The ShotSpotter alert that jailed Michael Williams began as a firecracker

The Sunday ride that ended in a murder charge

On a Sunday evening in late May 2020, Safarian Herring, 25, flagged down a car on the South Side of Chicago and asked the driver, Michael Williams, for a ride. As Williams drove, Herring was shot in the passenger seat, and he died at a hospital days later.¹

That August, roughly three months after the shooting, police arrested Williams, 63 by his own account, and charged him with first-degree murder.² The case prosecutors built leaned on two pieces of machine evidence: a surveillance video that showed Williams's car, and a loud bang that a network of acoustic sensors had recorded and classified as a gunshot.¹ Williams spent nearly a year in Cook County Jail, about eleven months, before prosecutors asked a judge in July 2021 to dismiss the charge for insufficient evidence. He was never tried and never convicted.¹

How a bang becomes a gunshot alert

ShotSpotter, the gunshot-detection system then sold by the company now called SoundThinking, is a network of microphones fixed to poles and buildings across a city. When the sensors pick up a loud, impulsive sound, software estimates its location by comparing the moment each sensor heard it, and classifies whether the sound was likely gunfire.³ A human then makes the final call: an acoustic expert in the company's around-the-clock review center listens to the recording and decides whether to confirm gunfire and alert police.⁴

That human step is where the Williams alert came apart. Court filings, reported by the Associated Press, show two decisions a person made after the sound was recorded.¹

The system had first placed the noise at 5700 South Lake Shore Drive. A senior ShotSpotter engineer, Walter Collier, later changed the reported location to the intersection where the surveillance video showed Williams's car, about a mile away.¹ The company does not deny either change. It casts them as the review step working as intended, confirming and correcting the automated read against the sensor data.⁴ Its chief executive, Ralph Clark, told the AP the details of the algorithm were "not really relevant" because "anything that ultimately gets produced as a gunshot has to have eyes and ears on it."¹ What no one disputes is that a person set the label and the location by hand.

Williams's case was not the only one. In 2016, a ShotSpotter engineer testified in Rochester, New York, that an employee had reclassified a sound the system flagged as a helicopter into a gunshot at police request. The AP reported that employees "can and do modify the location or number of shots fired at the request of police, according to court records."¹

Why an alert rarely means a gun crime

A sensor tuned to catch a rare event, and tuned so it rarely misses, will flag that event far more often than it truly happens. Most of what it flags will be something else.

The city's Office of Inspector General examined every ShotSpotter alert confirmed and sent to police between January 2020 and May 2021: 50,176 of them. Of those, 41,830 ended with a recorded disposition, and 4,556 turned up evidence of a gun-related crime. That is 9.1 percent of every alert police responded to.³

A separate study by the MacArthur Justice Center, covering about 21 months of city data, reached the same shape. It found 89 percent of ShotSpotter deployments turned up no gun-related crime, and 86 percent produced no report of any crime at all, more than 40,000 dead-end deployments in all.⁵

There is a real argument for the sensors. Most gunfire is never reported: a Brookings study cited in the inspector general's report found that in Washington, DC, only 12.4 percent of shots-fired incidents produced a 911 call.³ A system that hears what no one phones in could send help where help would otherwise never arrive. But that is why a single alert says so little. Gunfire is common enough to trip a sensor, and a recoverable gun crime is comparatively rare. So any one alert is only a lead worth checking, well short of proof that a crime happened where it points. The charge against Williams rested on one of these alerts, treated as if it had located a shooting.

The test nobody has run

The critics and the company are not measuring the same thing. The inspector general's 9.1 percent and MacArthur's 89 percent measure an outcome downstream of the alert: whether police, once they responded, found evidence of a gun crime. They do not measure whether the sensors correctly heard gunfire in the first place.³⁵

The company measures something narrower. SoundThinking reports a 97 percent aggregate accuracy rate across its customers for 2019 through 2021, with a customer-reported false-positive rate below half a percent.⁴ But that rate depends on who reports back: an alert counts as a false positive only when a police customer says no gunfire occurred, a report the MacArthur study says is rarely filed.⁴⁵ The company also rejects the outcome measures themselves, saying that when an alert turns up no physical evidence, the cause is a slow response or an uncooperative witness rather than a detection error.⁶

Both framings step around the one measurement that would settle it: an independent, controlled test of whether the sensors can tell gunfire from firecrackers, backfires, or fireworks. The MacArthur study says no such test has been published.⁵ Until one is, the 97 percent and the 9.1 percent describe different things, and neither can tell a court whether the bang outside Williams's car was a gunshot.

The inspector general found one more effect. Where alerts were frequent, some officers came to rely on that frequency itself, using how often an area lit up to justify stopping the people in it.³ An alert that means little on its own can still change who gets approached on the street.

A flag that outlived the machine

Chicago eventually pulled back. In February 2024 the city said it would not renew its SoundThinking contract, which expired that month, and it switched ShotSpotter off on September 22, 2024.⁸ Research had given the decision cover: a study of 68 metropolitan counties from 1999 to 2016, cited by the AP, found gunshot-detection systems produced no measurable drop in gun homicides or arrests.¹

Williams's case closed in stages. The charge was dropped in 2021, and in February 2026 the city agreed to pay him $500,000 to settle his federal lawsuit.⁹ SoundThinking, which was not a party to the suit, then claimed the opposite role for its technology. In a public statement, it said its own later analysis found ShotSpotter is not built to detect gunfire inside an enclosed vehicle. That analysis, it said, is what led prosecutors to release him.¹⁰ Both accounts can be true. The first alert and its human edits helped build the charge in 2020, and the company's later work may have helped undo it in 2021. Neither shows the alert ever earning the certainty a murder charge is supposed to require.

The microphones are gone from Chicago, but the failure that jailed Williams was never tied to one machine. A system produces a probabilistic flag. A person treats the flag as a fact. The fact walks into a case as evidence. The same shape recurs wherever an automated signal — an acoustic alert, a face-recognition match, a risk score — enters a case as though someone had already confirmed what it claims.