When an acoustic gunshot detection system fires and police respond but find no evidence of a shooting, the event vanishes from the crime record. That gap — between detection events and recorded incidents — has specific, measurable consequences for anyone consuming crime data professionally.
The Detection-to-Record Pipeline
Acoustic gunshot detection systems like ShotSpotter (rebranded as SoundThinking in 2023) triangulate audio signals to identify potential gunshots within seconds. A confirmed alert dispatches police. If officers arrive and find evidence of a shooting — spent casings, a victim, shell fragments — a police report is filed and the incident enters the crime record. If officers find nothing, the event typically generates an activity log but no crime report.
This is not a bug in the crime data pipeline. It is, by design, the correct outcome: no crime report is filed because no crime was confirmed. But the volume of these non-recording events, and what they imply about measured crime rates, is large enough to matter to anyone using crime data professionally.
SoundThinking adds its own filtering layer before any alert reaches a dispatcher. Company analysts review raw audio detections and reject those they classify as non-gunfire — fireworks, backfires, construction noise. The figure transmitted to police represents alerts that cleared the company's own internal review. That pre-filtered population is still the one that generates mostly non-recording events.
The Numbers from Chicago
The most comprehensive independent audit of ShotSpotter accuracy was published by the MacArthur Justice Center in May 2021, covering a 21-month period across Chicago. Key finding: ShotSpotter generated 40,453 alerts in that period. Police responded to each. In 89 percent of cases, officers found no evidence of a gun-related crime. The remaining 11 percent produced evidence of gunfire — shell casings, victims, or property damage.
The Chicago Police Department conducted a separate audit covering a different period, which found a higher evidence recovery rate of roughly 25 percent. The spread between these figures is itself a methodological finding: what counts as “evidence of a crime” varies by how strict the threshold is set.
What both analyses agree on: the large majority of ShotSpotter alerts in major urban deployments do not produce crime reports. The question crime data consumers should ask is: what is the denominator? In cities with active gunshot detection, the gap between acoustic detections, dispatched police responses, and filed crime reports is large and structurally invisible in published crime data.
What “No Crime Recorded” Actually Means
The standard logic of crime data is that a crime report is filed when an officer concludes a crime occurred. ShotSpotter alerts that produce no crime report are, from a crime data perspective, noise — events that generated police activity without a confirmed crime.
But that framing conceals a meaningful ambiguity. Gun violence undercount in crime data is a well-documented problem. The FBI's National Incident-Based Reporting System counts only crimes reported to law enforcement. Shootings where no one calls police, incidents where victims refuse to cooperate, and drive-bys with no witnesses all create a gap between gunshots fired and crimes recorded. Gunshot detection was, in part, justified as infrastructure to close that gap.
The empirical record on whether it does is mixed. A 2022 study in the Journal of Urban Economicsfound that ShotSpotter adoption did not significantly increase gun-related arrests in the 68 cities studied. A 2023 Rutgers study found a modest increase in victim-present shooting reports in some deployment areas — the alerts that produce the clearest physical evidence. Neither result supports the strong form of the claim: that gunshot detection substantially closes the gap between shootings that occur and shootings that are recorded.
The City Discontinuation Pattern
Several major US cities have canceled ShotSpotter contracts following independent reviews:
- Chicago terminated its contract in February 2023 following a city inspector general report and the MacArthur analysis. The contract had cost approximately $33 million over eight years.
- Cleveland let its contract lapse in 2022 after a review found insufficient evidence of crime reduction.
- San Francisco declined to renew in 2022.
- Detroit briefly deployed and then discontinued.
SoundThinking disputes the accuracy of these city-level assessments and has published its own efficacy data showing higher evidence recovery rates and measurable reductions in violent crime in retained deployments. The company currently operates in approximately 150 US cities and municipalities.
What the departure pattern does establish: independent accuracy assessments and the vendor's internal metrics disagree substantially. That disagreement has direct implications for how crime data derived from these cities should be interpreted — particularly in before/after trend comparisons that cross a contract termination date.
Three Data Artifacts Developers Should Know
For crime data API consumers, ShotSpotter's deployment creates three categories of noise worth understanding before building on top of incident feeds from covered cities.
1. Geographic concentration bias
ShotSpotter deployments are not random. They concentrate in high-violence neighborhoods, typically under political pressure following high-profile incidents. This means covered areas have both more audio sensors per block and more police responses per potential event. The coverage gap between a deployed neighborhood and an adjacent non-deployed one is a structural artifact of resource allocation, not a reflection of differential violence.
