Synthetic Identity Risk

The threat is operational, not theoretical

Voice deepfake fraud incidents grew approximately 680% year-over-year through 2025. The U.S. saw more than 100,000 attacks in that year, with cumulative reported losses crossing $2.19 billion. The technology threshold has collapsed: a usable voice clone now requires three to ten seconds of clean audio. A CEO’s last earnings call, a CFO’s podcast appearance, a client-facing webinar — any of these is sufficient.

The deployment pattern in 2026 is no longer "send a phishing email." It is: an agent calls the target on a real phone line, in a real-sounding voice the target recognizes, and conducts a real-time interactive conversation that adapts to what the target says. More than 50% of CISOs in a recent survey reported a successful deepfake-based intrusion in the prior 18 months. The number was approximately 10% the year before.

Avatar video deepfakes are the visual equivalent of the same trajectory, lagging voice by maybe twelve months. Open-source diffusion models for face animation are now usable on consumer GPUs. Real-time avatar synthesis on a video conference is achievable. Detection by the human eye is not reliable.

The legitimate use cases are also large

This is the difficult part. The same technology is being used responsibly to:

• Render product walkthrough videos in seven languages without re-shooting.
• Automate Tier-0 IT helpdesk — a video agent walks an employee through password reset or VPN setup at 2am.
• Run sales-development outreach with a real human spokesperson’s likeness, with that person’s contracted consent.
• Turn a written knowledge-base article into a 90-second explainer video an end-user actually watches.

The market for this is real. HeyGen, Tavus, ElevenLabs, and Synthesia together represent hundreds of millions in annualized revenue. None of them, as default products, give an MSP the controls needed to ship the technology to a regulated client. That gap is where AiT Avatar Studio and AiT Voice Concierge live.

What the regulatory wave actually requires

EU AI Act Article 50. Fully applicable 2 August 2026. Any AI-generated audio, video, or image must carry a machine-readable mark identifying it as such. The obligation flows through to downstream integrators of GPAI systems, including MSPs reselling avatar or voice technology. The June 2026 Code of Practice is being finalized now and will specify the technical formats — almost certainly C2PA-compatible content provenance plus a steganographic watermark fallback.

TRAIGA (Texas HB 149). Enforceable since 1 January 2026. Penalties up to $200,000 per violation. Affirmative defense: substantial compliance with the NIST AI RMF. This is not optional in Texas.

ELVIS Act (Tennessee) and right-of-publicity laws. States are individually enacting consent requirements for the use of any individual’s voice, likeness, or name in AI-generated content. Retention requirements for the consent record vary; the longest is currently 7 years. Tennessee’s ELVIS Act is the loudest, but at least 12 other states have proposed or enacted similar statutes.

SOC 2 + ISO 42001. AI-specific control families are now part of mainstream attestations. Drummond, Schellman, and A-LIGN all offer AI-aware audits. By 2027, an AI control family is likely to be standard.

Why we excluded Synthesia from our stack

This is a specific, unflattering, and important data point. Synthesia is a strong product. Their terms of service, however, explicitly prohibit white-label resale — meaning an MSP cannot legally embed Synthesia in a product the MSP markets under its own brand. We discovered this during the build of AiT Avatar Studio, after evaluating their API and pricing. We respect that this is their commercial choice. The consequence is that we use HeyGen, Tavus, and ElevenLabs as the underlying providers and have a documented exclusion of Synthesia in our vendor-onboarding playbook. MSPs evaluating Synthesia for resale should read the ToS carefully.

What we built

AiT Avatar Studio and AiT Voice Concierge are companion products with a shared compliance foundation:

• Consent ledger. Every individual whose likeness or voice is used in any output signs a consent video that is archived for 7 years. The signature is non-repudiable. The archive is exportable to the auditor in the format the auditor wants. The retention policy is enforced by the storage layer, not by a calendar reminder.
• Watermark every output. EU AI Act Article 50 compliance is implemented at the rendering layer, not bolted on. C2PA content credentials are embedded by default. A steganographic backup is layered on top so that re-encoded outputs still carry the marker.
• Disclosure language. Every output has a machine-readable disclosure tag and a default human-readable tagline ("This message was generated by AI on behalf of [Company]") that the tenant can customize but not remove.
• Allowed-use boundary. The system refuses prompts that map to known fraud patterns — impersonation of a third party without their consent, banking-account-change instructions, urgency-with-money-transfer scripts. The refusal is logged in the audit trail, including the prompt, the user, and the matching rule.
• Voice-print verification. For Voice Concierge, every voice-render is fingerprinted at synthesis time. If a malicious render of the same voice surfaces in the wild, the fingerprint allows attribution and provides discovery evidence for the rights holder.

What this looks like in practice

A regional bank wants to send onboarding videos in three languages from a video-message of their CEO. The CEO records a one-minute on-camera consent statement granting the bank’s use of his likeness for onboarding content for 24 months. The consent is logged. Avatar Studio renders the three localized videos with watermarking embedded and disclosure language captioned. The Trust Portal entry shows: who consented, when, what they consented to, what was rendered, when, and where it was distributed. Two months later the bank’s auditor asks for the evidence. It exports in three clicks.

A second example: a manufacturing client wants 24x7 voice-based IT helpdesk for their three plants. AiT Voice Concierge deploys with a stock voice (no specific person’s likeness involved). Refusal patterns block any prompt that asks for a financial transaction or claims to be a specific human. Every call is logged and transcribed; the manufacturing IT lead reviews the daily digest. Tier-0 deflection rate at six months: 41%. After-hours call volume to the human team: down 60%. Compliance posture: clean.

What you should ask any avatar/voice vendor

1. Where is consent recorded, and what is the retention policy?
2. How is Article 50 watermarking implemented, and is it removable?
3. What disclosure language ships with each output by default?
4. What fraud patterns does the system refuse, and where can I see the refusal audit log?
5. What happens to a render if it’s captured by an attacker and re-encoded — can I prove provenance?
6. What does your ToS say about white-label or MSP resale?

Where this fits

AiT Avatar Studio and AiT Voice Concierge are part of our AI Customer Surface capability cluster. They depend on the AI Gateway for governed model access, AiT Coord for cross-tenant arbitration, and the Trust Portal for client-facing transparency.