AI Transparency

Our role, and yours

The Act splits responsibility, and the split matters because part of it lands on you.

• We are a providerof the AI systems in this product — caption writing, image and video generation, voice synthesis, and the assistants. Obligations to mark our output land on us.
• We are a deployer where we point AI at people on our own behalf: our chat and voice assistant, and our automated support replies.
• You are a deployerwhen you publish. The duty to disclose a deepfake or unreviewed public-interest text to your audience is yours, not ours — but you can only meet it with controls we build. That is what most of this page is about.
• We are not a provider of general-purpose AI models. We train no models. The obligations that attach to model providers are not ours.

Telling people they are talking to AI

Article 50(1). Anyone interacting with an AI system must be told, at the start, clearly.

• Our voice assistant discloses before the microphone opens. The disclosure is sent by the server, not spoken by the model, so it holds even if the model never says it.
• Our chat assistantcarries a persistent marker beside its name for the life of the conversation — not only while it is idle.
• Automated replies sent on your behalf identify themselves as AI. On social channels the wording is branded to youraccount, not to us: your follower is told they are talking to your brand’s AI assistant, not to a product they have never heard of.
• Suggested replies you send yourself are not disclosed, and should not be. A person pressed send, so no AI system interacted with the recipient.

Marking AI-generated content

Article 50(2). Synthetic output must be marked in a machine-readable format, using solutions that are effective and interoperable so far as is technically feasible.

• Images and videoare generated with an invisible watermark (SynthID) applied by Google at generation. The models we use apply it as a non-configurable default, so we inherit it rather than request it — they expose no parameter to request it with. What we do control is routing: every generation checks the model that actually served against a list of models we have confirmed mark, and an unrecognised one raises a warning naming the model. That is aimed at the real risk, which is a silent model swap onto something that does not mark.
• Every file we store carries provenance in its own metadata: what produced it, and which model. Rendered graphics carry it in the image file itself; video carries it in the container.
• Uploads are marked too, as not AI-generated. A real photograph that says so explicitly is more useful than one that says nothing, because silence is ambiguous.
• Text is the honest exception, and it carries no mark at all.An earlier version of this page said no watermark in AI-written text was “detectable by any third party”, which implied a mark existed that could not be checked. That was wrong. No watermarking is available for text generated through the APIs we use, so there is nothing to detect. Rather than imply otherwise, we hold the provenance of every generated caption in our own systems and rely on the visible label at the point of publication. We keep a written record of that reasoning and will revisit it if text marking becomes available to us.

Labelling deepfakes and published text

Article 50(4). This is your obligation as the publisher. The product is built so you can meet it by default rather than by remembering to.

• Synthetic likeness is labelled automatically. When a post carries a generated or cloned face or voice, the disclosure is placed at the startof the caption — where it survives the “… more” fold and reaches the viewer at first exposure, which is what the article requires.
• The label is narrow on purpose.A product on a generated backdrop is not a deepfake, and labelling every AI image would train audiences to ignore the label that matters. We label a real, identifiable person’s synthesised likeness.
• You can turn it off, and we record why. The Act allows for evidently artistic or satirical work. Suppressing a label requires a written reason, which is kept with the post.
• Unreviewed text on matters of public interest is labelled.Text a person reviewed is exempt, so we record whether a human actually reviewed each post — as its own fact, distinct from whether it was approved by an automated workflow.
• You see the label before you publish, not afterwards in your own feed.

Evidence you can produce

If a regulator, a client or your own legal team asks what you published and how it was made, you can answer without asking us. Available on Agency plans and above:

• A provenance report— one row per published post: what generated its media, whether a person reviewed it, which label it carried, and any suppression with its recorded reason.
• An audit trail— disclosures served, labels applied and labels turned off, in order, across any period you choose.

What we do not claim

There is no certification scheme for Article 50. Conformity assessment and CE marking exist for high-risk AI systems, which social content generation is not. No vendor can hand you an Article 50 compliance certificate, and you should treat one sceptically.

We are a signatory to the Code of Practice on Transparency of AI-generated Content, for both sections, and you can check that against the Commission’s published list. Adherence is a recognised route to demonstrating Article 50 compliance. It is not certification, and we are not describing it as one.

So we do not describe ourselves as certified, audited, or compliant. We describe what the product does and let you check it. Three specific limits worth stating plainly:

• We do not yet attach C2PA Content Credentials. Our marks are not cryptographically bound to the file, and they do not travel with it once it leaves our storage. We are building it, and we are putting a date on it rather than saying “on our roadmap”: Q1 2027. If that slips we will change this date rather than quietly drop it.
• A watermark embedded in an image can be weakened by heavy editing or re-encoding, including by our own compositing. The provenance we write alongside it is a parallel record, not a repair.
• We cannot discharge your obligations for you. You decide what to publish and to whom. We give you defaults that point the right way, controls when the default is wrong, and the records to show what happened.

Contact

Vertical AI Inc. Questions about this page, or a request for evidence about content published through the product: privacy@verticalinc.ai.