Post-trust: the simple phone call, undone

Or from something that sounded like me, asking for a quick transfer. Neither call came from my number.

Both of them clocked it within seconds and nothing happened. The part that stayed with me wasn't the attempt, it was how little effort had gone into it. Someone had cloned my voice, worked out who around me might act on that voice, and then not bothered to spoof my number. They didn't need to. The voice on its own was worth a try.

I assume this is happening constantly now. But it left me with a question I hadn't thought about properly in years: what actually makes a phone call trustworthy?

We got "post-truth" in 2016. It fixed nothing, but it gave us a way to describe something that had already happened to political argument, which was that shared facts had stopped working as the anchor. Something similar is happening to the phone call. Post-trust, if you want a word for it. For a hundred and fifty years the phone call has underwritten banking, medicine, law, insurance and negotiation, and it did that on the strength of an assumption that is coming apart.

Six struts

Trust in a phone call was never held up by one thing. It rested on six struts:

• Your voice, which was expensive to fake.
• The number, which was tied to a person.
• Caller ID, which the carriers were supposed to vouch for.
• The network, which only a handful of operators could originate a call on.
• The device, which was a phone and nothing else.
• The shared meaning of "a phone call", which was the same thing at both ends of the line.

Every one of those is weak on its own. Together they were strong, because to fake a call convincingly you had to beat all six at the same time, and until roughly 2023 nobody could do that cheaply. What we were calling trust was really the cost of beating six things at once.

Nearly all six are now loose.

How each one gave way

Voice

Voice went first. Faking one meant a studio, weeks of source audio and someone with real skill. Now it's thirty seconds off any podcast and a laptop. Whatever your voice was worth as proof of who you are, it's worth a fraction of that today.

The number

The number went next. SIM swaps broke the link between a number and a person. Cheap VoIP broke the link between a number and a place. Spoofing is a service you can rent by the month. A number that calls you is a claim now, not evidence.

Caller ID

Caller ID was supposed to be rescued by STIR/SHAKEN, and it sort of has been. In practice a full attestation tells you the originating carrier believes the caller is entitled to use that number. That is not the same as the call being real, and it says nothing at all about who, or what, is doing the talking. Adoption is patchy, fraudsters route through the gaps, and international interconnect is a mess. Still standing, but bending.

The network

The network stopped mattering a long time ago. Once anyone with a laptop and an API key can originate a call from anywhere, the network isn't a filter, it's a pipe. We just carried on behaving as though it was doing more than that.

The device

The device is the odd one. The thing in your pocket isn't a phone. It's a general purpose computer running seventy apps, any of which can listen, and the phone is one of them. That's been true since about 2010, and we still call it a phone and mostly act like that means something.

The shared meaning of "a phone call"

The sixth is the one nobody really wants to say out loud, which is that "a phone call" used to mean the same thing at both ends of the line. When you talk to a customer service agent that turns out to be a model, is that a phone call? When two agents negotiate a freight load between them? When a clone of your CFO's voice tells you to move money and you do it, what was that, exactly? Read it strictly and the shared category has quietly gone.

So: six struts, most of them loose, one or two gone. None of them was holding your trust up on its own. The scaffolding was, and the scaffolding wobbles.

Two curves crossing

What makes this worse than it might sound is the timing. The cost of impersonating someone by voice has collapsed in the same few years that legitimate autonomous callers, AI voice agents deployed by real institutions, have started showing up in volume. Two curves that used to be nowhere near each other are crossing. Fraud has never been cheaper to attempt, and there have never been more legitimate machine callers to hide among. Trust has to come from somewhere other than "the voice sounds right."

Why this can't just be lived with

Worth saying, because a lot of what voice AI is doing is good. Small farmers who were never going to fill in a grant application on a laptop can now do it by voice, in their own language, at the pace of a conversation. There are more people ageing than there are carers to look after them, and a voice agent can be a medication reminder, a bit of company and a bridge back to family at a scale no human care system is going to reach. Voice gets to people that literacy-first interfaces never did. I don't want to lose that.

Which is why the collapse isn't something we can just live with. If voice is going to be the layer through which vulnerable people get at grants, care and expertise, that layer has to be trustworthy. Otherwise every good use case ships with a fraud stapled to it.

Fraud follows the surface

Fraud follows the surface, by the way. That isn't a moral observation, it's just what the record shows.

• Credit cards were rare, so card fraud was rare.
• Online commerce arrived and payment fraud arrived with it.
• Mobile banking spread and SIM swap fraud spread with it.

Voice AI is the new surface and voice fraud is right behind it. The mirror doesn't lag. It turns up at the same time.

Patching the exceptions when the exceptions are the norm

Detection on its own doesn't solve this, and the detection companies will tell you so themselves.

Detection tells you after a call that it was probably fake, and "probably" is carrying a lot of weight in that sentence. Disclosure rules, the EU AI Act and the state laws that came in through 2025 and 2026, assume the caller volunteers that it's AI. Callbacks assume there's a safer channel to call back on.

All three start from the idea that a phone call is basically trustworthy and try to patch the exceptions. The exceptions are now the ordinary case.

Email already did this

We have done a version of this before. Email went through the same collapse in the early 2000s, when spam and phishing broke every assumption about who a message came from. It took about fifteen years to fix, and what fixed it wasn't better detection, although that came too. It was new infrastructure built into the substrate: signed provenance, verifiable senders, messages that fail safely when the signal is missing and can be trusted when it's there. Nobody thinks about SPF or DKIM today, which is exactly how you know they worked.

What voice needs

Voice is at the front end of the same fifteen years. What it needs is its own trust infrastructure.

Identity that can be verified independently of the number. Provenance that doesn't rest on caller ID or on someone recognising a voice. Some way for legitimate humans and agents to establish who is actually behind an interaction.

I should say that I run a voice infrastructure company, so discount me accordingly, but I think the shape of that is reasonably clear by now. The detail isn't, yet.

Who is at either end

The phone network worked out how to connect two endpoints a very long time ago. What it can no longer reliably tell you is who, or what, is sitting at either end of it.

Voice AI didn't create that weakness. It has made it impossible to keep ignoring, and it means the next layer of voice infrastructure won't be about agents that sound better, respond faster or cost less. It'll be about trust.