An Anne Arundel County Fire Department dispatcher responds to an emergency 911 call from his department’s dispatch center in Glen Burnie, Maryland, in 2020. 911 dispatchers have long handled crash calls, where callers report a bomb threat, mass shooting or other type of life-threatening emergency with the intent of initiating a massive emergency response. Now, AI-generated voice calls pose a new challenge.
Alex Edelman/AFP via Getty Images
hide caption
Alex Edelman/AFP via Getty Images
In August, a caller made a bomb threat to Wheatland Union High School in Yuba County, California.
The police intervened during the night and classes were canceled. There was no bomb and no one was hurt.
Authorities said it was part of a series of scam calls — also called swatting — made to schools across the country over the summer.
Calls may be made to 911 or police departments, with callers reporting a bomb threat, mass shooting or other type of life-threatening emergency with the intent of initiating a massive emergency response, according to the FBI.
911 dispatchers have long faced this problem. But the incident in Yuba County was different: The call came from an AI-generated voice, a new and growing challenge for emergency responders.
“These voices (of knocking callers) were pretty easy to recognize, but now they sound natural. They can insert pauses. They can use words like ‘uh’ or ‘maybe, uh, let me think about that.’ And it feels very natural,” says Mel Maier, CEO and executive director of APCO International, the Association of Public Safety Communications Officers.
The purpose of canceling calls is to harass or retaliate against the targets of the calls, such as a public official. They typically trigger a massive, chaotic police response, redirecting public safety resources from real emergencies.
People have also been killed and injured in crush attacks, Maier notes.
Removing calls made by people using AI-generated voices or background sounds makes these scam calls even more realistic and makes it harder to determine what is real and what is not.
APCO and NENA: The 9-1-1 association, which both represent 911 dispatchers, say incidents like the one in Yuba County reflect their growing concerns about how AI could be used by bad actors.
APCO and the Public Safety Threat Alliance, an information sharing and cybersecurity group, studied the prevalence of attacks on 911 call centers. They released their findings in June, based on 766 responses from public safety communications professionals. About 14% of dispatchers reported being targeted using AI tools. Maier says he thinks that number will continue to rise with more and more advanced tools
Signs of an overwhelming appeal
Companies like ElevenLabs, an artificial intelligence company creating synthetic voice software, offer AI tools that are incredibly cheap and readily available, according to Brandon May, an assistant professor of forensic psychology at the Florida Institute of Technology.
“I can clone my voice on ElevenLabs for less than $20 and I can change the intonations. I can change the depth or the high pitch of my voice. I can change the pauses,” he says.
And that makes it difficult for anyone to detect these AI fakes.
“It’s no longer easy to just look at an image, hear a video, or hear a voice and say, ‘That’s AI.’ People are constantly fooled by this,” he says.
According to May, sophisticated fraudsters using AI chatbots are even tricking people into handing over their savings.
And the increasing sophistication of these fake voices and sound effects makes it especially difficult for emergency dispatchers to deduce what is real or not during what May calls “high-stakes, high-consequence” calls. “It’s going to be almost impossible to detect,” he says.
Despite the difficulties, APCO and NENA are developing ways to train their members to deal with these new threats.
April Heinze, NENA vice president and chief of 911 operations, says there are certain characteristics of an overwhelming call.
For example, most swatters do not call 911 directly.
“The reason is that 911 provides a location. It provides a call-back number. It provides the name of the person who owns the phone. It provides a lot of additional data and it’s really impossible to spoof your location to 911,” she says.
Swatting calls are usually made to police departments or courthouses, Heinze said. These callers are then ultimately transferred to 911 centers. It is therefore impossible to trace these calls, she said.
Training evolves to face the new threat of AI
Heinze and Maier decline to describe the specifics of the training aimed at identifying potential AI crash calls. The more information available, they say, the easier it will be to train AI to fool dispatchers in the future.
Maier notes a training idea. APCO is working to create caller interrogation training criteria, where the dispatcher asks the caller specific questions at “such a granular level that an AI can’t pull that information from a public database, can’t capture it from a social media post, or just make guesses,” he says.
Despite growing concerns about the use of AI, Maier and Heinze emphasize that dispatchers will always assume that calls reporting a life-threatening situation are real — and will continue to send emergency resources to any calls for help.
“It’s important to understand that no matter what happens, public safety will have a very strong response,” Heinze said. “Even though there are indicators that it’s a crash incident. They always signal a life-or-death situation, and if it’s a life-or-death situation, you can never take chances.”
Gn headline
