AI-powered apps alert Boulder County area residents to fires, shootings that don’t exist

On Jan. 30, some Frederick area residents acquired a notification that firefighters have been battling a “commercial blaze” downtown.

A screenshot of the alert was shared to Facebooksparking concern that was shortly dismissed by one commenter who wrote: “I’m sitting in an office in that building and there is nothing going on.”

An emergency notification app that makes use of synthetic intelligence had misinterpreted radio visitors from a coaching train that simulated a construction hearth in downtown, in accordance to a post from the Frederick-Firestone Fire Protection District.

“This incident is a good reminder of the importance of verifying information through multiple reliable sources before sharing or acting on it,” the submit reads.

Summer Campos, a spokesperson for the district, mentioned she wasn’t certain how the app had entry to the channel firefighters have been utilizing. In the long run, the district can be utilizing a tactical channel that does not air publicly, she mentioned.

Campos could not affirm what app had alerted residents to the false scenario. But CrimeRadar, an app that makes use of AI to summarize publicly accessible dispatch audio, had a submit that described a fireplace in downtown Frederick.

Such false alerts usually are not distinctive to Frederick. In Boulder and Longmont, AI-driven emergency notifications have unfold false data that, in some situations, has sparked very actual concern.

Incorrect alerts

On Wednesday, the CrimeRadar despatched out an alert reporting an condominium hearth in Longmont. Linked dispatch audio did not embrace a location, and Rogelio Mares, a spokesperson for the town, mentioned the town hadn’t acquired reviews of any condominium fires. The submit later disappeared from CrimeRadar.

CrimeRadar displays publicly accessible dispatch audio after which makes use of AI to generate summaries, in accordance to its web site. Users entry alerts on the CrimeRadar web site or by downloading the CrimeRadar app, which additionally sends push alerts.

While Longmont police radio transmissions are encrypted, Longmont Fire radio visitors is aired, Mares mentioned.

In Boulder, the app not too long ago reported that a firefighter was taken to the hospital after a medical emergency in Boulder. The abstract mentioned “the firefighter’s condition was not disclosed.”

No Boulder firefighters have been injured that day, in accordance to Jamie Barker, a Boulder Fire-Rescue spokesperson. The app appeared by mistake “fire rider” — a time period generally utilized by dispatchers to describe firefighters driving in ambulances to present care to sufferers — for “firefighter.”

“It took information that it heard incorrectly, and then it summarized it incorrectly, and then also made an assumption,” Barker mentioned.

On Wednesday, the app posted about 60 reviews of incidents within the Longmont area, together with summaries that precisely summarized linked dispatch audio describing medical emergencies and a reported assault. When a fireplace sparked within the Colo. 119 median east of Longmont on Wednesday, a CrimeRadar submit precisely described the scenario.

When apps interpret dispatch audio appropriately, there may be nonetheless an opportunity the data might be inaccurate or incomplete, as dispatch communications might be incomplete and unverified if crews usually are not but on scene, Barker mentioned.

“The scanner can be an exceptional tool and resource, but the scanner also only ever tells half of the story,” Barker mentioned.

Such false alerts might be dangerous, particularly if the corporate accountable for the system places it forth as correct, in accordance to Casey Fiesler, a professor of data science on the University of Colorado Boulder who researches AI ethics.

“If someone gets an alert saying that there’s a fire, that’s going to be very upsetting,” she mentioned.

“People often think that machines are less biased or more accurate than humans, so for this reason, I just think it’s really, really important that systems like this have very strong disclaimers about how information might not be accurate,” Fiesler added.

CrimeRadar mentioned in an announcement it’s “constantly improving” its system to “ensure higher precision.”

When a person clicks right into a CrimeRadar alert to see extra particulars, a disclaimer is posted above the AI-generated abstract that reads, “Not official report. AI-generated from public dispatch audio. Verify with official sources.”

“Our goal is to make communities safer by making emergency information accessible, which is why our disclaimer has been a core feature since day one,” the CrimeRadar crew wrote within the assertion. “It serves as a constant reminder to users that dispatch calls are unconfirmed and to always rely on official sources for final confirmation.”

Misinformation spreads shortly

Nextdoor, which makes use of the AI-powered Samdesk to generate alerts, has additionally despatched out incorrect data to Boulder County communities.

Last failed, a Nextdoor alert reporting an energetic shooter at a federal facility in Boulder shortly generated concern. A screenshot was shared to Reddit, Boulder dispatch started getting a number of calls from involved group members, and police have been scrambling to get correct data, Dionne Waugh, a Boulder police spokesperson, mentioned.

The Nextdoor alert turned out to be false. The AI ​​had scraped data from an outdated on-line submit, Waugh mentioned.

The panic generated by these false alerts might be notably dangerous in Boulder, which has seen large-scale tragedies lately, together with the 2021 King Soopers shooting and, simply this summer season, a Targeted attack on Pearl Street.

“Every time something like that happens, it takes people back to the real moments,” Waugh mentioned.

“We understand how distressing a false alert can be for residents, and we regret the concern these incidents caused,” a Nextdoor spokesperson wrote in an announcement.

The firm labored straight with each Boulder and Longmont police after the departments expressed issues in late 2025.

Nextdoor alerts are processed by means of a assessment system and “augmented by human oversight as needed,” in accordance to the assertion.

The app has not too long ago added extra verification layers for alerts that contain incidents like mass shootings and wildfires, and given native public companies entry to an alerts map so “boots on the ground” specialists can flag or right inaccuracies inside their jurisdictions.