AI Disclosures

In short

SuperPost writes, renders, and publishes marketing content using AI. Every post we publish on your behalf carries an AI disclosure — on platforms that provide a native AI-content flag we set that flag, and everywhere else we append a short disclosure line to the post itself. There is no configuration that turns disclosure off.

This page describes what is running today, not what is planned. Where something is not built yet, it says so.

Models and providers we use

Text generation and critique: Anthropic (Claude). Text embeddings, used for the per-workspace quality critic and drift scoring: Voyage AI. Voice synthesis, when you have opted in to voice cloning: ElevenLabs. Generative images for post visuals: fal.ai (Flux / SDXL family). GPU lip-sync video rendering, when a cloned voice is paired with a face image: Modal Labs.

Each of these is a listed subprocessor at superpost.io/legal/subprocessors, with its location and role. Model choice changes as providers ship better models; the subprocessor list is the authoritative, versioned record and carries a 30-day change notice.

We do not train on your content

Your repository content, drafts, edits, voice samples, and published posts are not used to train third-party foundation models, and we do not sell them. Our provider agreements are on zero-retention or no-training terms for API traffic.

We do use your own content to improve generation for your own workspace — your edits adjust prompt context, selection weights, and persona scoring for future drafts in that workspace only. That is per-workspace learning, not model training, and it never crosses a workspace boundary. The one exception is the anonymized cross-workspace winner library, which is strictly opt-in and off by default.

Media we render carries a C2PA provenance manifest whose `c2pa.training-mining` assertion declares AI training, AI inference, and data mining all `notAllowed`.

AI output can be wrong

Generative models produce fluent text that can still be factually wrong. SuperPost writes about your repository and your product, so a mistake reads as a claim you made. We run automated checks — a fact critic, a claim gate that holds posts whose stated facts we could not corroborate, and a persona guard — but none of them is a guarantee, and no automated check catches everything.

You remain responsible for what publishes under your name. Review drafts before approving them, and treat autonomous mode as a delegation of effort, not of accountability. Every post can be reviewed, edited, or vetoed before it ships.

How each platform is disclosed

Two mechanisms, chosen by what the platform's API actually supports.

Native flag — the platform has its own AI-content field and we set it from the label we resolved: TikTok (`post_info.is_aigc`) and YouTube Shorts (`status.containsSyntheticMedia`). The platform renders its own label; we do not additionally alter your caption.

Disclosure line — the platform has no such API field, so we own the disclosure and append a short line as the final block of the post: X, LinkedIn, Instagram, and Facebook. Meta's "AI info" label has no Graph API parameter (Meta applies it platform-side), and X and LinkedIn have no field at all, so the caption line is the only publish-time control we hold. The lines are short by design: "🤖 Written with AI", "🤖 AI-generated image", "🤖 AI-generated video", "🎙 AI-narrated", "🎙 AI voice clone", or "🤖 AI-assisted, human-edited" — whichever matches how that specific post was made.

An autonomously published post can never carry no label. If nothing upstream stamped one, the publish path applies "🤖 Written with AI" as a floor rather than shipping the post undisclosed.

Assisted posting: you paste, you confirm

Some platforms forbid fully-automated publishing of public content (TikTok's Content Posting API is the current case). There we upload the video to your in-app drafts or hand you a caption package to paste, and you complete the post yourself.

The caption package includes the disclosure line, and when you mark the post as done we ask you to confirm the line stayed in what you actually posted. That confirmation is required: without it the record is refused, because we will not log a post as disclosed when we cannot see the caption that shipped. Where a platform's draft API does not accept the AI flag — TikTok's inbox upload takes no `post_info`, so we cannot set `is_aigc` on a draft — you may also need to set the platform's own AI toggle in its app.

C2PA provenance on media

Every image and video we publish carries a C2PA manifest — a signed, machine-readable record of what produced the asset: the render actions, the provider and model, and, for voice-derived renders, a reference to the specific row in the voice-consent ledger that authorized it.

Today those manifests are signed with our own key and published as a detached JSON sidecar alongside the asset. They are signed, not yet verifiable against a public trust chain: that requires an operator certificate from a recognized issuer, which we have not provisioned. We will say "verifiable" on this page when that is true and not before. Embedding the manifest inside the media file itself (JUMBF) is also pending, gated on tooling availability.

Not everything visual is AI-generated

Much of what we render is a designed, programmatic composition of your real product — a screen capture, your actual UI, your own copy, animated to a template. That is design tooling, not generative media, and over-labelling it would be its own kind of inaccuracy.

Synthetic elements — an AI narrator, a cloned voice, a diffusion-generated image or b-roll — are genuinely AI-generated and are labelled as such. Where our provenance data cannot yet distinguish the two, we currently apply the more conservative label. Improving that resolution so the label matches the artifact is active work, and this section will change when it lands.

EU AI Act Article 50

Article 50 of Regulation (EU) 2024/1689 obliges providers and deployers of generative AI systems to make it clear when content is artificially generated or manipulated, in a machine-readable form where feasible, and to inform people they are interacting with an AI system. Those transparency obligations apply from 2 August 2026.

Our position: SuperPost is a deployer of third-party general-purpose models rather than a provider of them, and the content it produces is marketing communication published under your name. On that basis we treat the Article 50 marking and disclosure duties as applying to every post we publish, and we discharge them through the per-post labels described above plus the C2PA manifests on media — machine-readable marking where the format allows it, plain-language disclosure in the post where it does not. Where a platform's own AI label is the mechanism the audience actually sees, we set it.

This is our own reading, published so you can see the reasoning rather than take a claim on faith. It has not yet been reviewed by external counsel, and the provider-versus-deployer allocation between us, you, and the model vendors is one of the specific questions in our counsel engagement. If review changes our position, this page changes with it.

What is still yours

Platform policies and local law may require more of you than we can do from our side. Some platforms require the account holder to set an AI toggle in their own app; some jurisdictions regulate advertising claims regardless of who wrote them. Our disclosure is a floor, not a substitute for your own compliance — and where you are unsure, ask before you publish rather than after.

If you believe a post published through SuperPost was not properly disclosed, tell us at legal@superpost.io. We keep an audit record of the disclosure applied to every post — the label, the mechanism, and the mode — and we can tell you exactly what shipped.