AI Disclosure Policy (draft outline)
Updated October 10, 2026
Draft — pending legal review; not yet in effect
This document is a working draft prepared for counsel. It may change before launch and is not yet in effect.
Draft for legal review — not in effect.
This outline records how the product handles AI and endorsement disclosure today (from docs/PLATFORM_POLICIES.md, the QA engine in packages/qa, the persona gates and the composer) so counsel can turn it into a public policy. The practical guide for customers is AI disclosure in practice.
1. Principle
- "Transparent fiction, not deception": AI personas and AI-generated media are allowed when they are disclosed and do not mislead.
[COUNSEL: public statement of the principle].
2. Where disclosure happens in the product
| Layer | Behaviour today |
|---|---|
| Persona profile | A persona's social bio must say it is AI or fictional. The app checks the bio and shows "Disclosure missing" until a marker is present; a missing disclosure is a persona gate. |
| Content (caption, on-screen text, speech) | For content with a synthetic person, an AI voice or a digital twin — or when the workspace policy or a rule pack requires it — AI review requires a disclosure marker in the caption, on screen or in speech. Missing ⇒ a hard gate that nobody can override. |
| Plain AI-generated media (no persona) | Not a hard gate; native platform AI labels are applied per network at publish time. [COUNSEL: decide whether a default visible "AI-generated" disclosure is required for all AI video — open founder policy decision since Phase 1] |
| Native network labels | The composer sets the network's AI-generated and branded-content labels wherever the account allows it through the API. Required labels are locked on. When an account cannot set a required label through the API, the post is exported with a checklist to apply it manually. |
| Endorsements | Classification sponsored / gifted / affiliate / own product; sponsor tag and affiliate disclosure fields; missing endorsement disclosure is a hard gate. |
| Provenance | Every asset records model, prompt version, sources and edits. Provider watermarks are preserved (Google Veo outputs carry an invisible watermark). C2PA Content Credentials are planned: the manifest field exists, but no signing certificate is configured [Founder: C2PA certificate — BRIEF §123]. |
3. Jurisdiction and platform rules
- Rule packs are versioned records in the database (severity INFO / WARNING / BLOCK) with source and verification date; BLOCK findings are hard gates.
- EU AI Act Article 50 transparency obligations apply from 2 August 2026; the registry notes a transition to 2 December 2026 for marking obligations of systems placed on the market before that date.
[COUNSEL: our role (provider vs deployer), required marking, deepfake disclosure wording]. - US endorsement guidance (FTC): AI personas are not assumed exempt.
[COUNSEL: review the US endorsement rule pack]. - High-risk verticals (health, finance, legal, alcohol, politics/elections, minors): education-only, claims tagged and substantiated, specialist review required before publishing.
4. Real people
- Digital twins and cloned voices require a recorded consent (subject, allowed uses, territories, expiry); minors are blocked; public-figure look-alikes are blocked. Revoking consent invalidates open approvals of affected content.
5. What the customer must do
[COUNSEL: customer obligations — keep disclosures intact when downloading or exporting content, apply manual labels when the export checklist asks, do not remove watermarks].
6. Exports and downloads
- Content downloaded from the app leaves the platform's publishing gates.
[COUNSEL: customer responsibility for disclosure on content published outside InfluenceBot].
7. Review cadence
- The policy registry is re-verified before the first Live adapter, after incidents and at least quarterly; every verification creates a registry revision.
[COUNSEL: public change log].
