Gartner® names Reality Defender a Market Shaper in deepfake detection, June 2026

Get the report

\

Insight

\

Who’s calling 911 now?

Alex Lisle

CTO

Scientists and engineers are terrible at understanding the impact of technology. If you want to understand what a new technology will do to the world, don’t ask us. Our track record reads a litany of underestimations and overestimations.

New technology is presented to the world, and the world adapts to it in ways beyond most engineer’s ken. The internet, mobile phones, social media, all these things shift our world in ways unforeseen.

The internet has become a world of the synthetic. There are theories and strong indicators that most of what is posted and shared on the internet is generative. Some of it obvious, some of it subtle, most of it falling somewhere in between.

How bad is it? Well, model companies increasingly leverage telltale markers in their generative output so they don’t accidentally use their own output to train their own models. Amazon is buying and allegedly burning books to train. These are things you don’t do if the internet is a ripe place of untapped human output. The internet is dead, long live the internet. Truthfully the internet isn’t dead, it just wasn’t what it was. If Heraclitus was alive, I am sure he would have something to say about it.

Generative media is upon us. And like other technologies unleashed onto the world, how it is being used, and why it is being used is evolving and changing, both the technology and more importantly us.

We don’t trust what we see on the internet, but what we hear, that is different. And more importantly, unlike the internet where the idea of bots has been around for a very long time, calls were inherently human. One side of the communication was always assumed to be human.

And now it isn’t. The worst thing to go wrong is when something which is ubiquitous, obvious, stops being true. When something we rely on no longer holds true. It’s in that where monsters be.

Ben wrote last week about what this looks like when it lands close to home. I want to go one layer down: how these calls are actually built, who they’re aimed at, and why the usual ways of catching them don’t really work any more.

Who’s on the line

Swatting never made a huge amount of sense to me, but then again, I suppose it doesn’t have to make sense for it to happen. Generally though, it had an inherent aspect to it that was personal. Someone wanted to inflict pain and suffering to a person. Like a lot of these types of attacks, it used a third party, removing a lot of the risk associated with it, while offering a range of outcomes from mild inconvenience to loss of life.

It was one person, attacking another person, leveraging a third party. The FBI says that has shifted. The targets now are hospitals, government buildings, houses of worship, transit centres, and schools, and the calls often arrive in coordinated waves across several places at once.

It points to something else: it’s no longer against the singular person, but the institutions. It requires a larger response, more effort, the outcomes are less severe in some ways, and more in others. It’s to waste the resources needed to face a threat against such a larger potential victim, either for fun or for profit.

How the calls are built

In security we talk a lot about trust boundaries. There are explicit trust boundaries, these are good, and there are implicit trust boundaries, these are less good. Like anything which is implicit, it isn’t defined.

911 has an implicit trust boundary. You trust the person who is calling, you simply have to. Why? Because there is asymmetric risk. If you are right, and ignore it, you save resources. If you are wrong, something very bad and very preventable happens.

The first generation of swatting using generative voices was to add an extra step from the attacker so they’re less likely to be caught or create incriminating evidence of their own voice.

There are now other reasons to do it.

The voice. The older swatting voices were fairly easy to spot. The new ones aren’t. Callers adjust pitch and pauses and lay AI-generated background sound underneath. As Mel Maier, CEO of APCO International, told NPR: “They can use words like, ‘uh’, or ‘maybe, um, let me think about it.’” And one forensic psychologist says he can clone his own voice for less than $20.

The number. The FBI’s warning signs are numbers that show up as all zeros, all nines, or blocked, sometimes from outside the US. Those are, I suppose, the careless ones. The harder case is a number that looks completely ordinary. Callers spoof caller ID so the screen shows someone else’s perfectly real number, and internet phone services let them spin up a fresh number and throw it away in seconds.

The route. Most swatters don’t dial 911, because 911 can pinpoint where a caller is. They ring a police non-emergency line or a courthouse, and the call gets transferred into 911, where it’s very hard to trace. Nobody designed it that way. It just sort of accumulated. It reminds me of phone phreaking, when the right tones got the network to connect you for free. Now it isn’t a tone being reproduced. It’s a person.

