AI App Invents a Murder — It Was Really Just a Traffic Stop

What happened

On Wednesday, November 4, 2026, the CrimeRadar app posted an alert that could hardly have sounded more alarming: a man had been killed behind an apartment complex in Gillette, Wyoming. The entry came with a location, a timestamp (3:18 p.m. local time), and a category assigned by the software — "Fatality Report." According to Cowboy State Daily, the text read:

"As discussed in the dispatch call, radio traffic indicates a possible homicide of a man behind an apartment complex, with multiple units en route to the scene."

A day later, Lt. Dan Maul of the Campbell County Sheriff's Office looked up what had actually taken place. His verdict: "No, not even close." There was no body, no homicide investigation, no crime scene. Instead: "Yes, we had a traffic stop there at 3:18. It looks like we were dealing with a woman at that location." The woman had an outstanding warrant and was arrested. Maul said that looking at the entry, he couldn't see anything that would have turned the incident into a murder.

What CrimeRadar does

CrimeRadar is a third-party app that lets you listen in on public police radio and read automatically generated summaries of calls. This is no niche product: the app has an estimated 4.6 million Android downloads and ranked among the most popular free news apps in the US in early November. Clearly there is strong demand for instant information about incidents, accidents and emergency calls in your own neighborhood — even if the technology parsing the radio traffic can be wildly wrong.

Below the Gillette alert sat a note saying the text was AI-generated, that it was not an official police report, and that readers should verify the information with official sources. On its website, the service says it does not independently verify live radio information. The entry also showed that a user had edited it — though it's unclear whether the headline, the description or the classification was changed after the fact. Cowboy State Daily asked the company how the summaries are produced, what safeguards exist and how user corrections are handled. No reply had arrived by publication. Under the entry itself, at least one person had already commented "WOW."

Why police radio is especially hard for software

Maul draws a comparison many people will recognize: the auto-generated captions on YouTube videos that turn ordinary sentences into nonsense. Police radio stacks several additional problems on top — poor audio quality, static, overlapping transmissions, numeric codes, abbreviations and unfamiliar names. "It's not the best audio quality and a lot of codes and whatnot. So it's probably even worse," he says.

Byron Oedekoven, a former Campbell County sheriff and longtime head of the Wyoming Association of Sheriffs and Chiefs of Police, knows the problem from the pre-AI era: even people with scanners frequently misread the clipped jargon officers use to communicate quickly without tying up the channel. That a computer would now conjure a murder out of radio traffic still impressed him: "Oh my gosh! If an AI error can say 'homicide' — that's number one on the list of serious crimes. That's a little different than a call about a cat up a tree."

Even a perfect transcription wouldn't be the whole truth, Maul stresses. Over the radio, officers often relay what frightened, confused or simply mistaken witnesses have told them; the situation only becomes clear on scene. "Just because we think we hear something over the radio doesn't mean that's exactly what's happening." In the Gillette case, though, the problem ran one level deeper: the alert described a killing that never occurred. At least there was no panic — Maul said he knew of no calls about the supposed murder apart from the newspaper's inquiry.

Not an isolated case — and what the research says

Similar incidents have been documented repeatedly:

In the Virginia case, EMS responded and worried parents contacted the school district. Jason Grant, the district's communications chief, sums it up in one line: "These apps are fast, but fast is not the same as accurate."

The phenomenon has a name: hallucination. It means an AI system doesn't just make mistakes but confidently invents things that appear nowhere in the input data. Cornell researcher Allison Koenecke and colleagues examined more than 13,000 audio clips run through OpenAI's transcription system Whisper for their 2024 study "Careless Whisper: Speech-to-Text Hallucination Harms." About 1 percent of the transcripts contained entirely fabricated sentences or sentence fragments. Of the hallucinations found, 38 percent included potentially harmful material — violent language, false attributions or faked authority. In one particularly unsettling example, the software invented references to terror, a knife and a person killed, none of which were anywhere in the recording. "With hallucinations, artificial intelligence makes something out of nothing," Koenecke explains.

A tool, not a decision-maker

Oedekoven is no AI opponent. He sees potential, especially as agencies look for new ways to communicate and process information — but he draws a clear line between gathering information and deciding what is true. "I think it's important to understand AI as a tool. It's not a decision tool, it's an information tool that you then have to use to make an informed decision."

The biggest weakness, he says, is the system's urge to produce something, anything: "AI gives you something when, in this case, it would have been better to give you nothing." Maul's advice is simply to be cautious about AI-generated content that lacks corroboration. And Oedekoven expects that caution to become an everyday skill: "We're going to have to verify more, be aware of malfunctions and possible errors, and make sure we check and double-check."

Our take

The case is small but instructive, because it shows where automated summaries get dangerous: not in chatbot answers, which you read skeptically anyway, but where a machine shows up in a news-like format with a timestamp, an address and an official-sounding category. That packaging is exactly what lends credibility the content doesn't have. A school weighing a lockdown and paramedics racing to a nonexistent overdose are real consequences.

Technically, the chain of causes is plausible: hard-to-parse audio, speech recognition with a known tendency to hallucinate, and on top of that a text generator writing summaries — every step can add errors, and what comes out the other end is a sentence with no connection to the original transmission. A disclaimer in the fine print simply shifts the problem onto users. What remains unclear is what the company is actually doing about it: CrimeRadar didn't respond to questions about its processes and safeguards, and how the Gillette entry was edited by a user is anyone's guess. The practical lesson for readers: neighborhood and police-scanner apps deserve the same treatment as any AI text — for serious claims, check the official source first, then get worked up.