In January 2024, Chicago police officers responded to a ShotSpotter alert on the South Side, saw a flash in the dark, and one officer fired at a boy who turned out to be lighting fireworks. The department’s oversight agency later confirmed there was no gun, no weapon recovered, and no reason for the stop. That is not a rare malfunction of police AI. The audits that exist say it is closer to the normal case.
Here is the answer up front. Police AI in 2026 means two things above all: gunshot detection (mostly SoundThinking, formerly ShotSpotter) and predictive policing (mostly a company called Geolitica, formerly PredPol, which no longer exists). Both got sold to cities for years on numbers that the vendors produced about their own products. The few independent audits that were actually allowed to happen found that most alerts lead to nothing, that the crime-reduction claims do not survive outside scrutiny, and that the systems rarely worked as advertised.
The bigger problem is not that the software fails. It is that almost nobody checked, and the places that did check mostly dropped the tool afterward. Here is what the record shows, question by question.
What counts as police AI in 2026?
Four things get lumped under the label:
- Gunshot detection. Acoustic sensors on poles estimate where a shot was fired and dispatch officers. SoundThinking markets ShotSpotter to more than 100 agencies.
- Predictive policing. Software that forecasts where crime will happen or who will commit it, so patrols go there. PredPol was the best known; the category is now largely dormant in the United States.
- Plate readers and face recognition. Camera networks that identify cars and faces. This is a different story with a different paper trail, covered in our companion piece on public-space surveillance.
- Report writing and triage. Newer tools that draft reports or route calls. They are newer, so they have even less of an audit record than the first two.
The two with real deployment history, and real audits, are gunshot detection and predictive policing. They are also where the claims and the evidence diverge the most.
Does gunshot detection actually work?
ShotSpotter is the test case, because Chicago ran one of the largest deployments in the country. It covered more than 100 square miles with over 500 sensors under a three-year, $33 million contract, according to reporting by Wired in February 2024. If acoustic detection could cut gun violence anywhere, it should have shown up there.
It did not. The numbers from independent reviewers point the same direction:
- Chicago, 2021: the MacArthur Justice Center reviewed more than 40,000 ShotSpotter dispatches over about 21 months and found that 89% led to no evidence of a gun-related crime, and 86% led to no crime at all.
- Chicago, 2021: the city’s Office of Inspector General, in an August 2021 report, found officers had begun stopping and searching people solely because they were in an area with many ShotSpotter alerts.
- New York City, 2024: a Comptroller audit of June 2023 found the NYPD responded to 940 alerts, that 82% could not confirm a shot was fired, and that 5% were unfounded or false positives.
- Houston, 2023: a Houston Chronicle investigation found more than 80% of alerts were false positives or otherwise useless. Nearly 5,500 alerts produced 99 arrests, about half for misdemeanors.
- Chicago and Kansas City, 2024: a multi-agency study published through the federal Office of Justice Programs found the systems did not reduce shootings or crime, or increase clearance rates.
- Nationwide, 2024: a Northeastern University study found the technology increased detection of gunfire but did not reduce gun violence or improve the clearance of gunshot cases.
Chicago stopped renewing the contract in February 2024, and its coverage ended later that year. Houston’s mayor called the system a gimmick in May 2024 and tried to end the deal early. More than a dozen cities have now walked away.
The honest reading is not that the sensors never work. They do detect gunfire. The problem is that a detected bang is not the same as a solved crime, and the gap between the two has been documented everywhere someone looked.
Does predictive policing work?
PredPol began as a Los Angeles Police Department project with a UCLA professor, and its algorithm was borrowed from a model that predicts earthquake aftershocks. The idea: heavy policing in small boxes where crime just happened. By 2020 it was the most widely used predictive policing product in the country.
The evidence never caught up to the sales pitch. In March 2019, the LAPD’s own internal audit concluded there was insufficient data to say whether PredPol reduced crime. In April 2020, the department ended the program without ever being able to measure whether it worked.
