Telnyx maps EU AI Act Article 50 onto voice AI

Telnyx has a compliance guide for EU AI Act Article 50, which becomes enforceable on 2 August 2026 and requires that AI-generated voice be watermarked and that deployers be able to tell authentic audio from a deepfake. The sharpest point: the obligation “falls on the deployer, not the AI provider” — buying from a compliant vendor doesn’t transfer it — with penalties up to €15 million or 3% of global turnover.

It’s useful as a map of a fast-approaching deadline that sits squarely in the deepfake-disclosure lane. The post flags that May 2026’s Digital Omnibus pushed the provider-side watermarking deadline to 2 December 2026 while leaving the deployer deepfake-disclosure obligation at 2 August, and notes the UK’s parallel Ofcom Online Safety Act duties (up to £18 million). The rest is a Telnyx argument that a “stitched stack” of separate speech-to-text, LLM, text-to-speech, and telephony vendors breaks Article 50’s chain-of-custody requirement, because an application-layer watermark gets applied only after the audio has already passed through several hands.

The regulatory summary is genuinely useful; the architecture argument is a pitch for owning the stack at the carrier layer. Both can be true. The part worth internalizing is the deployer-not-vendor liability split — it’s the same pattern as controller obligations under GDPR, and it means “we use a compliant vendor” won’t be a defense.