THE PROVENANCE WALL .. On the Deepfake Crisis, the Symptom-Level Remedies That Cannot Certify, and the Architectural Answer: Verify the Artifact, Not the Human

In October 2026, Sam Altman told a Federal Reserve conference that voice-printauthentication is still in use at some financial institutions, and that AI can now replicate aperson's voice in seconds. https://x.com/i/status/2115193109864989034 He called the situation a "significant impending fraud crisis." The warning is correct. The remedies being deployed are not. This paper makes three claims: First — the crisis is a provenance failure, not a detection failure. A voice call carries nosignature, no trace, and no named source. The question "is this really the person they claimto be?" has no checkable answer. It can only be estimated. And estimation, on anunbounded input space, cannot be certified. Second — the deployed remedies reproduce the disease. Liveness detection, adversarialrobustness testing, and behavioral biometrics are all probabilistic classifiers answering anunbounded question. They are stochastic systems defending against stochastic systems.They move second by construction, because every improvement in detection is a trainingsignal for the next generation of forgery. Third — the architectural answer is to stop authenticating the human and start authenticating the artifact. A signed message from a known key cannot be deepfaked — not because detection improved, but because the forgery fails at the signature layer. The question changes from "does this look real?" to "does this verify against a named key?" The paper states what this solves, what it does not solve, and how it could be wrong. It doesnot claim that an architecture ends deepfakes. It claims that the class of interaction whichcan be certified is the class that carries provenance — and that everything else should be treated as unverified by default.