911 groups develop dispatcher training to counter AI-generated emergency hoaxes

A convincing voice is becoming less useful for judging an emergency call. APCO International and NENA, two organizations representing 911 dispatchers, are developing training to address AI-assisted hoaxes, NPR reported November 11. One approach would have dispatchers ask detailed questions that a fabricated caller could not easily answer using public information.

A synthetic voice, a real school closure

In August, an AI-generated voice delivered a bomb threat against Wheatland Union High School in Yuba County, California. Police responded overnight, and classes were canceled. There was no bomb, and no one was hurt. Authorities described the incident as part of a summer series of school hoaxes across the country.

These attacks are known as swatting: false reports of bomb threats, shootings or other life-threatening situations intended to trigger a major emergency response. They can divert police and other public-safety resources from real emergencies. Mel Maier, APCO’s CEO and executive director, noted that swatting attacks have also killed and injured people.

Maier said synthetic voices have become harder to recognize because they can reproduce ordinary hesitation. Pauses and fillers such as “uh” or “let me think about it” make the speech sound natural. AI-generated background sounds can add another layer of realism, making it harder for dispatchers to distinguish a hoax from an actual crisis.

An earlier survey offers a measure of reported exposure. APCO and the Public Safety Threat Alliance, an information-sharing and cybersecurity group, published findings in June based on responses from public-safety communications professionals. The result describes those respondents’ experiences, rather than the share of all emergency calls that involve AI.

About 14%Respondents reporting AI targeting

The June findings drew on 766 responses from public-safety communications professionals. About 14% said they had been targeted using AI tools.

Questions that go beyond public information

APCO’s proposed response includes detailed caller questioning. Maier said the organization is developing training criteria for questions whose answers would not be readily available in public databases or social-media posts, or easily guessed. The aim is to probe the caller’s knowledge of the reported situation, rather than rely only on whether the voice sounds human.

Maier and April Heinze, NENA’s vice president and chief of 911 operations, declined to disclose the specifics of training to identify potential AI swatting calls. Both said that making more information public could help attackers train AI to fool dispatchers. Their response leaves the detailed methods undisclosed while the organizations develop the training.

The call’s route can also offer clues. Heinze said swatting callers usually contact police departments or courthouses, which then transfer them to 911 centers. She contrasted that route with dialing 911 directly, which provides dispatchers with information such as a location, callback number and the name of the phone’s owner.

Suspicion will not cancel the response

The tools creating this challenge need not be expensive. Brandon May, an assistant professor of forensic psychology at the Florida Institute of Technology, told NPR he could clone his voice using ElevenLabs for less than $20. He said he could adjust its intonation, pitch and pauses—features that complicate judging authenticity by listening alone.

Even better questioning will operate within a firm safety boundary. Maier and Heinze said dispatchers will continue treating reports of life-threatening emergencies as real and sending resources, including when there are signs of swatting. Heinze’s explanation was direct: a possible life-or-death situation is not a chance public safety can take.