Securing Phone Calls in the Age of AI Voice Cloning

With the rise of AI-generated voices, it's getting harder to trust who you're actually speaking to over the phone. Verifying whether a number really belongs to the service you're dealing with is key to avoiding scams. To cut down on social engineering—especially for older folks—we might need to move away from direct phone calls and instead use an AI middleman to relay conversations. This could make communication safer, more reliable, and much harder to manipulate.

A solid way to verify if someone has access to something is through the internet, and the safest approach is likely a website-based method. We could add a text field to our domains—something like a public key—that acts as proof of identity. The real challenge is authentication over a voice call, where one option could be combining both phone numbers, adding a salt, and encoding it into binary sounds for transmission. To securely send data, signing it with a private key and verifying it using the public key on a domain ensures the other side can confidently say, "Yes, this is a bank," or "Yes, this is the right person."

Just like we can verify that an email belongs to a specific domain, we can do the same for phone numbers, linking them to domains for authentication. There's already plenty of precedent for this, and as scams keep rising, people will likely shift away from pure phone calls toward platforms like WhatsApp or Telegram, where verification is built in. These apps could introduce a verified tick to confirm that a number is legitimately tied to a domain—done automatically. The only risk is users misreading domain names, much like mistaking a phishing email, but at the end of the day, some of the responsibility still falls on the user.

The main risk here is user error, but if a person had their own AI chatbot handling calls, it would be far less likely to misread or overlook small details. If someone claims, "Hey, I'm from [Company]," the AI could vet that claim instantly. Sure, hallucinations are still a factor, but as the tech improves and LLMs become more reliable, this system only gets stronger. The real key isn't just LLMs—it's LLMs combined with deterministic tools like function calling or structured outputs, allowing the AI to monitor, relay, and filter conversations in real-time, even shutting things down if they start looking sketchy.

This kind of system wouldn't just improve security—it would make communication more accessible, not less, especially for older people who are most vulnerable to scams. In Australia, scams are a massive problem, and many people lose money because of them. If we had widespread adoption of verified call systems, it would completely change expectations, making normal phone calls feel outdated and even suspicious. Just like how blue ticks on X used to indicate verification, we need a clear, trusted way to tell if a caller is legit—or if it's just another scam. And with transcription tools improving fast, we're getting closer to making this a reality.