Comment on the Robocall Mitigation Database FNPRM: Content Provenance as a Complement to Call Authentication
Box Commons · 30 N Gould St Ste N, Sheridan WY 82801
- A call can carry a full A-level STIR/SHAKEN attestation and still deliver audio that is entirely AI-generated or cloned. The signaling plane is authenticated; the media plane is not.
- Three integration pathways extend the existing architecture: a PASSporT extension type, Rich Call Data integration (RFC 9795), and media-plane binding (RFC 8862), delivered out-of-band to avoid the TDM stripping problem.
- Independent certification turns narrative RMD self-certification into auditable evidence, and gives providers deploying AI call analytics the documented standard of care insurers now require.
+ Jump to Section
I. Executive Summary
The FNPRM appropriately strengthens the Robocall Mitigation Database's filing integrity and provider accountability mechanisms. However, the Commission's call authentication framework, including the RMD and STIR/SHAKEN, addresses only the signaling plane. It authenticates who is calling and whether the caller has the right to use a telephone number. It does not, and structurally cannot, authenticate the media plane: the actual audio content of the call.
A call that is fully authenticated under STIR/SHAKEN, properly signed with an A-level attestation, can still deliver audio that is entirely AI-generated, cloned, or manipulated. This media blindness is the architectural vulnerability that AI-enabled robocall fraud exploits.
Box Commons recommends that the Commission recognize independent, third-party content provenance certification as a complement to signaling-layer authentication within the RMD framework.
II. STIR/SHAKEN's Structural Blind Spot
The PASSporT specification (RFC 8225) defines claims for originator identity, destination identity, and issuance time. It contains no representation of the audio payload, no cryptographic binding to the media stream, and no mechanism to verify the nature or provenance of the voice content.
The Commission's February 2024 Declaratory Ruling (FCC 24-17) correctly classified AI-generated voices as "artificial or prerecorded" under the TCPA, and the pending NPRM (FCC 24-84) proposes disclosure and consent requirements for AI-generated calls. But neither the Commission nor voice service providers possess a standardized, verifiable mechanism to detect whether a given call uses AI-generated voice content. The disclosure requirement depends entirely on the good faith of the caller.
The RMD's narrative self-certification model compounds the gap: a mitigation plan filed at a single point in time cannot represent a provider's continuously evolving technical capabilities.
III. Content Provenance as a Complement to Call Authentication
We do not propose replacing STIR/SHAKEN or restructuring the RMD. We propose extending the existing architecture with C2PA-aligned content credentials: tamper-evident manifests that would assert whether voice content is human, AI-generated, or hybrid; whether AI voice detection was performed; whether AI disclosure was delivered; and a cryptographic binding to the audio stream itself.
Three integration pathways:
- PASSporT extension type carrying a content provenance claim, signed with the provider's existing STI certificate.
- Rich Call Data integration (RFC 9795), surfacing a "Verified Human Voice" or "AI-Generated, Consent Verified" indicator on the handset.
- Media-plane binding via RFC 8862, establishing a verifiable chain from caller identity to the voice payload.
Out-of-band SHAKEN (RFC 8816) is the natural delivery vehicle: manifests published alongside PASSporT tokens avoid both SIP size limits and the TDM stripping problem.
IV. The Insurance Forcing Function
Providers are required to deploy advanced analytics to identify and block illegal robocalls. Yet ISO exclusionary endorsements CG 40 47 and CG 40 48 broadly exclude AI-related liability from standard commercial general liability coverage. A false positive that blocks a legitimate appointment reminder or fraud alert creates exposure that the provider's carrier has likely excluded.
Specialty AI insurers require auditable governance and data provenance verification as conditions of coverage. Independent certification of content provenance practices establishes the defensible standard of care those markets demand.
V. Recommendations
- Recognize content provenance as a component of robocall mitigation. RMD filings should permit, and the Commission should encourage, descriptions of content provenance and AI detection capabilities.
- Initiate a technical proceeding on content provenance for telephony, coordinated with the IETF STIR Working Group, ATIS, and the SIP Forum.
- Condition any recognition on governance criteria: multi-stakeholder participation with balanced voting weight, royalty-free IP defaults, separation of standards development from assessment, and public certification registries.
- Coordinate with the AI voice rulemaking (FCC 24-84) so that machine-readable provenance can operationalize disclosure and consent requirements.
- Engage international standards alignment as EU AI Act Article 50 machine-readable marking requirements take effect.
Contact:
Brice Love, Acting Executive Director
Box Commons
[email protected]
Content Integrity Notice: This comment was authored by the Box Commons Policy Working Group. Generative AI was used for research synthesis and drafting support. All policy positions, recommendations, and normative claims were formulated and reviewed by human authors.
Related Analysis
Related Filings
Comment on the CDC Data Intermediaries RFI: Data Quality Certification as a Universal Intermediary Baseline
Health data exchange has independently certified cybersecurity but no equivalent certification for data quality, accuracy, or provenance. Recommends the CDC include independent data quality and provenance certification in the proposed Universal Intermediary Baseline, building on the NCQA Data Aggregator Validation program.
Colorado AGComment on Colorado ADMT Proposed Rules: Training Data Certification
Recommends that Colorado's ADMT proposed rules recognize independent data certification as a mechanism for developers to satisfy training data documentation obligations, with a rebuttable presumption that certified datasets meet disclosure requirements. Backed by a live 22-station broadcast radio certification pipeline.
CPPAPreliminary Comment on CPPA Opt-Out Preference Signals
Addresses AB 566, AI systems processing opt-out signals, the verification gap, and independent third-party credentialing as a scalable compliance pathway for California's expanding privacy framework.