The script. The scenarios are written in advance to force an immediate response: a shooting in progress, a bomb in the building. With the right tools, one caller can hit several neighbouring counties with the same number. Why people do it is rather harder to say. Sometimes it’s a gaming dispute or a vendetta, sometimes panic for its own sake, sometimes politics. From the receiving end, though, it all sounds exactly the same.

The automation. This is the bit I find most interesting from my point of view. You no longer need a human on the line at all. An AI voice agent can run the whole call on its own, following a branching script and answering whatever questions come back, with a person able to take over if the call goes somewhere unexpected. Building something like that used to be non-trivial. It isn’t any more.

The cost to emergency response

I’m generally wary of arguing from numbers, but this one is worth knowing. In a June survey by APCO and the Public Safety Threat Alliance, about 14% of 911 dispatchers said they’d been targeted with AI tools. I suspect the real figure is higher, for the simple reason that you can’t report what you didn’t detect, and most centres don’t have a detection protocol in place.

The more useful way to think about it is in time. Every fake call pulls officers, dispatchers, and responders away from something real.

In August, a string of AI-generated robocalls threatened schools across two north Georgia counties. Deputies and bomb dogs searched six campuses in Jackson County, which cancelled classes across the whole district, and a high school in neighbouring Hall County was evacuated. Nobody found anything and nobody was hurt, which is the good outcome. It still cost one district a full day of school, and investigators believe the calls were linked.

Why ears and laws don’t catch it

The FBI’s guidance for spotting a swatting call is sensible enough: a caller who sounds calm when they shouldn’t, who sounds like they’re reading from a script, or who is vague on the details. The trouble is that each of those is something AI is quite good at removing. The calm becomes panic, the script becomes a conversation, and the vague details get filled in. We’re asking dispatchers to win a battle of wits against something that doesn’t get tired and doesn’t get nervous.

Penalties are the other answer people reach for, and I’m fine with tougher penalties. My only thing really is that a penalty needs a person to attach to, and a caller on an internet number, routed through a courthouse line, possibly from another country, is very hard to name.

What changes

I don’t think the answer is better-trained ears, I’m afraid. Human perception doesn’t scale, and it’s the very thing being attacked. The way we view it is that detection has to sit inside the call itself rather than with the person listening to it. The audio is analysed while the call is happening, and the result arrives alongside it.

Just as important is where that signal goes. It goes to whoever is handling the call, but also to the systems around them: routing, dispatch, and escalation. A call flagged as likely synthetic might get a second check before units go out, or go straight to a supervisor. The people and the systems keep the decisions. Detection gives them something to decide with, which at the moment they don’t have.

The person on the other end

I said at the start that I found the dead internet theory a bit overstated. I don’t think the same is true of the phone.

Technology can change the world in a day. Realistically, emergency services don’t change in a day, and nobody should expect them to rebuild themselves around this. But the assumption at the centre of every 911 call, that there’s a person on the other end, is no longer one anyone can safely make. That’s the real problem. It isn’t that every call is fake. It’s that you can no longer be sure which ones to trust.

If you run a 911 centre, or any of the places on the FBI’s list, it’s worth looking at where your own trust boundaries sit. If it would help, we’re happy to do a risk assessment with you, or you can simply contact us.

Frequently Asked Questions

Frequently asked questions

Swatting is a false emergency report made to send police or other responders to a location. AI swatting uses generated or cloned voices to make those calls, and increasingly AI voice agents that can run the whole call with no person on the line.

According to the FBI, targets have shifted from individuals to institutions: hospitals, government buildings, houses of worship, transit centres and schools. The calls often arrive in coordinated waves across several places at once.

Not reliably. The FBI’s warning signs, such as a caller who sounds too calm, scripted or vague, are exactly what AI is good at removing. Current voices add filler words, natural pauses, and background noise.

Callers spoof caller ID to show someone else’s real number, use internet phone services to create and discard numbers in seconds, and route calls through non-emergency or courthouse lines that transfer into 911, where they are hard to trace.

Detection needs to sit inside the call, analysing the audio as it happens and passing the result to the call-taker and to routing, dispatch and escalation systems. A call flagged as likely synthetic can get a second check before units go out, or go straight to a supervisor. Real people still make the decision.