// Contents

At 3:18 p.m. on November 4, 2026, sheriff’s deputies in Gillette, Wyoming, pulled over a car outside an apartment complex and arrested a woman with an outstanding warrant. In the CrimeRadar app, that same minute became a “Fatality Report”: a man killed behind the building, a possible homicide, multiple units responding. A few hours earlier, the same app had turned an abandoned car in a Wisconsin parking lot into a shooting.

“No, not even close,” Campbell County Sheriff’s Lt. Dan Maul told Cowboy State Daily after checking the records. Nobody died, and there was no homicide investigation. There was a traffic stop.

CrimeRadar listens to open police, fire and EMS channels, turns the radio chatter into text and has an AI summarize it. The result is a map of incidents with alerts about what is happening near your home. Cowboy State Daily put its Android downloads at about 4.6 million. On the App Store it is rated 4.9 out of 5: people like hearing about a shooting down the street before the news does.

How “I’m on scene” becomes a school shooting

You can trace CrimeRadar’s mistakes word by word, and most of them play out like a game of telephone.

// What was said on the radio and what users read

On the radio

“Show me out at…”, the phrase an officer uses to tell dispatch they have arrived at an address. Mount Vernon, Missouri, April 2026.

“Fire alarm”: a kindergartner had pulled the handle by accident. Streetsboro, Ohio, December 2025.

“Shop with a Cop”, a holiday charity event where officers take kids shopping. Oregon, late 2025.

In the app

“Shooting at…”, followed by a report of a shooting at Mount Vernon Elementary.

“Firearm”, and a post reading “Firearms discharged at elementary school in Streetsboro.”

“Shot with a cop”, and the suggestion that an officer had been shot in Bend, although the event was in neighboring Redmond.

Four days after the Mount Vernon alert, the developer apologized: “This is a serious mistake and we have taken steps to stop it from happening again.” The company promised to update its audio processing and to tighten checks on any incident involving schools and firearms. Four months earlier, after Streetsboro, it had said nearly the same thing: the system “misunderstood ‘fire alarm’ as ‘firearm,’” and the audio protocols had been improved. Streetsboro Mayor Glenn Broska compared the post to yelling fire in a movie theater.

Why a machine hears gunfire that isn’t there

CrimeRadar doesn’t say how it works under the hood. But it has owned up to the first stage: after Streetsboro, it called the mistake “a serious transcription error,” meaning the step that turns sound into text. Police radio is hard on speech recognition even on a good day. The audio is compressed and hissy, officers talk fast, swallow words and use jargon nobody outside the job would recognize. Any dispatcher knows what “show me out at” means. In everyday text, though, the phrase is probably rare, while “shooting” turns up in the news all the time. The recognizer picks the continuation it finds most likely, and in these cases that turned out to be the scariest one.

If the app works the way such services usually do, a second stage then kicks in: a language model takes a scrap of transcript and writes a tidy post around it, with a headline, a category and details. Nobody seems to check the posts against each other. At 12:29 p.m. on December 9, 2025, CrimeRadar published two posts about the same call in Streetsboro: one mentioned a fire alarm, the other a shooting at an elementary school. The false one came down only the next day, after users complained.

Cpl. Dylan Davis of the Grand Chute Police Department told WBAY that the app seems to latch onto keywords like shooting, drugs or medical even when the dispatcher said nothing of the kind. Police don’t know exactly how an abandoned-vehicle call turned into gunfire.

Researchers showed back in 2024 that speech recognition can make things up. A team led by Allison Koenecke of Cornell ran more than 13,000 audio clips through OpenAI’s Whisper. About 1% of the transcripts contained whole phrases or sentences that were never spoken. 38% of those inventions contained something harmful: violence, false associations between people, made-up appeals to authority. The model made things up more often on recordings of people with aphasia, a language disorder that leaves long pauses in speech. Police radio is full of pauses.

You can’t simply carry the study’s numbers over to CrimeRadar: nobody knows which model it runs. The mechanism is probably the same, though. Silence and static don’t stay empty; the model fills them.

Speech recognition fills the gaps in radio audio with words nobody said
Speech recognition fills the gaps in radio audio with words nobody said

Two hours for a screenshot to cross town

A false post doesn’t live long, but it doesn’t need to. On November 4, shortly before 1 p.m., a police dispatcher in Grand Chute, Wisconsin, sent an officer to the Fleet Farm parking lot, where someone had left a car. CrimeRadar reported a possible shooting in that same lot. Police heard about it through social media and phone calls, including one from the store itself. Spotting the post and issuing a correction took about two hours. By then, screenshots were already making the rounds.

In the spring, the company promised this wouldn’t happen again. Since then, Cowboy State Daily has counted several more cases. In August, in the small town of Sumrall, Mississippi, the app reported a shooting with multiple victims. Residents called the police, a school asked whether it should lock its doors, and there was nobody to lock out: police found no victims, no suspects and no weapons. A month earlier, in Waco, Texas, several apps including CrimeRadar reported that an officer had been shot during what was actually a training exercise. On October 29, in Albemarle County, Virginia, CrimeRadar said an 11-year-old had overdosed at school and stopped breathing. There was no overdose, and the student never lost consciousness. Six days later came Grand Chute and Gillette.

After the “Shop with a Cop” alert, Bend police communications manager Sheila Miller told Central Oregon Daily News that such posts frighten the whole community, police families most of all. And there is no accountability for them, she said. Davis in Grand Chute worries about something else. “We don’t want this AI app causing people to become immune to these reports,” he said. Police need people to believe them when something real happens.

When false alarms keep coming, the real one gets lost
When false alarms keep coming, the real one gets lost

The fine print under a frightening headline

Technically, CrimeRadar doesn’t deceive anyone. Its store listing says the content “is not an official report or verified information” and tells users to look to official sources for confirmation. The Gillette “fatality” entry carried a note that the text was AI-generated and should be verified. It also showed that a user had already edited it, though there was no way to tell what had changed: the headline or the category. By the time the story ran, the company had not answered Cowboy State Daily’s questions about how it builds the summaries or what safeguards it has. On the App Store, the developer is listed as Scoopz Inc.

The most awkward part: nobody knows how accurate the app is. CrimeRadar doesn’t publish how many of its posts turn out to be false, and police only notice the loud mistakes. Nobody counts the correct alerts about crashes and fires, or the quiet errors that never prompted a phone call.

Police keep giving the same advice: listen to the recording before you share. Every CrimeRadar post comes with the original radio audio, one tap away.

Former Campbell County Sheriff Byron Oedekoven summed up the Gillette case in a sentence that applies far beyond Wyoming: “AI will give you something, even though it would have been better in this particular case to have given nothing.”