Two investigations made the picture worse. Reporting by Gizmodo and The Markup in December 2021 found the software targeted Black and Latino neighborhoods while generating almost no predictions for wealthier white areas, which is the opposite of a bias-free system. In October 2023, The Markup found that predictions the algorithm generated for a Plainfield, New Jersey police department were accurate less than half of one percent of the time.
The company itself did not survive the scrutiny. Geolitica, as PredPol was renamed, ceased operations at the end of 2023. Its engineering team, patents, and customer list were absorbed by SoundThinking, the maker of ShotSpotter. Read that twice: the two flagship police AI vendors collapsed into one company, which now largely shapes the numbers the industry reports about itself.
So why are the audits so thin?
This is the part that matters most, and it is structural.
Almost every accuracy figure attached to these tools comes from the company selling them. ShotSpotter has claimed a 97% accuracy rate; internal testimony in a 2017 case described the guarantee as a marketing product rather than an engineering one. The economist Jennifer Doleac told Voice of San Diego that the company resisted rigorous independent evaluation, and noted it clearly found it could get cities to sign contracts without such evidence.
Departments often do not keep their own data either. A 2022 investigation found Savannah, Georgia had spent $489,000 on ShotSpotter without tracking whether it worked. The Brennan Center has argued that even basic audit logs, who runs a prediction and who sees it, are rarely kept or disclosed, which makes outside review impossible even when someone wants to try.
That leaves a short list of real audits. The Chicago Inspector General report and the New York City Comptroller audit are the exceptions, and both found most alerts went nowhere. When the rare independent look happens, the tool looks worse, not better.
Regulators have started to treat this as the point. The European Union’s AI Act bans predictive policing built on profiling people and real-time remote biometric identification in public spaces, with those prohibitions taking effect from February 2025. The law’s premise is simple: systems that expand police power need proof before deployment, not after.
The cameras are the other half of the story
Gunshot sensors and prediction software are only one layer of automated policing. The larger network is plate readers and face recognition, and it has its own audit trail of wrongful stops and national search rules.
What should the record change?
Two lessons stand out. First, a detection is not an outcome. Finding a bang, a face, or a hot spot is the easy half; the hard half is proving it made anyone safer, and that half was skipped for a decade.
Second, the audit gap is the product, not a bug. A tool that resists measurement is not a neutral tool. It shifts risk onto the people in the coverage area, who absorb the stops and the false alerts while the vendor keeps the headline number. If a city cannot independently test a police AI system, the honest assumption is that it has chosen not to.
The Short Version
- Gunshot detection: Chicago’s own reviews found roughly 89% of alerts led to no gun-related crime, and NYC’s 2024 audit found 82% of alerts could not confirm a shot. Chicago ended its contract in 2024.
- Predictive policing: the LAPD shut its program down in 2020 without ever proving it worked, and the vendor, Geolitica, shut down in 2023.
- The accuracy numbers: almost all come from the vendors. The rare independent audits find most alerts lead nowhere.
- The audit gap: many departments never tracked effectiveness at all, and audit logs of who uses the predictions are rarely kept.
- The stakes: the failure mode is not wasted tech, it is armed officers dispatched on bad signals into the same neighborhoods, again and again.
FAQ
Is ShotSpotter the same thing as AI?
Partly. The sensors are acoustic hardware, but the step that classifies a sound as gunfire and estimates its location is an automated model, and it has been the subject of disputes over reclassification of evidence.
Do any of these tools reduce crime?
The best available independent studies found gunshot detection increased the detection of gunfire without reducing gun violence or improving case clearance. The crime-reduction claims rest mostly on vendor figures.
Has any city kept the system and shown it works?
Some cities still run ShotSpotter. None has published an independent evaluation that survives the kind of review Chicago and New York produced.
Who is watching now?
Rarely anyone independent. Vendor self-reporting is still the norm, which is exactly why the European Union wrote a ban into law rather than a reporting requirement.
About The Author
