The Quiet Boom in Predictive Policing Nobody Voted For

Somewhere in America right now, a software company is selling your city a crystal ball. It costs six figures a year, runs on data you never agreed to share, and comes with a dashboard that promises to tell cops where crime will happen before it does. Predictive intelligence targeting teams are the hottest product in law enforcement procurement, and almost nobody is asking who actually benefits. Here's the pitch, stripped of jargon. Feed years of arrest records, 911 calls, and "intelligence" into a machine learning model. The model spits out lists of people and places flagged as likely to be involved in future violence. A "targeting team" then pays those people a visit, offers services, or applies pressure. The sales deck calls it proactive. The fine print calls it something closer to pre-crime. On paper, the results look impressive. Cities from Chicago to New Orleans have touted double-digit drops in shootings after deploying these systems. Departments love the narrative: we're not just reacting, we're predicting. Politicians love it more, because a falling homicide number is the only crime statistic voters actually notice. Now the skeptical part. Correlation in a city that's also flooding a neighborhood with overtime officers, federal task force money, and a new violence-intervention nonprofit is not proof that the algorithm did anything. Several of these programs launched alongside massive pandemic-era federal grants. When the money and the model arrive together, attributing the drop to the software is a choice, not a finding. Then there's the data problem. These models train on arrests, not crimes. Arrests happen where police already patrol. So the machine learns to flag the same over-policed blocks and the same names that were already in the system, often from juvenile stops that never led to a conviction. Feed in bias, get out bias, just faster and with a vendor's logo attached. Who profits? Follow the contracts. Palantir, SoundThinking (formerly ShotSpotter), and a growing crop of smaller analytics firms are cashing checks from departments that can't easily evaluate what they're buying. Many contracts are sole-sourced, because the methodology is proprietary. That means no independent audit, no public accuracy score, and no way to know the false positive rate. If a model flags 500 people and 495 are wrong, nobody publishes that number. The human cost is real. Being on a "target list" can mean unannounced visits, extra scrutiny, and a permanent data trail. A 2023 investigation into one such program found that people flagged had often committed no crime at all. They were just statistically unlucky. Look, nobody wants to defend violent crime. But "it might work" is a thin justification for building a surveillance apparatus that can't be turned off and can't be inspected. The vendors will keep selling certainty. Cities will keep buying it. The rest of us will keep paying for it, in dollars and in liberty. So before your mayor signs the next contract, ask three questions. What's the false positive rate? Has anyone independent checked it? And if the model is wrong about you, what's your recourse? If the answers are vague, you're not buying safety. You're buying a receipt. The uncomfortable truth is that predictive targeting isn't failing because it's evil. It's succeeding because it's convenient. Convenience is how surveillance becomes permanent, one budget cycle at a time.