2. Crime report inflation in high-alert zones
When ShotSpotter alerts produce police responses that find evidence, that evidence would often not have been found without the response. Police find spent casings; a report is filed for an incident that, without the technology, would have gone unrecorded. This is the intended upside — and it produces real data. But in a before/after comparison, the “after” dataset has higher completeness in deployment areas. Year-over-year trend analysis conflates detection improvement with actual changes in crime rates.
3. Temporal spikes from system expansion
When a city expands ShotSpotter coverage to new neighborhoods, the recorded crime rate in those neighborhoods typically rises before it falls — a detection effect rather than a crime effect. Models that interpret this as a crime increase are miscalibrated.
The Calibration Problem
This is structurally similar to the AI calibration problem documented in classifier research. Academic work on LLM calibration using NEISS injury data found that model confidence scores are systematically overconfident: the model is more certain than its accuracy warrants. Crime data systems built on detection infrastructure carry a similar risk: the data appears complete because every dispatch produces a record, but the record population reflects detection architecture as much as underlying crime rates.
A well-calibrated system knows what it doesn't know. Crime data in a ShotSpotter-covered area looks more complete than in an uncovered area. That completeness is real — but uneven, and not labeled. The API response does not tell you whether a given incident was initiated by an acoustic alert, a 911 call, or officer observation. All three produce identical-looking CAD exports.
Where This Fits in the National Picture
For context on where detection debates sit in the broader crime landscape: USAFacts data puts the US violent crime rate at 359 per 100,000 in 2024 — down 5.4 percent year-over-year and approximately 49 percent below the 2001 rate. Firearm homicide hit a 30-year low. The national downward trend is clear.
But national averages mask the block-level variation that makes address-level crime data useful. ShotSpotter deployments concentrate in the highest-variance neighborhoods — the areas where ZIP-code statistics and block-level reality diverge most sharply. That is precisely where the detection-to-record gap is most consequential and most likely to distort trend signals.
What a Crime Data API Should Do With This
The honest answer: most crime data APIs do not handle this systematically, because the underlying municipal feeds do not label their source mechanisms. A police incident report filed after a ShotSpotter response looks identical, in a standard CAD export, to one filed after a 911 call. The data structure does not record whether detection technology initiated the response.
Practical mitigations for developers building on crime data feeds from ShotSpotter-covered cities:
- Treat coverage-expansion periods as potential data artifacts. If a city expands gunshot detection coverage, crime data in those areas may rise in the short term due to detection effects, not true crime increases. A 90-day smoothing window will not fully absorb this.
- Do not run before/after trend analysis in newly covered areas without noting deployment dates. The baseline completeness is different. Year-over-year comparisons that straddle a deployment start will overstate the crime increase.
- Use homicide data as an anchor.Homicides are documented regardless of detection technology — bodies are found, victims are identified, medical examiners file reports. If property crime and assault trends diverge significantly from homicide trends in a detection-expanded area, the divergence likely reflects detection effects rather than true changes in underlying violence.
- Be skeptical of sharp single-neighborhood spikes. Technology deployment produces data spikes that look like crime spikes. Context about when a city acquired or expanded coverage is a necessary input to trend interpretation.
The Transparency Gap
The broader issue mirrors the police data transparency debates that define this space. ShotSpotter's alert-to-report data is not publicly accessible in most cities. The company treats its algorithm, its internal review process, and its city-level accuracy metrics as proprietary. Independent analyses depend on public records requests — which, as Chicago's Inspector General documented, can take years and require sustained pressure to produce.
Cities that deploy ShotSpotter are augmenting their crime data infrastructure with a component whose methodology is not documented in the public record. The outputs of that component — when they generate crime reports — feed into the public crime data that developers, researchers, and the public rely on. The inputs do not.
The lesson for anyone building products on crime data is not that ShotSpotter-sourced incidents are unreliable or that the technology serves no legitimate purpose. The lesson is that the completeness and consistency of crime data is a function of the detection infrastructure behind it — and that infrastructure varies across cities, changes over time, and is not labeled in the data itself. Building without knowing that is building on a foundation you haven't fully inspected.
Access Address-Level Crime Data
Real-time incidents · SpotScore™ safety ratings · 36-month trends · 22,000+ US cities. Normalized and verified — because raw data isn't enough.