In more than 140 US cities, ShotSpotter’s artificial intelligence (AI) algorithm and complex network of microphones evaluate hundreds of thousands of sounds a year to determine whether they are gunshots, generating data that is now being used worldwide. Used in criminal cases.
However, a confidential ShotSpotter document obtained by The Associated Press outlines something the company doesn’t always advertise about its “precise surveillance system”: that its employees can quickly override and reverse the algorithm’s determinations, and They have wide discretion to decide whether a sound is a gunshot, fireworks, thunder, or something else.
Such rollbacks happen 10% of the time, according to the company’s calculations for 2021, which experts say could lead to subjectivity in increasingly relevant decisions and conflict with one of the reasons AI is used across devices. Human.
“I listen to a lot of recordings of gunshots, and it’s not an easy thing to do,” said Robert Maher, a national expert on gunshot detection at Montana State University, who reviewed the ShotSpotter document. “Sometimes it’s clearly a gunshot. Sometimes it’s just a ping, ping, ping…and you can convince yourself it’s a shot.”
Marked “WARNING: CONFIDENTIAL,” the 19-page operational document outlines how staff at ShotSpotter review centers should listen to recordings and algorithmically evaluate potential shots’ conclusions, which may require a ruling decision based on a number of factors. including whether the sound features the timing of the gunshots, whether the audio pattern resembles “a Christmas tree on its side”, and whether “the gunshots have 100% certainty in the reviewer’s mind.”
In a statement to the Associated Press, ShotSpotter said the role of humans is one of positive validation of the algorithm and that the “plain language” document reflects the high standards of accuracy its reviewers must meet.
“Our data, based on millions of incident reviews, demonstrates that human review adds value, accuracy and consistency to the review process on which our customers and many shooting victims depend,” said Tom Chittam, the company’s vice president of analytics and services. ” Forensic.
Chittum said the company’s experts have testified in 250 court cases in 22 states and that its “97% overall accuracy rate for real-time detection for all of its customers” has been verified by an analysis firm hired by the company. has gone.
Another part of the document outlines ShotSpotter’s longstanding emphasis on speed and decisiveness, a commitment to classifying sounds in less than a minute and alerting local police and 911 emergency service operators so they can locate agents. Can send on
Titled “Embracing a New York State of Mind”, it refers to the NYPD’s request to shotspotters to refrain from posting sound alerts as “shot likely” and to simply make definitive classifications as shots or no shots. .
“Bottom line: Train reviewers to be decisive and precise in their classification and try to avoid questionable publications,” the document reads.
Experts say such guidance under time pressure could encourage ShotSpotter reviewers to err on the side of classifying the sound as a gunshot, even when the evidence is not strong, potentially The number of false positives is increasing.
“Humans are not being given a lot of time,” said Geoffrey Morrison, a speech recognition scientist in Britain who specializes in forensics. “And when people are under enormous pressure, they are more likely to make mistakes.”
ShotSpotter reports that it posted 291,726 gunshot alerts to its customers in 2021. That same year, in Associated Press’s comments added to a previous article, ShotSpotter claimed that its human reviewers agreed with shots’ ratings more than 90% of the time. machines, but the company invested in its own team of reviewers “for 10% of the time they didn’t agree with the machine.” ShotSpotter didn’t answer a question on whether this ratio still holds true.
ShotSpotter operations documents, which the company argued in court for more than a year were a trade secret, were recently released by a protective order in a Chicago court case in which police and prosecutors used ShotSpotter data. did as evidence. manslaughter in 2020 for allegedly shooting a man inside his car. Michael Williams spent nearly a year in prison before a judge dismissed the case for insufficient evidence.
Evidence at Williams’ pre-trial hearing showed that ShotSpotter’s algorithm initially classified the noise picked up by the microphone as a firecracker, making a determination with 98% confidence. However, the ShotSpotter reviewer who rated the sound immediately relabeled it as a gunshot.
The Cook County Public Defender’s Office noted that the operational document was the only document ShotSpotter sent in response to multiple subpoenas for guidelines, manuals or other scientific protocols. The publicly traded company has long resisted calls to open its operations to independent scientific inquiry.
Fremont, California-based ShotSpotter acknowledged to the Associated Press that it has other “extensive training and operational materials” but considers them “confidential and trade secrets.”
ShotSpotter installed its first sensors in Redwood City, California in 1996, and relied solely on local 911 operators and police to review every possible shot until adding its own human reviewers in 2011.
Paul Green, a ShotSpotter employee who frequently testifies about the system, explained in a 2013 exploratory hearing that employee reviewers have addressed issues with a system that “has been known to sometimes give false positives”. Because it “has no ears to hear”.
“Classification is the most difficult element of the process,” Green said at the hearing. “Simply because we don’t have … control over the environment in which the shots are fired.”
Green said the company prefers to hire ex-military and police officers familiar with firearms, as well as musicians because “their ears are more developed.” His training includes listening to hundreds of audio samples of gunshots and even visiting a shooting range to become familiar with the characteristics of the shots.
As cities weigh the promise of the system against its price tag — which can reach up to $95,000 per square mile per year — company employees have detailed how its acoustic sensors on utility poles and streetlights detect strong pops, It picks up explosions or hits and then filters the sounds through an algorithm that automatically classifies whether they are gunshots or something else.
But until now, little was known about the next step: how ShotSpotter’s human reviewers in Washington, D.C. and the San Francisco Bay Area decide what’s a gunshot versus any other noise 24 hours a day.
“It is important to listen to the audio downloads,” reads the document, written by David Valdez, a former police officer and now-retired supervisor at one of ShotSpotter’s review centers. “Sometimes the audio is so close to a gunshot that it can override all the other features.”
Burke reported from San Francisco.
Garance Burke is on Twitter as @garanceburke
Michael Tarm is on Twitter as @mtarm