There is a knock at the door and no crime behind it. In Chicago, for years, a computer produced a "heat list" — a ranked roster of people the algorithm judged most likely to be involved in gun violence, as shooter or victim. If your name climbed high enough, officers came to your home to tell you the machine was watching you. You had done nothing. The prediction was the event. You were now a person of interest to a system that could not point to a single act you had committed — only to a number that said you might.
This is not a dystopian film. This is a procurement line item. And once you see how the trick works, you can never un-see it in the daily news.
The statistics laundry
The pitch is seductive: crime isn't random, it clusters, so let the data show us where to send the cars. PredPol — the best-known vendor, now rebranded Geolitica — took historical crime reports, ran them through an earthquake-aftershock model, and printed little red boxes on a map. Police were told to patrol the boxes. Objective. Data-driven. Free of the messy prejudice of human cops. That was the promise.
Here is the rot inside it. The model is trained on reported crime and recorded arrests — not on crime itself. And arrests are not a mirror of wrongdoing; they are a mirror of where police already went. Send more officers to a neighborhood and you record more arrests there, for the ordinary reason that you can only catch what you're standing next to. Feed those arrests back into the model and it says: send more officers here. Which produces more arrests. Which trains the model harder on the same blocks. The system does not discover where crime is. It discovers where you have been looking, and commands you to look there forever.
Engineers have a plain name for this shape: a feedback loop with no ground truth. The output becomes the next input, the error compounds, and the model grows more confident precisely as it grows more wrong. A study of a PredPol model applied to Oakland drug-use data found it would have sent police overwhelmingly into Black neighborhoods, though survey data showed drug use spread evenly across the city. The bias of the past didn't get corrected by the math. It got encoded, amortized, and automated — printed on a map with the authority of a decimal.
From place to person
A red box on a map is bad enough. The next move is worse: predicting people.
Chicago's Strategic Subject List scored individuals on their supposed risk. Reporting found that a large share of the people on it had never been arrested for anything, yet there they were, ranked, watched, doorstepped. In the UK and across Europe, similar "risk assessment" tools score whether an individual is likely to reoffend — and courts have used such scores to shape bail and sentencing. In the US, the COMPAS risk tool became infamous when analysis suggested it flagged Black defendants as future criminals at roughly twice the false-positive rate of white defendants. The number wore the robe of objectivity. Underneath, it was the same old prejudice, now unchallengeable because "the algorithm found it."
Sit with what that means. In the old order, suspicion had to be earned — you had to do a thing, and then be observed doing it, and only then were you a suspect. Predictive policing inverts the arrow of justice. Suspicion arrives first, generated by a correlation you never chose to be part of, attached to you because of where you live and who you know and what a model inferred from ten thousand strangers who share your pattern. You are presumed guilty of the future.
Our record: The scale of Maat weighs a deed — a heart against a feather, an act that was actually committed, named and answerable. It cannot weigh a deed that has not occurred; there is nothing on the pan. Predictive policing puts a phantom on the scale — not what you did, but what a machine guesses you might — and calls the tipping of that scale justice. This is the signature of Isfet: it keeps the ritual of judgment and discards its only lawful object. Apophis, the serpent of entropy, does not need to break the law of Maat directly. It only needs to convince the weighers to judge shadows, and let the innocent hang for the crimes of a probability distribution. A world that punishes the not-yet-done has stopped weighing and started guessing — and dressed the guess in a badge.
Why it spreads even when it fails
You'd think a tool this broken would be abandoned. Some cities have dropped it — Los Angeles ended its PredPol program, others quietly followed. But the logic keeps returning under new names, and the reason is not that it works. It is that it launders responsibility.
A police chief who over-patrols a neighborhood on a hunch owns that decision. A police chief who over-patrols because "the model directed resources there" has outsourced the blame to a black box. The algorithm becomes an accountability sink — a place to pour hard decisions so no human has to sign them. That's the real product. Not better prediction. Deniability at scale. And deniability is addictive to any institution that would rather not be asked why it does what it does.
The lever
Refuse the phantom, because this machine has one structural weakness: it runs on your data and hides behind your ignorance, and both can be taken away from it.
Attack the input. Predictive policing is only as powerful as the surveillance feeding it — the license-plate readers, the phone dumps, the social-media scraping, the fused location trails. Every camera you help pull out of a public contract, every data-broker sale you help outlaw, every "smart city" sensor you help subject to consent starves the model at its root. No data, no prophecy.
Attack the opacity. Demand that any predictive tool used against citizens be auditable — its inputs, its error rates by group, its false-positive count published, or it doesn't get deployed. Several jurisdictions now require exactly this, and vendors hate it, because a model forced to publish its miss rate usually can't survive the embarrassment. Sunlight is a scandalously effective disinfectant against a scared vendor.
And hold the line no algorithm can cross: a person is answerable for deeds, never for correlations. Refuse, loudly and in every venue, the premise that a probability is a cause and a pattern is a crime. That refusal is not soft. It is the exact boundary between a society of laws and a society of odds.
You cannot be guilty of tomorrow. Weigh the deed, not the shadow it might cast.