# Predictive Policing Predictive policing applies forecasting to crime data. Two families: **Place-based.** Forecast where offenses will occur and direct patrol accordingly. Borrows earthquake-aftershock models — the near-repeat hypothesis, that an offense at a location raises short-term risk nearby. **Person-based.** Score individuals for likelihood of involvement in violence, as victim or offender, from arrest records, associations, and network position. Chicago's "Strategic Subject List" was the best-known American deployment. ## The feedback loop The central technical objection is not that the models are unsophisticated. It is that **the training data measures police activity, not crime.** Recorded crime is a function of two things: offenses committed and police presence to observe them. For offenses discovered through patrol — drug possession, low-level disorder, weapons found in stops — recorded incidence is dominated by the second. Send more officers somewhere and recorded crime there rises whether or not more crime occurred. So: 1. Historical data shows more recorded offenses in over-policed neighborhoods. 2. The model directs patrol to those neighborhoods. 3. Additional patrol generates additional recorded offenses. 4. The new records enter the training data, confirming the model. **The system is not predicting crime. It is predicting its own past deployment, and each cycle raises its apparent accuracy.** From inside, this looks like validation. The model's forecasts are borne out. That the model caused what it forecast is invisible in the metrics, because the metric is recorded incidents. This is [[Statistical Discrimination]]'s self-fulfilling loop with an unusually tight cycle and an institutional actor with the power to close it. The loop is weakest for offenses with high reporting rates independent of patrol — homicide is reported regardless of where officers are stationed. **This is why the honest defense of place-based prediction restricts it to serious reported crime, and why deployment on drug and disorder offenses is the indefensible case.** ## The person-based problems Person-based systems add distinct failures: **Network contamination.** Scores incorporating associates' records mean a person can be flagged for whom they know. In segregated, heavily-policed neighborhoods, having arrested acquaintances is close to universal. **Opacity to the subject.** People placed on such lists have generally not been told, not been given the basis, and had no route to contest it. **Ambiguous purpose.** "At risk of involvement in violence" conflates likely victim and likely offender, and the intervention differs enormously depending on which is meant. In practice the list functioned as a police-attention list regardless. Chicago's list was ultimately discontinued after evaluation found no detectable reduction in violence among those on it, alongside sustained criticism of its opacity and racial composition. ## Where it stands Several major departments — Los Angeles, Chicago, and others — have discontinued named programs after audits, litigation, or evaluations found unproven benefit and documented disparate impact. Some cities have restricted the practice by ordinance. But the capability migrated rather than disappeared. Gunshot detection, license-plate reader networks, facial recognition, social-media monitoring, and integrated data platforms (see [[Palantir]]) perform overlapping functions without the "predictive policing" label that attracted scrutiny. **The critique attached to a name, and the name was retired.** Tracking the function rather than the branding is the necessary discipline here. ## The general lesson Predictive policing is the sharpest illustration of a principle that generalizes across this branch: **when a model's outputs influence the world that generates its training data, ordinary validation breaks down.** Accuracy measured against subsequently-collected data cannot distinguish a good forecast from a self-executing one. Establishing whether such a system works requires an experimental design — randomized deployment, external outcome measures — that is expensive, politically awkward, and almost never done before procurement. ## Related [[Statistical Discrimination]] · [[Criminal Risk Assessment Instruments]] · [[Credit Scoring]] · [[Fairness Impossibility Results]] · [[Algorithmic Bias]] · [[Surveillance Capitalism]] · [[Palantir]] · [[Mathematics and Power Hub]]