CRAIVIS AI Transparency

1. What is AI-generated

2. What is not AI-generated

• Your logo, your product photos and any image you upload. They are composited by templates, never redrawn by a model.
• Headlines, prices, offers, hours, CTAs and other text on the design. Rendered by our template layer from your brand profile and the approved caption, never requested from the image model. This is why prices in our posts are legible and are the prices you set.
• Occasion dates and sensitivity flags. Curated by people from public government lists and a licensed feed, reviewed by hand; not computed by a model.
• Analytics. Numbers come from the platforms' APIs.
• Your approvals and edits. A person decides what publishes.

3. Labels and Content Credentials

3.1 Inside the file: C2PA Content Credentials. Every exported image, motion post and reel is signed with a C2PA manifest stating that it was generated or edited with AI, with the tool named as CRAIVIS. Manifests are signed with a certificate from a certificate authority; until that certificate arrives manifests are self-signed and carry the same assertions. Where an upstream model already embeds provenance (Google's SynthID, a provider's C2PA manifest), we preserve it; we never strip it.

3.2 On the platform. Where a platform reads Content Credentials or offers an API flag, the label is applied at the platform:

3.3 In the caption. For platforms without a label of their own, and as a second signal everywhere else, the caption carries a short disclosure line ("Made with AI", localised to the caption's language). The in-product name of the setting is "Show 'Made with AI' label" and it is on by default for every brand, including every agency-managed client brand.

3.4 Visible label on the design. Where the law of the brand's country requires a visible label on the content itself (India's SGI rules), our templates reserve a label area and the reel pipeline burns the label in. The exact size the Indian rules require is being confirmed with counsel; until then we use a clearly legible label no smaller than the caption text.

3.5 In the product. Every review card shows which parts are AI-generated, the model tier used, and the "why" behind the post, so approvers know what they are approving.

4. Provenance record

For every generated asset we keep an internal, immutable provenance record: the model and provider, the tier (Draft / Standard / Cinematic), a hash of the prompt, the licence flags of the route used, the C2PA signer, the reviewer (automated gates and, where applicable, the person who approved), and timestamps. The record travels with the asset for its life, is available to you in the post's details, and is produced to a platform or regulator on a lawful request. It is never edited after creation.

5. Override policy

5.1 Default: on, for everyone. The label is on for every brand from creation, including brands an agency creates for its clients.

5.2 What an override can do. A brand Owner or Admin may switch off the caption disclosure line for a brand. Every override is logged (who, when, the reason given) and shown in the brand's audit trail. An agency that overrides for a client brand confirms in the dialog that the client agreed.

5.3 What an override can never do. It never removes C2PA Content Credentials, never removes the platform's own label or API flag, never removes a visible label the law requires, and is not available at all for content that contains a real person's likeness. There is no "strip metadata" tool anywhere in the product, and using another tool to strip it is a breach of the Acceptable Use Policy.

5.4 India. For a brand whose country is India, the SGI rules require synthetically generated content to carry a label and embedded provenance. On Instagram, Facebook and Threads the platform's label satisfies the visible requirement; on platforms with no label of their own the caption line cannot be switched off for Indian brands.

5.5 Why we keep the override at all. Some markets and formats do not require a caption disclosure and some customers prefer the platform label alone. The override exists for those cases, logged, and it is reviewed under our policy-watch process whenever a platform or a law changes.

6. Human approval and safety gates

6.1 Approve-then-post is the default. Nothing publishes until a person taps Approve. Autopilot is opt-in per brand after 10 approvals; every post is shown at least 24 hours before it publishes, you can veto it, daily caps apply, and a one-tap kill switch pauses the brand.

6.2 Six gates run on every post, in every mode:

1. Prompt moderation before any paid generation.
2. Output review: brand fit; factual claims (prices, hours) against your brand profile; regulated-industry claims; competitor names; language quality; text legibility; logo present when required.
3. Real-person likeness blocked without a consent record.
4. Novelty check against your last 90 days and against sibling brands in your organisation.
5. Sensitivity-day check from the Occasion Graph (no promotions on mourning or solemn days).
6. Platform specification validation, C2PA signing and platform AI flags.

A failed gate triggers one automatic regeneration, then the post waits in Draft with the reason. In Autopilot every gate still runs and a per-post audit record is kept. Gates 1, 2, 5 and 6 are live at launch; gates 3 and 4 ship with Autopilot in Week 8. Until gate 3 and the consent record ship, prompt moderation (gate 1) refuses every request to depict a real, identifiable person, so no real person's likeness is generated at all before then; until gate 4 ships, Autopilot is not available, so every post is seen by a person who can spot repetition.

