Telnyx argues voice-AI trust belongs at the network layer

In a companion piece, Telnyx makes the case that voice-AI trust has to live at the network layer. The hook is a Mobile World Congress 2026 demo by Resemble AI: 140 attendees tried to tell AI-generated audio from real recordings and averaged roughly 50/50 — a coin flip — with a top score of 8 out of 10.

The argument follows from that result. If trained listeners standing in a booth can’t reliably spot a deepfake, voice-AI security has to start from the assumption that inbound audio might be synthetic. Telnyx and Resemble announced a carrier-layer deepfake-detection feature that fires a webhook on every call, flagging AI-generated voice within four seconds — early enough for an application to drop the call, escalate to a human, or trigger step-up authentication. They claim the detection model was trained across more than 150 text-to-speech architectures and 52 languages, and argue that both detection and watermarking belong at the carrier interconnect, before the codec transformations that degrade signals applied at the application layer.

It’s vendor framing, but the underlying instinct lands: “trust lives at the network layer” is the same logic behind STIR/SHAKEN — put the verification where the traffic actually flows, not bolted on at the edge. The open question is whether carrier-layer deepfake detection becomes a real product category or stays a single-vendor pitch.