing requires reallocating patrols shift by shift to locations that change weekly. Environmental remediation requires scanning 44,000 vacant lots and triaging them by violence risk. Swift-certain-fair requires catching every violation and responding within days, every time, for years.

AI solves none of the political problems. It solves every one of the logistics problems. And in a field where the evidence is clear but the implementation is stuck, logistics is the bottleneck that matters.