6.3 Kill switches. Our trust-and-safety role can pause generation or publishing per region and per platform in minutes, for example during a disaster or a platform incident. Pauses are shown on the status page.

7. Real people, deepfakes and consent

• We do not generate or alter the likeness of a real, identifiable person unless you record that person's consent in the product (who, when, proof). Sexualised or demeaning synthetic depictions are refused regardless of consent.
• We do not generate content that presents a synthetic event as real, or false information about elections, public health or public safety.
• Content containing a real person's likeness always carries the AI label; the override in Section 5 is unavailable for it.
• Uploading a real photo of a person is your responsibility (Acceptable Use Policy, Section 3.4).

8. Which models and providers we use

We route each task to a model by quality tier, licence and price through our own media router, and we may change a route at any time. We make no single-vendor promise in the product; what we promise is the tier and the licence conditions. As of 2026-09-17:

Vendor terms are a hard filter before price. We use only providers whose terms give commercial rights to outputs at our tier and do not train on customer inputs; we record indemnity, branding and territory conditions per route. Excluded, always: Sora, Midjourney, Suno and Udio. Runway only under an Enterprise agreement without training and with indemnity. HunyuanVideo excluded for EU, UK and Korean tenants. Luma and Higgsfield as last-resort fallbacks only.

9. Music, voices and copyright

• Music beds come only from licensed generators (Google Lyria, Eleven Music, Stable Audio) or a negotiated library, with the licence recorded per asset. Some licences allow social posts but not film, television or games; the licence travels with the reel's provenance record. We never use Suno or Udio and we never add "trending audio" through an API; where you want a platform's trending sound, we hand you the reel with a "post this now" link to add it in the platform's own app.
• Voices are synthetic and never imitate a real person's voice without consent.
• Copyright in AI output is uncertain and varies by country. You can use what we make commercially; nobody can promise you own it. The Terms of Service, Section 8, sets this out.

10. Your data and model training

We do not use your content or outputs to train general-purpose AI models, and we route only to providers whose terms say the same. Your own approvals, rejections and edits personalise your Brand Brain and nobody else's. Our providers are listed in the DPA's sub-processor list.

11. Limits of the technology

• AI makes mistakes. Prices, hours, names and claims in a draft are proposals until you approve them; the output review reduces errors but does not remove them.
• Images may contain artefacts; text inside AI imagery is avoided by design (we render text with templates).
• The novelty check reduces repetition; it does not guarantee uniqueness across the internet.
• Occasion dates that follow lunar calendars are reviewed by people and cross-checked against a licensed feed; you see and can edit occasions three weeks ahead.
• Analytics depend on what each platform's API returns; where it returns nothing (LinkedIn member profiles) we say so.

12. Laws we follow

13. Reporting a problem and takedown

13.1 Report any post we published that is wrong, harmful, infringing, or depicts someone without consent: e-mail contact@adrologic.com with the link to the post, or use the "Report this content" link on any magic-link page or brand page. Tell us who you are, what the content is, and why it should come down.

13.2 Speed. Our internal target is to act within 2 hours of a valid report on everything. Where India's SGI rules set a statutory limit (3 hours, and 2 hours for the categories the rules single out) we meet it. Acting means: removing the post where the platform's API allows (Facebook, Threads, X, Bluesky; TikTok and YouTube when live), or, where it does not (Instagram), instructing the customer to remove it and pausing the brand's publishing until they do, and unpublishing from any page we host. We keep a record of every report and action.

13.3 Counter-notice. The customer whose content was removed can reply at contact@adrologic.com within 14 days; we reinstate only where the law allows and the report was mistaken.

13.4 Consent withdrawal. A person who appears in content and withdraws consent: we stop generating their likeness at once and remove content on the timeline above.

14. Keeping this page current

Platform label rules and AI laws change several times a year. Our policy watch reviews the watched sources (Meta, TikTok, YouTube, LinkedIn, X, Google, Apple, Google Play, MeitY, the EU AI Office, RBI, our AI vendors' licence terms) every quarter and within 7 days of any announcement, records each change in a public change log, and updates the label rules in configuration the same week. This page is versioned; prior versions are at craivis.com/legal/ai/history. Our engineering charter reads: "Official APIs only. OAuth only. Publishing and analytics only; no engagement automation. Labels on by default. When a platform changes the rules, we change the config the same week."

15. Contact

Questions about AI use: contact@adrologic.com. Reports, takedowns and consent: contact@adrologic.com. Grievance Officer: named in the Privacy Policy.