Shield AI — AI System Card

This document discloses, in compliance with EU AI Act Article 50, where Shield AI uses AI in its services, what models are involved, and how end users can identify AI-generated or AI-modified content.

1. Scope

Article 50 of the EU AI Act applies to generative AI systems that produce or interact with synthetic audio, image, video, or text content. The obligations become applicable on August 2, 2026, with a potential December 2, 2026 grace period under the provisional agreement announced on May 7, 2026 (the "Digital Omnibus"). Shield AI publishes this card ahead of that date to give creators, fans, partners, and regulators clear visibility into our AI surface.

The following Shield AI services use AI and therefore fall in scope of this card:

• Deepfake and synthetic-content detection — a multi-model consensus pipeline combining NVIDIA NIM classifiers, Resemble AI voice analysis, and C2PA manifest verification.
• Voice clone consent attestation — the Twin server records and verifies that a voice clone was authorized by the speaker before any synthetic audio is rendered.
• Content provenance and C2PA manifest signing — cryptographic provenance assertions attached to AI-touched assets so downstream readers can verify how an asset was made.
• Brand-safety pre-publish scanning — automated review of draft content for trademark, defamation, and platform-policy risk prior to publication.
• Compliance routing — generation of FTC §255 disclosure language and EU AI Act Article 50 labels for content that requires them.
• Culture intelligence and signal classification — classification of cultural moments along the forming-to-archived lifecycle used to time creator activations.
• Team-pipeline role attribution (S22) — attribution of contributions across a creator team pipeline so that each role can be credited correctly.
• DMCA evidence assembly and takedown drafting — assembly of infringement evidence and drafting of takedown notices. All legal actions are gated by human-in-the-loop review before submission.

2. Models

The table below lists the models and providers Shield AI relies on, the services they support, and whether their output is shown to end users directly.

3. Disclosure Mechanisms

Article 50 requires that AI-generated or AI-modified content be detectable by end users and downstream readers. Shield AI uses the following overlapping mechanisms so that detection does not depend on any single signal:

• C2PA manifest assertions are attached to every AI-touched asset. A reader who pulls the manifest can verify which model touched the asset, when, and under what authorization.
• Inline labels — creator-facing surfaces show an inline #aior "AI-assisted" label on drafts and outputs that involved a model.
• Article 50 disclosure tokens — Shield AI maintains a canonical token set used to mark assets in the product UI and the audit ledger: ai-generated-3d-space, ai-generated-content, and ai-persona.
• The Article50Disclosure banner appears on fan-facing surfaces where AI-touched material is rendered, so that the audience sees the disclosure at the point of consumption.
• The ui_disclosure_log audit ledger is an append-only record of disclosure events. It is queryable on request under GDPR Article 15, so a person who appears in or interacts with an asset can ask what disclosure was shown and when.

4. Limitations

No detection or disclosure system is perfect. Shield AI is transparent about the limits of the ones described above.

• C2PA can be strip-attacked. A manifest can be removed by an adversary or stripped by an intermediate platform. The absence of a manifest is not, on its own, proof that an asset is forged or AI-generated.
• Multi-model consensus is probabilistic, not deterministic. Deepfake and synthetic-content detection produces a confidence band. False positives and false negatives are possible, and results should be read alongside other evidence rather than as a final verdict.
• Voice clone classifiers degrade on novel synthesis methods. When a new generation technique appears, classifier accuracy drops until the model is retrained. Consent receipts attest that a voice clone was authorized; they do not, on their own, prove that a given audio file was produced by an authorized pipeline.
• Article 50 implementation guidance is still moving. Final technical standards and codes of practice are expected between June and August 2026, and may be revised again under the Digital Omnibus. This card is versioned and will be re-issued on any material change.

5. Contact

6. Version History

• v1.0 — 2026-05-19 — initial publication.